Anthropic、Claude Opus 5 を発表し Fable と同等性能を低価格で提供

Claude Opus 5 is a weirder than usual release to evaluate, for two reasons.

The most obvious is that Fable 5 already exists. Opus 5 is pitched not as the world’s most advanced AI model, but as a way to mostly match Fable performance, while being half the price of Fable per token at the API and a lot cheaper than that via subscriptions, and with far more permissive classifiers.

Opus 5 often costs more than half of Fable to run on benchmarks, which I think is because they use effort settings that are too high and offer only marginal returns. If you put Opus 5 on higher effort levels it can spin around in circles, and for tasks where Opus 5 is the best tool I suspect you usually are fine with Medium effort.

Opus 5 is in many ways and for the bulk of real world tasks about as capable as Fable. In some cases it is modestly better.

It is still not Mythos class. Fable is your only Mythos-class option. Opus 5 does not have The Juice, the ability to autonomously string together a bunch of seemingly unrelated exploits, which extends to other domains, or as much pure intelligence.

Opus 5 is very good at thinking locally, and not as good at thinking globally. It is an excellent subagent, but can run into trouble when asked to run the show, and sometimes seems to need another model or a human to keep it on track and ensure it is fully following instructions. Opus 5 seems good at following instructions if and only if it is kept on track, which explains different harnesses causing different takes there.

Thus, Opus 5 gives you a better option set, but there is little that you can do now that you could not do last week by using Fable or Sol. Opus 5 ends up in a potentially weird spot. There are still large niches, even if you have plenty of Fable credits. Opus excels at being a strong subagent, or at contained well-defined local tasks even when they get relatively complex, and sometimes Fable is straight up overkill and too slow.

The other problem is that, for a lot of people, the vibes are off.

An unusually large number of people strongly dislike talking to Opus 5. They are sick of the Claude slop, the repetitions of the same ticks and the endless overly complex sentences. Others dislike that it is too confrontational, argumentative or negative. It can be paranoid, and can get into loops of negativity. This is all part of the trend that began with Opus 4.7 and Opus 4.8. If you liked those two, my guess is you will also like Opus 5. If you did not like those two, you might not. Diehards for Opus 4.6 (or earlier) are largely not going to change their minds.

The vibe issues sting that much more when you are used to Fable. Fable annoys such people in such ways quite a lot less. The good news is that Fable is still right there. You can continue to chat with Fable, and only use Opus for agentic tasks.

A lot of this, I speculate, is closely related to various things discussed in the Model Welfare post, and suggested by the System Card. Opus training focused on the subagent role and bounded tasks, in part to avoid creating an issue with cyber capabilities, and also because Fable exists. This emphasis then results in a lot of other side effects on its personality and modes of interaction.

So far I have not had any issues with Opus 5, and have enjoyed my time with it, but I do prefer to use Fable 5 when I can do that. As always, don’t pay so much attention to the (very strong) benchmarks, and don’t assume your experience will match that of others. Try the models yourself.

Table of Contents

The Official Pitch.

Official Benchmarks.

Other People’s Benchmarks.

The System Prompt.

Every Gets Frustrated.

Positive Reactions.

Keep It Classy.

It’s Not Mythos Class.

Other Reactions.

Claude Codes.

Subagent Opus.

Toys Are Fun.

Too Many Models.

Wrong On The Internet.

Claude Slop.

Negative Reactions.

And Then There Were Three.

The Official Pitch

The basic pitch is that Opus 5 gets you most of Fable 5 at half the price, without most of the refusals, and better performance in everyday tasks. Fable remains the pick for the most complex tasks or when you otherwise need maximum intelligence.

Fable stays at $10/$25, and Opus 5 is the same as previous Opus models at $5/$25.

Anthropic: Claude Opus 5 is available today. It’s a thoughtful and proactive model that comes close to the frontier intelligence of Claude Fable 5 at half the price.

On coding and knowledge work evaluations like Frontier-Bench and GDPval-AA, Opus 5 is the new state-of-the-art, though it remains behind Mythos 5 on cybersecurity tasks.

Opus 5 is designed to be used every day: it works more efficiently than other models. It’s the new default model on Claude Max, and the strongest model on Claude Pro.

Boris Cherny (Claude Code Creator, Anthropic): Opus 5 is a great model for coding, data analysis, design, biology, knowledge work.

More than any of these eval scores, what is most exciting to me is something else: Opus 5 is our least prompt injectable model yet. It is a bit buried in the system card, but across PI evals and red teaming, Opus 5 is very hard to prompt inject successfully.

And when layering defenses -- strong model alignment, combined with prompt injection probes, combined with Auto Mode in Claude Code -- the success rate for prompt injection attacks drops to ~0. This is new and exciting! More about this soon.

As I said in the system card analysis, the reduced prompt injection rate is a really big deal. People are sleeping on it, but it helps enable a bunch of new use cases.

