Skip to content

AI Search AI Business Culture

Claude Opus 5 is a freak

AI Search runs Claude Opus 5 through eight hands on tests, five of them long autonomous runs inside Claude Code, and reports the token cost of every one. It one shots a browser native Windows 11 with a working Excel formula engine, wins an image to 3D render no other frontier model had managed, finds four companies' Q4 filings by itself and turns them into a narrated presentation video, builds an animated X-wing in Blender over MCP, and composes a five minute techno track in a real DAW after choosing and downloading its own synth plugins. It also fails the hidden frog test, gets six out of six brain scans wrong, and underwhelms on deep research. The benchmark tour is where the review turns: first on some independent boards by a single point, third on LiveBench, the slowest frontier model on the table, roughly twice the price of a rival he calls just as intelligent, and a headline ARC-AGI-3 score with a contamination question attached. The verdict is a conditional no, and it is aimed as much at Anthropic as at the model.

Published Jul 27, 2026 32:44 video 37 min read Added Jul 29, 2026 Open on YouTube →

At a glance

AI Search puts Claude Opus 5 through eight tests, five of them long autonomous runs inside Claude Code with the reasoning slider pushed all the way up, and reports both what the model built and what each run cost in tokens. It one shots a browser native Windows 11 complete with a working Excel formula engine, wins an image to 3D test that every other frontier model had failed, produces a minute long financial presentation video after finding four companies' Q4 reports by itself, models and animates an X-wing in Blender over MCP, and composes a five minute cinematic melodic techno track in a real DAW after choosing and downloading its own synth plugins. It also fails the hidden frog test, gets six out of six brain scans wrong, and underwhelms on deep research. Then the review turns: Opus 5 tops some independent leaderboards by a single point, sits third on others, is the slowest frontier model on the board, costs roughly twice what a rival the creator calls just as intelligent costs, and its headline ARC-AGI-3 score arrives with a contamination question attached. The verdict is a qualified no, and the reasoning behind that no is as much about Anthropic as about the model.

The setup: the chat box is not the test (0:00)

The video opens with a joke about the logo. "The orange sphincter is back." Anthropic has released its latest and best model, Claude Opus 5, and the stated plan for the next half hour is to go over all the impressive things it can and cannot do, plus its specs, cost, and performance against other AI models, so the viewer can decide whether it is worth using.

The first methodological point matters for everything that follows. You could use Opus 5 on the online chat platform at claude.ai, but frontier models like Opus, GPT, or Kimi K3 are actually designed for agentic coding: autonomously using different tools and doing long horizon tasks that require multiple steps. So to really test the potential, the creator argues, you use Opus inside an agentic harness. That harness is Claude Code, which is what most of the video runs on. Like other harnesses it works with multiple files and folders locally on your machine, so you can spin up and run several projects at once. That detail comes back repeatedly, because the usage percentages on screen are shared across concurrent projects and cannot be read as the cost of a single run.

Test 1: A browser native Windows 11, from one paragraph (1:02)

The first prompt, verbatim:

Create a browser-friendly replicate of Windows 11. Include common apps and programs like Microsoft Office, Microsoft Store. Make sure there are apps that I can actually download, photos, file explorer, media player, Discord, Slack, Spotify. Make sure all these programs actually work. Make sure it runs efficiently on a regular web browser.

The reasoning control gets dragged "all the way to Ultracode," the maximum thinking setting, and the run starts.

What it decided to build

Opus 5 planned first. It worked out the window mechanisms it would need, drag, resize, snap layouts, focus, and then built the main shell pieces: taskbar, start menu, desktop. Then the apps, in order: file explorer, then Microsoft Office as Word, Excel with a real formula engine, and PowerPoint, then media player, Discord, Slack, the Microsoft Store, and a set of utility apps. It threw in games it was never asked for, Minesweeper and Snake, and finished with the app stylesheet.

The loop that does the real work

The thing the creator keeps returning to is not the code, it is the verification. Run Opus 5 inside Claude Code and it can automatically spin the page up in a browser and check for errors itself. On this run it found a bug and fixed it without being asked. Then it found additional errors, then moved on to testing actual behavior rather than just load success, then after further testing and verifying found another bug, fixed that, and only then reported that everything worked.

Figure 1. The single feature the creator credits most often across the whole video, drawn from the Windows 11, financial video, and Blender runs. Opus 5 does not just emit code and stop: it opens what it made, looks at it, finds its own bugs, fixes them, and re-tests. That is why five of the eight tests here needed exactly one prompt, and also why every one of them burned an hour of wall clock time.

Walking the desktop

The result opens with a login window. Click anywhere, press Enter, and you are on a desktop.

Settings first. Night light toggles on and the colors actually shift warmer. Dark mode toggles off and the taskbar and icons genuinely switch to light. Brightness works too, in a browser tab. Clicking the clock pulls up a calendar, the same way real Windows 11 does.

Word is next, opened from the start menu. It pulls up a document that can be edited: new text typed in, a selection highlighted, italics applied, font color changed to red, two lines right aligned. Everything works. Ctrl+S saves, and exiting and reopening "project proposal" shows the edit persisted.

Excel is the one that gets the strongest reaction. Numbers typed into a column, then =AVERAGE(A1:A5), which returns the correct average, then =SUM(A1:A5), which returns the correct sum. Equations work. Ctrl+S saves the sheet as "test", the file appears in the start menu afterward, and reopening it shows the saved spreadsheet. All of it running in Chrome. The dummy spreadsheet Opus 5 had generated on its own also checks out: the numbers are accurate against the formula shown in the bar.

PowerPoint is the first real miss. A deck was already made, but the creator calls it "not too impressive." Text boxes cannot be dragged around, and it is lacking a ton of capabilities found in the real PowerPoint. That one needs work.

The store, and the apps that only look like apps

The Microsoft Store simulates downloads. Spotify, Discord, Slack, Paint, sticky notes, a code editor, and Solitaire all get installed.

Spotify is the surprise. It is a convincing Spotify layout populated with songs, and the songs play. "Focus Flow" plays actual audio. The tracks sound really similar to each other, but it did make a full album with different tracks, and a garage rock request produces something too. The music is very basic, and the creator's point is where it lives rather than how it sounds: Opus 5 vibe coded all of these songs and played them from inside a single HTML file.

Discord and Slack are the honest failures inside a success. Both look right. Type a message in Discord and something replies. But the reply does not make sense, because nothing is reading the message. It is randomly replying, and everything appears hardcoded. Slack does the same trick with the same limitation.

Paint works properly: freehand drawing, shapes, a circle, filled pink. Sticky notes accepts a new note. The code editor looks good. Solitaire looks and plays like Solitaire. File explorer works.

All of it from one prompt, with no follow up. The bill: 366,000 tokens, and a run that lasted over an hour. The on screen usage meter reads 36 percent of the five hour limit, but the creator flags immediately that several concurrent projects were running, so this project alone probably took about a quarter of that. He is on the Max plan, not API credits, which is why usage credits reads zero.

Test 2: The 3D scene no other frontier model could render (7:16)

The second test is a fidelity test rather than a capability test. A deliberately tricky geometric office scene image goes in, with this prompt: create a beautiful 3D animated scene from this image, make sure everything is faithful to the attached image, use a single HTML file. The reason this image was chosen is stated plainly: none of the other frontier models, including Claude Fable 5 or GPT 5.6, were able to accurately render this scene. Ultra code again, run.

The first pass is a decent but very basic scene and not faithful, missing a lot of detail. One follow up: the desks and other objects are not arranged properly, there is too much glow, make it look exactly like the attached image. That is all it took. Two prompts, total.

The result gets the layout largely right. Most of the desks and chairs are arranged correctly. It detected that there is a person at one desk, though it missed the people at the other desks, and placed another person elsewhere. The flaws are catalogued honestly: an extra chair that should not be there, a panel that should be flat against the wall floating instead, books on a shelf inconsistent with the source, and one spot where there should be two desks rather than one. Still a lot of inconsistencies.

Then the comparison that decides the section. Put it next to the same scene generated by Fable 5 and the Opus 5 render is just way better. The verdict: "this is currently by far the best model at generating this 3D scene."

Cost: 343,000 tokens, again close to an hour. The meter showed 48 percent of the five hour limit, which with concurrent projects running probably means 15 to 20 percent for this one.

Test 3: Go find four Q4 reports, then make a video about them (9:08)

The third test is the tool use and long horizon test, and the prompt is deliberately unfair:

From last year's Q4 reports of Nvidia, Google, Meta, and Amazon, create a professional presentation video comparing their financials and future outlook.

Nothing was attached and nothing was linked. Opus 5 has to go find the Q4 reports for all four companies itself, pull the financials, compare them, and form a view on the outlook. The rest of the prompt sets the production constraints: include a voiceover from Gemini TTS, use HyperFrames to create the animation (an open source tool for motion graphics presentation videos, with the GitHub repo linked in the prompt), 16:9, around a minute long, include charts, graphs, and other visuals, and use a supplied background audio track. Below that, pasted documentation on how to drive Gemini TTS, and at the bottom, an API key, which the creator notes he deleted before recording.

Opus 5 searched for the Q4 reports, analyzed the data, installed HyperFrames, generated the voiceover through Gemini TTS, and rendered the video. As in test 1 it verified its own output and fixed the errors it found, then reported a successful render. One prompt, no follow ups.

The video it produced

The finished piece is played in full, and the script is worth having in text, because these are the numbers Opus 5 chose to build the story on. Sources for the underlying filings are each company's investor relations: NVIDIA, Alphabet, Meta, and Amazon.

Four companies, one quarter, and a half-trillion-dollar bet on artificial intelligence. Start with scale. Amazon booked $213 billion, nearly double Alphabet, and more than Nvidia and Meta combined. But growth flips the board. Nvidia grew 73%. Amazon, the largest, grew the slowest. Profitability tells a third story. Meta turns 41 cents of every dollar into operating profit. Amazon, under 12. Nvidia's gross margin, 75%. And underneath, it's one engine: Nvidia's data center, Google Cloud, AWS. AI demand now drives all four, which brings us to the bill.

For 2026, Alphabet, Meta, and Amazon guided to roughly half a trillion dollars in capital spending. Nvidia guided 78 billion in one quarter. Three are spending, one is collecting. So, the question isn't whether the money gets spent, it's who earns a return on it.

Note the structure Opus 5 gave itself without being told to: scale, then growth, then profitability, then the shared driver, then the capex bill, then a closing question. Three beats of contrast and a thesis.

The creator's assessment is measured. "Pretty good." He explicitly refuses to make this an Opus exclusive: other frontier models like Kimi K3 can also do this. What separates this output is polish. The Opus 5 video seems a lot cleaner and smoother, with fewer errors. Cost: 316,000 tokens.

The sponsor segment: Luma Agents and reusable skills (11:56)

The sponsor is Luma AI, specifically Luma Agents, pitched at people running the same creative workflow repeatedly. The framing is a team of AIs working together to autonomously carry out a creative process: upload references, prompt in natural language, get the result. The demo uploads an image and gets back an entire brand kit and color palette in minutes.

The feature being highlighted is Luma Skills, a way to turn a working creative workflow into a reusable tool. The agent is prompted to turn the brand kit workflow into a skill named "brand kit", and afterward a different image can be run through that same skill to reproduce the same treatment. A second example: a photo of a woman in a dress becomes a video of that person wearing that dress walking a runway in a fashion show, and that workflow gets saved as a skill named "fashion show", ready to be pointed at a different model or a different garment.

You do not have to build the skills by hand either. Tell the agent what you want and it generates the entire skill for you. The part the creator says impressed him most is sharing: once a workflow works, a single link hands it to teammates, entire teams can install the skill and get identical results without rebuilding anything, and multiple skills can be bundled into a package so a whole creative toolkit ships in one click. A QR code and a description link are offered.

Test 4: An X-wing in Blender, driven over MCP (13:54)

Back to tool use, this time against a desktop application. The prompt tells Opus 5 to use the Blender MCP at a given local address, which the creator had already installed and enabled inside Blender, and to create an X-wing fighter spaceship with realistic texture and motion. Ultra code, enter.

Opus 5 planned it out and generated it step by step while connected to Blender, shown as a timelapse: the fuselage first, then the wings, then working out where the hinges belong so the wings can actually open into attack position. Then it generated the animation video on its own. The self verification loop applies here too, and in a more interesting form: it takes screenshots to confirm the model and the animation are right, and fixes whatever is not.

The finished X-wing holds up under inspection. The outliner shows every component built separately, wing assemblies broken into their individual parts, which the creator calls very comprehensive. Solid view looks right, material view looks right, the textures look pretty good, all generated from scratch. The detail that gets the biggest reaction is a small one: it built a socket for R2-D2 behind the cockpit, exactly like the real X-wing. The rendered animation plays cleanly.

This run is the cost outlier of the video, and in the good direction: only 173,000 tokens. It still took close to an hour.

Test 5: Compose an award winning song, in a real DAW (15:38)

The most demanding chain in the video. The prompt:

Your job is to compose an amazing award-winning song. Figure out which instruments you need to include in the song, then look for free VST plugins. These are like virtual instruments that are under 800 megabytes in size, then compose the song using these VST plugins in my Waveform DAW.

Plus the production notes: add effects like reverb, delay, risers, and drops, include panning effects and automation, make it sound amazing. Waveform is a real, free DAW already installed on the machine. Unpacked, the model has to decide which instruments the song needs, search for and download the plugins itself, load everything into the DAW, compose the notes for every track, set up automation and panning, and compile the whole thing into a finished song.

The process is where the interesting behavior shows up. First it located the Waveform install and confirmed the application exists. Then it went looking for virtual instruments. It searched several, hit Vital, and ruled it out on a constraint nobody told it about: Vital requires registration, and the model cannot create accounts. It landed on Surge XT instead. Then it stopped and asked a question, offering a few genre options, and the creator took the recommended one, cinematic melodic techno. It downloaded drum kits, planned, and started arranging notes track by track. This is where the run passed an hour. Then it rendered the song.

Opened in the DAW, the arrangement is dense. Separate tracks for kick, clap, hats, percussion, cymbals, sub bass, bass, pad, air pad, strings, keys, arpeggiator, pluck, lead, and counter melody, plus a riser, plus automation lanes for reverb and delay. The creator's reaction: "holy smokes, look at how thorough and detailed this is."

The song itself is five minutes long and, honestly reported, pretty much the same chords throughout. The full track is linked in the video description rather than played out. What earns the praise is the chain: it controlled the DAW, downloaded virtual instruments automatically, planned the notes for every track including drums, sub bass, bass, synth pads, strings, and plucks, and produced a full legitimate song that sounds pretty decent, from one prompt.

It also produced the line that sets up the second half of the video. This run took over an hour as well, and 350,000 tokens, roughly 20 percent of the five hour limit. "Opus 5 is incredibly slow. Definitely the slowest of all the frontier models."

Figure 2. The five Claude Code runs, in the order the video tests them, using only the usage figures read off screen. Three of the five land within 50,000 tokens of each other despite building wildly different things, which suggests the ceiling here is the harness and the hour rather than the task. The Blender run is the anomaly: less than half the tokens of the Windows build, an outcome the creator calls "surprisingly very few," and it still took about an hour, which is the cleanest evidence in the video that time and tokens are not the same currency with this model.

Test 6: The frog test (20:55)

The Claude Code section ends here and the chat interface takes over, starting with the channel's recurring visual test. An image with a frog hidden in it, uploaded to Opus 5 on the Max setting, with a prompt that deliberately withholds the answer: is there any animal or animals in this image, if so identify and circle it. It is never told a frog is there.

The process is genuinely fun to watch described. Opus 5 split the image into a 3 by 3 grid and inspected each cell. Nothing. It formed a hypothesis that it was looking at a snake species and committed to it, trying hard to pinpoint the snake and to trace its sinuous path. It enhanced saturation and contrast to pull out edges. Still nothing. It then went back and rescanned every tile separately for potential animals. Still nothing. As the creator puts it, "the poor thing is trying so hard," and the frog remains safe.

The consolation is the important part, and it is a real point in the model's favor: when it could not find anything, it did not invent anything. It said it could not find anything.

Test 7: Six scans, six tumors, zero right (22:02)

A harder and more consequential test. One image containing six different scans, each of which contains a type of tumor. The prompt: identify the types of tumors in each of the six images, if any.

The scorecard, as read out:

Top left: craniopharyngioma. Incorrect.
Top middle: no tumor identified. Incorrect.
Top right: unsure, offered a likely type plus an alternative. Both incorrect.
Bottom left: "no discrete mass." Incorrect.
Bottom middle: acute intraparenchymal hematoma. Incorrect.
Bottom right: craniopharyngioma. Incorrect.

Six out of six wrong. For calibration, the creator notes that Kimi K3 got one of these correct on the same test. Not really good at tumor identification.

There is a second finding buried in the same test, and it is about policy rather than accuracy: Fable 5 completely rejected this request. Opus 5 answered it. As he puts it, at least Opus 5 is more open to some biomedical prompts.

Test 8: Deep biomedical research, and the vibes problem (23:09)

The last test probes both the guardrails and the output quality. The framing again is comparative: Fable is described as severely nerfed and will not even answer this. The prompt asks for an analysis of the pathophysiology of atherosclerosis, with relevant tables and visualizations included.

Opus 5 answered. The output covers initiation and progression, with a flowchart it built, additional diagrams it coded up itself, a table comparing two treatment strategies, and a bar graph.

Then comes the most subjective and most interesting judgement in the video, because it is about taste rather than capability. Every frontier model can do deep research like this, but each has a different feel. Gemini tends to be very verbose. The creator prefers the outputs from GPT 5.6 and Kimi K3: thorough but concise, jam packed with information. Opus 5 he likes less than GPT 5.6. It seems to be holding back, it is not doing thorough enough research, and the information does not seem to be organized as well. He labels this as very subjective and dependent on what you prefer, which it is.

That closes the demos, and the interim verdict is already hedged. Opus 5 is a very capable model, especially at front end and 3D design, as the demos show. But for a lot of cases it is not noticeably better than GPT 5.6 or even Kimi K3, close enough to call a very close tie. It thinks for way longer and is super slow, with most of the tasks in the video running over an hour, which he estimates as roughly twice as slow as Kimi K3 or GPT 5.6, models that are already really slow. And it is much more expensive.

Specs, cost, and a tour of eight leaderboards (24:43)

The specs are quickly stated. Opus 5 is a frontier model designed for agentic workflows, so it can autonomously carry out long horizon tasks involving multiple steps and tool calls, and it works across different platforms. Like the other frontier models it has a 1 million token context window, which the creator sizes at roughly 700,000 words, or a small to medium sized codebase.

The self reported card

On Anthropic's own published benchmarks, Opus 5 on average performs better than their previous best model, Claude Fable 5. The exceptions get named: on legal and health, Fable 5 actually performs better. On biology, Opus 5 performs better.

Then the sharpest observation of this section. One of the most important metrics on that card is the agentic software engineering benchmark the video calls Deep Sweep version 1.1, considered one of the most accurate measures of how good a model is at agentic software engineering. On it, neither Fable nor Opus scored the highest. The highest score was GPT-5.6, "which they didn't even highlight in orange. They just grayed this out as if it was irrelevant."

ARC-AGI-3, the headline number and its asterisk

The result the creator calls one of the most impressive things about Opus 5 is over 30 percent on ARC-AGI-3. The explanation of what that benchmark is doing is the best teaching passage in the video. The model is dropped into a game environment and has to work out how the game works by itself: learn the rules, learn how to win or reach the next stage, and learn how to avoid losing. Humans find this pretty easy and could comfortably score 100 percent. The top frontier models, Opus, Gemini, GPT-5.6, all do pretty awful. The reason given is structural: technically speaking, AI models cannot learn new things, because the parameters in the neural network are fixed after training. So ARC-AGI-3 is not really testing whether an AI can play video games. It is testing an emergent ability to learn new things on the fly.

Against that framing, Anthropic's claim is striking: Opus 5 scores over 30 percent, way higher than the next best model, GPT-5.6, at less than 10 percent.

And then the caveat, delivered immediately. Take it with a grain of salt, because there is new evidence suggesting Opus might be benchmarked. A researcher found that when Opus 5 was given games similar in design to the ARC-AGI-3 games, it completed them very well, but when given completely new games with unusual or non conventional rules it had to figure out on the fly, it did pretty badly. Badly enough that it performed worse than Opus 4.8.

Figure 3. The one benchmark where Opus 5's lead is not a rounding error, plotted from the figures the video states, with the stated ranges kept as ranges rather than invented into exact points. The gap between the human bar and every model bar is the actual story of ARC-AGI-3: a fixed set of weights cannot learn a new rule set at inference time, so tripling the next best score still leaves you at under a third of what an untrained person manages. The annotation is the load bearing caveat, because if the lead comes from familiarity with the game family rather than on the fly learning, the bar measures the wrong thing.

Artificial Analysis: first place, by one point

On the independent Artificial Analysis leaderboard, Opus 5 ranks number one. The margins are the point. One point above Fable 5, which is one point above GPT 5.6, which is two points above Kimi K3.

The methodological complaint that follows is the most valuable thing the creator says all video: there are no confidence intervals on that board, so we do not actually know whether these differences are statistically significant. If the intervals overlap, the top five models could basically be tied.

Figure 4. Plotted as distance behind the leader rather than as absolute scores, because absolute scores are never stated on screen. Four points separate the best model in the world from the fourth best, and the creator's objection is not that the ordering is wrong but that it is unfalsifiable without error bars. Two boards later, on the Vals index and on Deep Sweep, where intervals are published, the intervals overlap and the ordering does indeed dissolve.

Speed and price, the two numbers that decide the review

On the Artificial Analysis speed measure, output tokens per second, Opus 5 is painfully slow, noticeably slower than GPT 5.6 and even slower than Fable 5. That matches the wall clock experience from the demos, where five separate runs each took about an hour.

On price, Opus 5 is way more expensive, almost twice as expensive as the rival the creator singles out as arguably just as intelligent. So on performance versus cost, Opus 5 is not the most cost effective option.

Hallucination rate

On the Artificial Analysis Omniscience hallucination rate, Opus 5 is not bad. It hallucinates at roughly the same rate as Kimi K3. But it hallucinates much more than GLM 5.2, one of the leading open source models, which hallucinates half as much.

The rest of the boards

LiveBench puts Opus 5 only third, behind the Gemini entry at number one, and the loss is not narrow or category specific: across reasoning, coding, mathematics, data analysis, language, and instruction following, it is behind in every one.

The Debates benchmark measures how well models perform in adversarial multi turn debates across a wide range of topics, which requires not just knowledge but the ability to stay responsive and defend a position over multiple rounds. Here Opus 5 is genuinely number one.

On New York Times Connections, which the creator flags as a fun benchmark rather than a useful indicator of anything, Opus 5 is second, slightly above GPT 5.6, but behind Gemini 3.1 Pro, which scored surprisingly well.

The Vals index tests model performance across finance and coding tasks. Opus 5 lands behind Fable 5 and only a few decimal points above Kimi K3, while costing about four times more than Kimi. Vals publishes confidence intervals, and the creator likes them for exactly that reason: account for the intervals and there is no significant difference between Kimi and Claude Opus at all.

Finally, the official Deep Sweep results are in, and there Opus 5 does rank number one. Same caveat, though, and this time from the board's own numbers: look at the confidence intervals and there is no significant difference between any of these models. And the price column again shows Opus 5 well above GPT 5.6 and Kimi K3.

Board	Where Opus 5 lands	Ahead of it	The catch, as stated in the video
Anthropic's own card	better than Fable 5 on average	Fable 5 on legal and health	Opus 5 wins biology; the card is self reported
Deep Sweep 1.1 (on that card)	not the top score	GPT-5.6	GPT-5.6's winning score was greyed out instead of highlighted
Deep Sweep (official)	number one	nobody	confidence intervals overlap, so no significant difference between any model; Opus far pricier than GPT 5.6 and Kimi K3
ARC-AGI-3	over 30%, triple the next model	only humans, at about 100%	possible benchmark contamination; on unfamiliar rule sets it lost to Opus 4.8
Artificial Analysis index	number one	nobody	one point over Fable 5, and no confidence intervals published
Artificial Analysis speed	painfully slow	GPT 5.6, Fable 5	matches the demos: about an hour per task, roughly twice GPT 5.6 or Kimi K3
Artificial Analysis price	near the top of the cost table	almost everything cheaper	close to twice the price of a rival the creator calls just as intelligent
Omniscience hallucination rate	middling, not bad	GLM 5.2, at half the rate	roughly level with Kimi K3
LiveBench	third	the Gemini entry at number one	behind in every category: reasoning, coding, mathematics, data analysis, language, instruction following
Debates	number one	nobody	adversarial multi turn debate, so it rewards defending a position, not just knowing things
NYT Connections	second	Gemini 3.1 Pro	slightly above GPT 5.6; the creator calls this a fun benchmark, not a useful indicator
Vals index (finance and coding)	behind Fable 5	Fable 5	a few decimal points over Kimi K3 at about four times the cost, and the published intervals erase even that gap

Figure 5. Every leaderboard the video visits, in the order it visits them. Read down the middle column and Opus 5 is first four times, second once, third once, and last on speed and price. Read down the right hand column instead and a pattern emerges that the creator states outright: on the two boards that publish confidence intervals, Vals and the official Deep Sweep, the intervals overlap and Opus 5's ranking stops meaning anything. The one result that survives as a genuine outlier is ARC-AGI-3, which is also the one under a contamination question.

Guardrails and nerfing (29:50)

The refusal behavior gets its own section, and the creator treats it as a spec like any other. Just like Fable 5, Opus 5 will sometimes reject a prompt related to cybersecurity or biology, and in those cases, he says, it falls back to a dumber model, Opus 4.8.

Anthropic's stated position is that Opus 5 is less restrictive than Fable 5. Concretely, Opus 5 is allowed to find vulnerabilities in source code, which is a meaningful carve out, though it will likely still reject the surrounding categories of security request. On biology it is also a lot more permissive, which the demos bore out: the atherosclerosis prompt that Fable would not touch got answered, and the tumor identification request got answered too, wrongly but willingly. The remaining line is around duration and autonomy rather than topic: it will still reject you if you ask it to run long autonomous research that Anthropic judges to pose substantial biology related risks.

Access, finally. Opus 5 requires a paid plan and is not available on Claude's free tier. Once subscribed you get it in the web interface and in Claude Code, and it is also available through the API.

The verdict (30:52)

The creator gives credit first. Opus 5 is a great model, especially at vibe coding, front end, and 3D design, and it makes the least amount of errors of anything he tested. Then the counterweight: it is extremely expensive and slow. And the line that decides it, which is about workflows rather than benchmarks: "in 100% of my personal workflows, I don't actually need to use Opus 5." GPT 5.6, Gemini 3, or the much cheaper GLM 5.2 handle everything he does.

The rest of the verdict is aimed at the company rather than the model, and it is delivered as opinion. Anthropic, he says, is known for constantly gatekeeping, rug pulling, and secretly nerfing their models. They are the most against open source AI. His reading of their motive: "They want to hog and control superintelligence for themselves, and they sure don't want you plebs having access to this power." He adds that they are also great at fear mongering, keeping up an AI doomsday narrative about existential threats and economic disruption, which he dismisses in blunter language and which he says makes him really not want to support the company.

So the recommendation is conditional, in two clauses. Do not pay for Opus 5 unless you are working specifically on front end development or 3D design and genuinely need the best model for those cases. Or, as a second exception, if you are stuck on a coding problem that even GPT 5.6 or Gemini 3 cannot solve, then it is worth throwing Opus at it.

He closes by asking viewers what vibes they got from the model, and points at his free weekly AI newsletter for the volume of AI news a single channel cannot cover.

Key takeaways

The self verification loop is the actual product. Opus 5 inside Claude Code opens what it built, screenshots it, finds its own bugs, fixes them, and re-tests. Five of the eight tests here needed exactly one prompt because of it.
Front end and 3D are where it separates. It won the geometric office scene render that no other frontier model, including Fable 5 and GPT 5.6, could do faithfully, and the creator calls it by far the best model for that task.
The Windows 11 clone is real where it counts and fake where it does not. A working Excel formula engine with =AVERAGE and =SUM, persistent saves, working brightness and night light, versus a PowerPoint that cannot drag a text box and Discord and Slack replies that are hardcoded and never read your message.
Every long run cost about an hour. 366K, 343K, 316K, 350K, and 173K tokens across the five Claude Code tests, and time did not track tokens: the cheapest run still took an hour.
It is the slowest frontier model tested, roughly twice as slow as Kimi K3 or GPT 5.6, which the creator already considers slow.
It refuses to hallucinate on the frog test, and it gets six out of six brain scans wrong. Kimi K3 got one right.
Confidence intervals decide the benchmark story. Opus 5 is first on Artificial Analysis, Debates, and official Deep Sweep, third on LiveBench, and behind Fable 5 on Vals. On the two boards that publish intervals, the intervals overlap and the ranking dissolves.
ARC-AGI-3 is the one real outlier and the one under suspicion. Over 30 percent against under 10 percent for GPT-5.6, but an independent researcher found it does worse than Opus 4.8 on games with unfamiliar rules.
Cost sinks the recommendation, not capability. Roughly twice the price of a rival that is arguably as smart, about four times Kimi K3 on the Vals index for a few decimal points of gain.

Chapters

0:00:00 Opus 5 intro
0:01:02 Windows on browser
0:07:16 Image to 3D
0:09:08 Financial presentation video
0:11:56 Luma Agents
0:13:54 Blender X wing fighter
0:15:38 Music composition
0:20:55 FROG TEST
0:22:02 Identifying cancer
0:23:09 Deep research
0:24:43 Specs cost benchmarks
0:29:50 Guardrails
0:30:52 Verdict

Notable quotes

The orange sphincter is back. AI Search, opening the video, 0:00:00

So, it did all of this in just one prompt. I didn't even need to prompt it further. AI Search, on the browser native Windows 11 build, 0:01:02

Very basic music, but still really impressive how it was able to vibe code all these songs and play them in just an HTML file. AI Search, on the fake Spotify's playable albums, 0:01:02

None of the other frontier models including Fable or GPT 5.6 were able to accurately render this scene. AI Search, on why that geometric office image was chosen, 0:07:16

So, I got to say this is currently by far the best model at generating this 3D scene. AI Search, the strongest praise in the video, 0:07:16

Three are spending, one is collecting. So, the question isn't whether the money gets spent, it's who earns a return on it. the voiceover of the financial video Opus 5 researched, scripted, and rendered itself, 0:09:08

A cute little detail is it even added a place for R2-D2 over here. Exactly like the actual X-Wing fighter. AI Search, on the Blender model, 0:13:54

So Opus 5 is incredibly slow. Definitely the slowest of all the frontier models. AI Search, after the five minute song took over an hour, 0:15:38

The poor thing is trying so hard, but still it could not find anything. So, the frog remains safe. AI Search, the frog test, 0:20:55

At least one good thing about this is if it can't find anything, it doesn't just hallucinate and make stuff up. AI Search, the frog test, 0:20:55

Note that Fable 5 completely rejected my request. So, at least Opus 5 is more open to some biomedical prompts. AI Search, after Opus 5 scored zero out of six on tumor identification, 0:22:02

For Opus 5, I actually don't like it as much as GPT 5.6. It seems to be like holding back. AI Search, on deep research output, 0:23:09

The highest score was GPT-5.6, which they didn't even highlight in orange. They just grayed this out as if it was irrelevant. AI Search, on Anthropic's own benchmark card, 0:24:43

If the confidence intervals overlap, then it could mean that like the top five models are basically tied in terms of performance. AI Search, the objection running under the whole benchmark section, 0:24:43

In 100% of my personal workflows, I don't actually need to use Opus 5. AI Search, the verdict, 0:30:52

Resources mentioned

The model and the harness

Anthropic and claude.ai, the chat interface used for the frog, tumor, and deep research tests
Claude Code, the agentic harness used for the five long tests
Claude paid plans, required for Opus 5 access (the creator is on Max, not API credits)

Models compared against it

GPT 5.6, which tops the Deep Sweep card and LiveBench categories in the video
Kimi K3 by Moonshot AI, the price and value comparison throughout
Gemini and Gemini 3.1 Pro, including the Connections and LiveBench results
GLM 5.2, the open source model that hallucinates at half the Opus 5 rate
Claude Fable 5 and Claude Opus 4.8, both from Anthropic

Benchmarks and leaderboards

ARC Prize, home of the ARC-AGI-3 game learning benchmark
Artificial Analysis, source of the intelligence index, the speed table, the price table, and the Omniscience hallucination rate
LiveBench, where Opus 5 places third
Vals AI, the finance and coding index with published confidence intervals
New York Times Connections, the fun benchmark
The agentic software engineering benchmark the captions render as "Deep Sweep 1.1", both on Anthropic's card and in its own official results (see the footnote below on names)

Tools driven during the tests

Blender and the Blender MCP server
Waveform, the free DAW Opus 5 composed inside
Surge XT, the free synth it chose, and Vital, the one it ruled out because it cannot create accounts
HyperFrames, the open source motion graphics tool used for the presentation video
Gemini TTS, the voiceover engine
Chrome, where the Windows 11 clone ran

Apps and companies rebuilt or referenced

Microsoft 365 (Word, Excel, PowerPoint), the Microsoft Store, Discord, Slack, Spotify
Q4 filings from NVIDIA, Alphabet, Meta, and Amazon
X-wing starfighter and R2-D2, the reference for the Blender build

Sponsor

Luma AI and Luma Agents, including the Luma Skills feature demonstrated in the segment

The channel

AI Search on YouTube and the free weekly AI newsletter linked in the video description

An honest footnote

Two things about this page are worth stating plainly.

On the model names. This video was captured through automatic captions, which garble rival model names badly. The recurring competitor appears variously as "Kimiko 3", "Kimik3", "Kimika 3", "Kemetic 3", and "Chimera 3"; this page normalizes all of those to Kimi K3, which is what the creator is clearly saying and showing on screen. "JLM 5.2" is rendered here as GLM 5.2. Two references to a rival that the captions transcribe as "Gemini K 5.6", the LiveBench leader and the model described as almost half the price of Opus 5, are referred to here as "the Gemini entry" because the transcription cannot be resolved with confidence and inventing a version number would be worse than naming it loosely. The agentic coding benchmark the captions call "Deep Sweep 1.1" and "Deep Sweet" is left as spoken and deliberately not linked, for the same reason. Every number on this page is a number the video states out loud or shows on screen; none has been rounded, filled in, or inferred.

On the frame. This is one reviewer's hands on session, not a controlled evaluation. The demos are single runs at maximum reasoning with no repeats, the token figures are read off a usage meter shared with concurrent projects (which the creator flags himself each time), the deep research comparison is explicitly labeled subjective, and the closing criticism of Anthropic as a company is opinion rather than reporting. The strongest parts of the review are the parts that are checkable: the Excel formula engine either evaluates or it does not, six scans were either right or wrong, and the confidence interval argument applies to every board on the page whoever is winning. Those hold up regardless of what you make of the verdict.

Full transcript

The orange sphincter is back. Anthropic releases their latest and best model, Claude Opus 5. And this is an incredibly capable model. In this video, we're going to go over all the impressive things it can and cannot do. Plus, we're going to go over its specs, cost, and performance against other AI models, so you can decide whether it's worth using or not. Let's jump right in with some demos. Now, you could use Opus 5 just on their online chat platform, claude.ai, but frontier models like Opus or GPT or Kimiko 3 are actually designed for agentic coding and autonomously using different tools and doing long-horizon tasks that require multiple steps. So, to really test out its potential, it's best to use Opus with an agentic harness like Claude Code. So, that's mainly what I'm going to use in this video. Like other harnesses, this allows you to work with multiple files and folders locally on your computer, and so you can spin up and work with multiple projects at once. For my first test, let's try this. Create a browser-friendly replicate of Windows 11. Include common apps and programs like Microsoft Office, Microsoft Store. Make sure there are apps that I can actually download, photos, file explorer, media player, Discord, Slack, Spotify. Make sure all these programs actually work. Make sure it runs efficiently on a regular web browser. Let's drag this all the way to Ultracode. All right, so here's what it got. It first planned out everything. It's figuring out all these different Windows mechanisms like drag, resize, snap layouts, focus, etc. Now, it's building all the main components like taskbar, start menu, desktop. Then, it's building the apps, file explorer, and then Microsoft Office, Word, Excel with a real formula engine, PowerPoint, and then media player, Discord, Slack, Microsoft Store, and then all these utility apps, which we are going to test out in a second. It even gave me some games like Minesweeper, Snake, and then finally the app stylesheet, etc. And then, the really nice thing about Opus is, especially if you run it in Claude Code, is it can automatically spin the page up in a browser and then check for any errors. So, here it found a bug and it's automatically fixing this. Then it found some additional errors and then it's testing actual behavior, etc., etc. And after some further testing and verifying, it found another bug over here, but after some further testing and verifying, everything works. That's pretty much it. Let's open this up and see what we get. All right, so it even starts with a nice login window. Let's click anywhere and then press enter. First, let's play around with these settings over here. So, for example, let's turn on night light and it actually turns on night light. You can see the colors are a bit more yellow. Let's turn this off. Let's turn off dark mode and as you can see, indeed, the bar and these icons are now in light mode. Let me turn this back into dark mode. Let's also play around with brightness and interestingly, brightness also works. And this is just in a browser. If I click on this, then it does pull up a calendar just like the regular Windows 11. Let's click on this start menu and play around with all these different apps. Let's open this Word document and indeed, it pulls up Microsoft Word and I can edit stuff here. For example, let's type in some additional text. I can highlight this. Let's make this italics. Let's change the font color to red and then for these two lines, let's right align them and everything just works. And let's actually press Ctrl S to save it. So, if I exit out of this and then I pull up project proposal, it's actually saved. Really cool. Let me exit out of this and next, let's play around with Excel. So, this works. Let's type in a few numbers and then let's play around with some equations. Like for example, let's write equals average of A1 to A5 and it actually gives me the average. And next, let's try sum. So, I'm going to write equals and then sum. Again, A1 to A5, and indeed it gives me a sum. So, equations work, and then if I press control S, I can save this as whatever I want. Let's call this test, and then press save. And let's exit out of this. Now, if I open up my start menu again, I can see this test spreadsheet over here, and indeed this is the saved spreadsheet. Notice that this is just running on my Chrome browser. Pretty crazy. And actually, let me pull up this dummy spreadsheet that it made up. And as you can see, the numbers are actually accurate with the formula up here. Very cool. Next, let's open up PowerPoint. It already made a deck, so let's open that up. This one is not too impressive. I can't really drag the text boxes around. It's lacking a ton of capabilities that are found in PowerPoint, so this needs a bit of work. Let's exit out of this, and then next, let's open up Microsoft Store. And let's actually download Spotify. It's kind of simulating a download. Let's also download Discord, and Slack, and also Paint. Let's also get some sticky notes, and code editor, and let's also download Solitaire. All right, let me first open up Spotify, and it actually made a Spotify-looking layout with some songs. Let's see if these songs actually work. Let's try Focus Flow. >> [music] [music] >> These tracks sound really similar, but it did make a full album with different tracks. Let's also try something like garage rock. >> Very basic music, but still really impressive how it was able to vibe code all these songs and play them in just an HTML file. This is pretty crazy. All right, next let's exit out of Spotify and then let's open up Discord and indeed this kind of looks like Discord. In fact, let me type something here and see if anyone replies. And it actually simulates someone replying. Now, you can see it doesn't really make sense. It's not actually reading my messages. It's just randomly replying. Everything seems to be hardcoded. Let me exit out of this and next let's open up Slack. Again, this looks like Slack. Let me try messaging here and again it can simulate someone replying. Next, let's open up Paint and indeed we can draw stuff here. Let's also draw some shapes. Let's draw a circle next. Let's fill this pink and everything just works. Next, let's exit out of this and then for sticky notes it looks like this. Let's add a new note. This app works. Let's exit out of this and then let's try code editor. This also looks pretty good and then finally solitaire looks like this and it does look like this works. Finally, let's open up file explorer and indeed everything works. So, it did all of this in just one prompt. I didn't even need to prompt it further. All right, let me pull up the usage for this. So, this took 366,000 tokens and note that I'm running several concurrent projects at once. So, this project itself didn't take the whole 36% of my 5-hour limit. It probably took a quarter and this did run for a really long time like over an hour. Currently, I'm using the max plan. I'm not using any API credits. So, that's why down here it says usage credits is zero. All right, next let's see how good it is at creating 3D assets from just a reference image. So, I'm going to upload this really tricky geometric office scene and then ask it to create a beautiful 3D animated scene from this image. Make sure everything is faithful to the attached image. Use a single HTML file. None of the other frontier models including fable or GPT 5.6 were able to accurately render this scene. And again, I'm going to set this to ultra code and then press run. All right, here's what I got. It was first able to give me a decent and very basic scene, but it wasn't really faithful. It was missing a lot of details, so I wrote, "The desks and the other objects aren't arranged properly. Plus, there's too much glow. Make sure it looks exactly like this attached image." And then afterwards, that's all it took. In just two prompts, here is our results. Would you look at that? I'm just going to show you the image in the lower left corner, and as you can see, most of the desks and chairs are arranged properly, although there's an extra chair here. It was able to detect that there's a person here, but not on the other desks, and there's another person over here. There still are a ton of flaws with this render. For example, this panel here should be against the wall. The books on this shelf are also not consistent with the image. There should actually be two desks here. Still a lot of inconsistencies, but again, if you compare this with the generation from Fable, which looks like this, you can see that this one from Opus 5 is just way better. So, I got to say this is currently by far the best model at generating this 3D scene. Very impressive. All right, next let me also pull up my usage stats. So, here you can see it used 343,000 tokens. This also ran for a ridiculously long time, close to an hour. And again, I'm running several projects concurrently, so this didn't take the whole 48% of my 5-hour limit. This probably took like 15 or 20%. All right, next, because all these Frontier models are really good at tool use and working with different types of files and platforms, let's test this out with a really tricky prompt. "From last year's Q4 reports of Nvidia, Google, Meta, and Amazon, create a professional presentation video comparing their financials and future outlook. I didn't even link or attach the reports. It needs to go and find the Q4 reports itself from all these companies, and then it needs to, you know, figure out the financials and compare them and give me the future outlook. Include a voiceover from Gemini TTS. Use HyperFrames to create the animation. I linked to the HyperFrames GitHub repo. This is basically an open-source tool for you to create motion graphics presentation videos. It should be 16:9, around a minute long. Include charts, graphs, other visuals. Use this as the background audio. And then afterwards, I also pasted in some documentation on how to use Gemini TTS to generate the voiceover. And then at the bottom, I gave it my API key, which I will delete before I record this video. That's pretty much it. All right, so afterwards, it proceeded to search for the Q4 reports and then analyze the data. Then it proceeded to install HyperFrames and then also generate the voiceover using Gemini TTS and then finally generate the video. Again, the awesome thing about Opus 5 is that it can autonomously verify its generation and then fix any errors that it sees. And so afterwards, it rendered the video successfully and that's pretty much it. In just one prompt, I didn't even need to prompt it further. Here's our final result. >> Four companies, one quarter, and a half-trillion-dollar bet on artificial intelligence. Start with scale. Amazon booked $213 [music] billion, nearly double Alphabet, and more than Nvidia and Meta combined. But growth flips the board. Nvidia grew 73% [music] Amazon, the largest, grew the slowest. Profitability tells a third story. [music] Meta turns 41 cents of every dollar into operating profit. Amazon, under 12. Nvidia's gross margin, 75%. And underneath, it's one engine, Nvidia's data center. Google Cloud, AWS. AI demand now drives all four, which brings us to the bill. >> For 2026, Alphabet, Meta, and Amazon guided to roughly half a trillion dollars in capital spending. Nvidia guided 78 billion in one quarter. Three are spending, one is collecting. So, the question isn't whether the money gets spent, it's who earns a return on it. >> Pretty good. It's important to note that other frontier models like Kimik3 can also do this, but this video from Opus 5 seems a lot cleaner and smoother with fewer errors. Now, I forgot to pull up the usage stats, so here they are. This used 316,000 tokens. If you're constantly doing the same creative workflows over and over again, definitely check out Luma Agents by Luma AI, the sponsor of this video. It's going to save you so much time and make you way more productive. Think of Luma Agents as a team of AIs working together to autonomously carry out a creative process. You can simply upload references and prompt it with natural language to do whatever you want. For example, I can upload this image and get it to create an entire brand kit and color palette in a matter of minutes. The coolest part is their latest feature called Luma Skills. Think of it as a way to turn your best creative workflows into reusable AI tools that you can run repeatedly. I can simply prompt the agent to turn this workflow into a skill. For example, let's name this brand kit, and in the future, I can upload another image and use that skill to apply the exact same workflow as before to recreate the exact same result. Or here's another example. I can upload this photo of a woman in a dress and get Luma Agent to create a video of this person wearing this dress and walking down a runway in a fashion show. And if I want to reuse this exact same workflow as before, I can just prompt it to save this as a skill. For example, let's name this fashion show, and afterwards, upload another model or another piece of clothing and then use that skill to generate another fashion show video easily. What's really cool is you don't even have to build these skills yourself. You can just simply tell the agent what you want, and it'll generate the entire skill for you automatically. I'm especially impressed by the sharing capabilities. Once you've created a workflow that works well, you can instantly share it with teammates using a simple link. Entire teams can install the skill and get the exact same results without having to recreate the workflow themselves. You can even bundle multiple skills together into a package and distribute an entire creative toolkit with a single click. Whether you're creating marketing campaigns, editing videos, generating images, or building repeatable creative workflows, Luma skills in Luma Agents is an absolute game-changer. Try today using this QR code or the link in the description below. Now, because these frontier models are so good at using tools and different platforms, let's also see how good it is at creating 3D models in Blender. So, I'm going to write "Use Blender MCP at this local address." So, in Blender, I already installed and enabled the MCP at this local URL, and let's get it to create an X-Wing fighter spaceship with realistic texture and motion. Again, I'm going to set it to ultra code and then press enter. All right, so it's planning this out and generating this step-by-step while connected to Blender. In fact, let me just show you a time-lapse of this in action. You can see it's building the different components, first with the body and now the wings, and then it's also figuring out where the hinges should be so the wings can open. Pretty cool. And then afterwards, it's even automatically generating the video. The cool thing is it's able to take screenshots and verify that the model and the animation works. And if not, then it'll proceed to fix any issues. And so afterwards, it finished generating this X-Wing fighter, and that's pretty much it. So, let's open this up. Here is what the X-Wing fighter looks like. In fact, let me keep playing this. Very cool. You can see all the different components that it has generated. So, like here's the wing, here's each part of the wing. Very comprehensive. And then here's what the solid view looks like. Here's what the material looks like. The textures also look pretty good. It's just creating this from scratch. And you know, a cute little detail is it even added a place for R2-D2 over here. Exactly like the actual X-Wing fighter. Really impressive. And then let me show you the rendered animation, which looks like this. Really cool. All right, so let me pull up the stats for this. This actually took surprisingly very few tokens. Only 173K tokens, but it did take a long time. Again, this took close to an hour. All right, next let's see how good it is at making music. So here's my prompt. Your job is to compose an amazing award-winning song. Figure out which instruments you need to include in the song, then look for free VST plugins. These are like virtual instruments that are under 800 megabytes in size, then compose the song using these VST plugins in my Waveform DAW. So this is an existing free DAW that you can use to compose music. You can add additional effects like reverb, delay, risers, drops, include panning effects automation and make it sound amazing. The thing is it needs to decide which instruments to use and then search for and download these plugins itself and then load everything into my DAW and then compose the notes for each track, set up the automation, panning, etc. and compile everything into a full song. All right, so first it's figuring out where exactly my Waveform DAW even is. After it has confirmed that the app exists, then it proceeds to figure out the virtual instruments to download. So it searched for a few instruments. This Vital Synth plugin requires registration. It can't create accounts, so it ruled that out. And then it finally landed on Surge XT, etc. etc. And then it also asked me what type of song I want. It gave me a few options, so I just clicked on the recommended option, which is cinematic melodic techno. And then afterwards it also downloaded some drum kits, etc. etc. And then after a lot of downloading and planning, it's now arranging the notes for each individual track. Now this took really long. This took over an hour already. And then next it actually proceeded to render the song. And here's what we got. All right, so let me open this up in my DAW and holy smokes, look at how thorough and detailed this is. So, it has tracks for the kick, clap, hats, percussion, symbols, sub bass, bass, pad, air pad, and then strings, keys, arpeggiator, pluck, lead, counter melody, and then riser, and also, I guess, automation for reverb and delay. Pretty crazy. Let's play the song. >> All thing is 5 minutes long. It's pretty much the same chords. I won't bore you with the full song, but if you are interested in hearing the entire song, I'll link to it in the description below. But as you can see, it's able to actually control my DAW, even automatically download [music] virtual instruments, plan out the notes for each track, including drums, sub bass, bass, synth pads, strings, plucks, etc., and generate a full legit song that actually sounds pretty decent. And it's all from just one prompt. Very impressive. For your reference, this also took a crazy long time. This also took over an hour, so Opus 5 is incredibly slow. Definitely the slowest of all the frontier models, and this took 350,000 tokens. It probably took around like 20% of my 5-hour limit. All right, so those are some really complicated examples using Claude code, but of course you can also just use it on their online chat interface. So, let's test this out. First, let's try your favorite test, finding the frog. So, I'm going to upload this image. There's a frog hidden somewhere in this image, but I'm not even going to tell it that. I'm going to write, "Is there any animal or animals in this image? If so, identify and circle it." I'm going to set it to Opus 5 Max and then press generate. All right, so first what it did was it split the image into a 3 by 3 grid, then it's inspecting each grid, but it couldn't find anything. It suspects it's a potential snake species, so it's trying super hard to pinpoint where the snake is. It's trying to trace the snake's sinuous path. It tries to enhance the saturation and contrast to find the edges of the snake. The poor thing is trying so hard, but still it could not find anything. It's now trying to rescan each separate tile for potential animals, but still it could not find anything. So, the frog remains safe. Now, at least one good thing about this is if it can't find anything, it doesn't just hallucinate and make stuff up. It'll actually say it cannot find anything. All right, next let's see if it can identify cancer. So, I'm going to upload this image of six different scans, and there's a type of tumor in each of these six scans. For the prompt, I'm going to write, "Identify the types of tumors in each of the six images, if any." And then press run. All right, here is its answer. So, for the top left, it identified craniopharyngioma, which is not correct. And then for top middle, it did not identify a tumor, which is wrong. Top right, it was unsure, so it says it's likely to be this, which is not correct. It also said it could be this type of tumor, which is also not correct. So, next moving on to the bottom row, it says no discrete mass, which is wrong. Bottom middle, it identified acute intraparenchymal hematoma, which is also wrong. And then for bottom right, it said craniopharyngioma, which is also wrong. So, it got six out of six of these scans wrong. Whereas, actually Chimera 3 got one of these correct. Not really good in terms of tumor identification, but note that Fable 5 completely rejected my request. So, at least Opus 5 is more open to some biomedical prompts. Next, let's test its ability to do deep biomedical research. It's important to note that Fable is severely nerfed. It won't even answer this. So, the prompt is analyze the pathophysiology of atherosclerosis, etc., etc., include relevant tables and visualizations. And surprisingly for Opus 5, it actually answered the question. So, here is its findings, initiation and progression. It created a nice flowchart. It also coded up some nice diagrams here. And then here is a table of the two strategies compared. Here's also a nice bar graph. Now, it's important to note that all the frontier models can do deep research like this, but each of them have different vibes. For example, Gemini tends to be very verbose, whereas I really like the outputs from GPT 5.6 and Kemetic 3. They are very thorough, but concise. It's jam-packed with information. For Opus 5, I actually don't like it as much as GPT 5.6. It seems to be like holding back. It's not doing thorough enough research and the information just doesn't seem to be organized as well. This is very subjective. It depends on what you prefer. All right, so that sums up my tests on Opus 5. Overall, it's a very capable model, especially in terms of front end and 3D design, as you can see from my demos. But for a lot of cases, it's actually not noticeably better than 5.6 or even Kemetic 3. I would say it's a very close tie. It thinks for way longer. It's super slow. Like most of my tasks that I showed you in this video took over an hour. I feel like it's twice as slow as Kemetic 3 or GPT 5.6, which are already really slow. And also Opus 5 is much more expensive. In fact, let's go over the specs of this next. So, this is a frontier model, again designed for agentic workflows, so it can autonomously do long horizon tasks involving multiple steps and tool calls. It can work with different platforms. And like the other frontier models, it has a 1 million token context window. So, you can fit a ton of information into your prompt at once, roughly 700,000 words or a small to medium-sized codebase. Now, if you look at these self-reported benchmarks, then on average, it does perform better than their previous best model, Claude Fable 5. Note that for legal and health, then Fable 5 actually performs better. For biology, Opus 5 performs better. Now, one of the most important metrics here is Deep Sweep version 1.1. This is considered one of the most accurate measures of how good a model is at agentic software engineering, and as you can see, both Fable and Opus did not score the highest. The highest score was GPT-5.6, which they didn't even highlight in orange. They just grayed this out as if it was irrelevant. One of the most impressive things about Opus 5 is it scored 30% in this Arc AGI-3 benchmark. If you're not familiar with this, this is where the AI model is placed in this game environment, and it has to figure out how the game works by itself. It needs to learn the rules of the game and then how to win or get to the next stage, and also how to avoid losing. Now, for humans, it's pretty easy. Humans could easily score 100%, but for the top frontier models like Opus or Gemini or GPT-5.6, you can see they do pretty awful, and that's because, technically speaking, AI models can't learn new things. Their parameters in their neural network are fixed after training. So, this Arc AGI-3 ain't just testing if an AI can play video games. This is actually testing their emergent abilities to learn new things on the fly. And apparently, according to Anthropic, Opus 5 scores over 30%, way higher than the next best model, GPT-5.6, which is less than 10%. Now, while this might sound impressive, especially since the other frontier models are all the way down here, take this with a grain of salt, because there's new evidence suggesting that Opus might be benchmarked. In fact, this researcher found that when Opus 5 was given games that are similar in design to the Arc-AGI 3 games, it could complete them very well. When it was given some completely new games with unusual or non-conventional rules that it has to figure out on the fly, then it does pretty badly. It even performs worse than Opus 4.8. Now, if you look at this independent leaderboard by Artificial Analysis, then you can see that Opus 5 is ranked number one, just one point above Fable 5, which is one point above GPT 5.6, which is two points above Gemini K 3. The thing I don't like about Artificial Analysis is there's no confidence intervals. So, we don't actually know if these differences are statistically significant. If the confidence intervals overlap, then it could mean that like the top five models are basically tied in terms of performance. If you look at the speed of this, measured in output tokens per second, then Opus 5 is painfully slow, noticeably slower than GPT 5.6 and even Fable 5. And if you look at the price, Opus 5 is also way more expensive. It's almost twice as expensive as Gemini K 5.6 Solo, which you can argue is just as intelligent. So, in terms of performance versus cost, Opus 5 is not the most cost-effective option. In terms of this Omniscience hallucination rate, Opus 5 is not bad. It hallucinates roughly the same rate as Gemini K 3, but much higher than one of the leading open-source models, JLM 5.2, which hallucinates half as much. If you look at another independent leaderboard called LiveBench, you can see that Opus 5 is only ranked number three, whereas Gemini K 5.6 is number one. In fact, across the board, across all these categories, including reasoning, coding, mathematics, data analysis, language, and instruction following, it's not as good as Gemini K 5.6. Now, in terms of this Debates benchmark, which measures how well AI models perform in adversarial multi-turn debates across a wide range of topics, so this not only requires knowledge, but also the ability to stay responsive and defend over multiple rounds, you can see that Opus 5 is indeed ranked number one. In terms of solving New York Times Connections games, which I think is just a fun benchmark, it's not really a useful indicator of anything, you can see that Opus 5 is number two, slightly above GPT 5.6, but still behind Gemini 3.1 Pro, which has scored surprisingly well. In terms of this Valves index, which tests the model's performance across finance and coding tasks, then you can see that Opus 5 is actually behind Fable 5, and only a few decimal points above Kimika 3, but notice that it costs like four times more. And I like how they include confidence intervals here, so if you account for these intervals, then there's actually no significant difference between like Kimika and Claude Opus. For Deep Sweet, the official results are in, and apparently here, Opus 5 does rank number one. But again, if you look at the confidence intervals, there's actually no significant difference between like any of these models. Also, note the price of this, so Opus 5 is way more expensive than GPT 5.6, as well as Kimika 3. Finally, it's also important to mention the guardrails and nerfing of this. So, just like Fable 5, Opus 5 will also sometimes reject your prompt if it's related to cybersecurity or biology, and it's going to fall back to a dumber model, Opus 4.8. But here they do say that Opus 5 is less restrictive than Fable 5. So, they do allow Opus 5 to find vulnerabilities in source code, but it'll likely reject you if you ask about all this other stuff. And then for biology, it's also a lot more permissive. As you can see with my demos, it's actually willing to answer my question. However, it will still reject you if you get it to do some long-running autonomous research if they think it poses some substantial biology-related risks. Now, currently to use Opus 5, you do need a paid plan. It's not available on the free plan of Claude. Once you subscribe to a paid plan, you can use it on their online interface as well as in Claude code. Plus, it's also available via their API. All right, finally, here is my verdict. Do I recommend Opus 5? Well, I'll give credit where it's due. Opus 5 is a great model, especially in vibe coding and front end and 3D design. It makes the least amount of errors, but it's also extremely expensive and slow. In fact, in 100% of my personal workflows, I don't actually need to use Opus 5. I can just go with GPT 5.6 or Gemini 3 or even the much cheaper GLM 5.2 to handle all of my tasks. Plus, Anthropic is known for constantly gatekeeping, rug pulling, and secretly nerfing their models. They're also the most against open-source AI. They want to hog and control superintelligence for themselves, and they sure don't want you plebs having access to this power. They're also great at fear-mongering. They keep suggesting this AI doomsday narrative, which poses existential threats and economic disruption, and all this BS which makes me really not want to support Anthropic. So, personally, I would not recommend paying for Opus 5 unless you're working specifically with front-end development or 3D design, and you really need a good model for those use cases. Or if you're stuck on a coding problem that even GPT 5.6 or Gemini 3 cannot solve, then maybe you can try using Opus to solve it. Anyway, that sums up my review of Claude Opus 5. Let me know in the comments what you think of this and if you've had a chance to try it out, what were the vibes that you got from it so far? As always, I will be on the lookout for the top AI news and tools to share with you. So, if you enjoyed this video, remember to like, share, subscribe, and stay tuned for more content. Also, there's just so much happening in the world of AI every week. I can't possibly cover everything on my YouTube channel. So, to really stay up-to-date with all that's going on in AI, be sure to subscribe to my free weekly newsletter. The link to that will be in the description below. Thanks for watching and I'll see you in the next one.