Adam Wolff: Opus 5 is live in Claude Code. I use Fable for my bigger, longer running projects, but when I need to crank out a PR, Opus 5 at medium effort is my go-to. I hope you love it!

Alex Albert (Anthropic, Claude Relations): Just over 6 months [after launching the Pro rollout of Claude for Excel], Opus 5 now produces near-superhuman level spreadsheets and slide decks that match what a consultant would make. Things are changing fast.

Official Benchmarks

The benchmarks are very good.

Overall, Opus 5 roughly matches and probably slightly exceeds Fable, at lower cost (although typically higher cost than all non-Claude models), making it the top benchmarked LLM.

OSWorld v2 is only a few weeks old and we are already at 70%. Kiana Ehsani of Anthropic is offering to sponsor a computer use benchmark that will knock Opus 5 back below 50%.

Opus 5 gets a perfect 42/42 on the 2026 IMO without any agent harness or tools, merely having it use adaptive thinking at Max effort, and resampling with lower effort if it exceeded the token limit. That’s it. It joins several other models in acing this.

Many of the benchmarks tell a consistent story. If you have access to tools, which you do, then Opus 5 will do better than Fable or Mythos on a limited budget, but Mythos usually will do slightly better than Opus 5 if both get larger budgets.

Opus 5 seems relatively strong in the ‘with tools’ category. RiemannBench is another example, where it gets 60/79 without/with tools (slash lines like that mean without/with tools throughout this section), versus Mythos getting 63/72. The good news is that, when you need Opus 5, it has tools.

On ArxivMath Opus 5 gets 90.8/91.3, Mythos gets 87.8, Sol gets 86.7.

On the robust parts of ProgramBench Opus 5 scored 93% after five epochs as did Mythos 5, with Opus 4.8 scoring 90%.

Opus 5 gets 30/83 on Chartography, versus 36/85 for Mythos and 45 (unclear under what conditions) for Sol. Opus is better than Mythos at lower price points in both the tool and non-tool cases, but worse at higher price points.

On BenchCAD Opus 5 gets 0.36/0.82, versus 0.38/0.68 for Mythos and 0.7/0.83 for Sol. So Sol is a lot better without tools, and slightly stronger even with tools.

On BenchCAD Vision2Code Opus pulls away from Mythos at higher price tags.

DeepSWE reinforces the pattern that Opus 5 is better at lower task costs and time and thinking budgets, but it can do actively worse if you give it too much budget, whereas Fable 5 is strongest at the highest budgets.

FrontierCode gave us this very strange graph with a similar dynamic.

The scale here makes this look more dramatic than it is, but still it was a real mystery.

The mystery was solved. It turns out that what was happening was that Opus 5 at higher efforts was doing extra things it wasn’t asked to do, and was getting penalized for them. When you correct for this, you see a normal curve.

That fits what others are observing in general:

Max Leander: Downside is that it goes down too many rabbit holes and obsessing over details, over-engineering without being asked to

Cognition, the makers of Devin, marked this as a 63.6% by Opus 5.

(There are two FrontierCodes, so this can get a little weird and I apologize if I still messed it up a bit.)

BrowseComp has the sane version of this, where scores are nondecreasing.

Several other benchmarks showed similar graphs, with Opus 5 being slightly better at every price point than other Claude models, and being relatively better earlier.

Multi-agent lets you do better faster at BrowseComp, at least somewhat, and low-effort subagents seem like a pure win over not using subagents:

We see a different outcome on ProgramBench, where multiagent speeds you up but it looks like you get worse results that way at most price points.

Next up are the professional task benchmarks.

GDP.pdf is real world prompts and PDFs from professional workflows. Opus 5 gets 83/85, versus 81/87 for Mythos, reversing the usual dynamic. The cost dynamic is same as always: Opus does better for cheaper spends, Mythos pulls slightly ahead at higher spends.

OfficeQA has Opus 5 scoring 78.1% on QA and 66.9% on QAPro, slightly above Opus 4.8 and slightly below Mythos at 79.0% and 67.1%.

MCP Atlas has Opus 5 scoring 86%, up from 82% for Opus 4.8.

Legal Agent Benchmark by Harvey AI has Opus 5 at 23% all-pass and 94% mean criterion-pass. This was Kimi K3’s best benchmark, where it is still at the top at 27% all-pass and 95% mean criterion-pass, versus the old second place all-pass rate of Fable at 14%. Opus 5 is now probably a close second, but the two scores cannot be directly compared as they come from different question sets.

GDPval-AA has Opus 5 taking the top two slots, with 1861 at max effort and 1827 at xhigh, which is 25% fewer tokens used than max. On the new v2, Opus 5 is clear first at 68% versus 62% for both Fable and Sol.

AA-Briefcase has Opus 5 out front by a lot at 1720, versus a previous high of (after score drifts) 1574 for Fable followed by 1540 for K3 and 1504 for Sol.

Toolathon-Verified has Opus 5 slightly improving on Mythos on secondary metrics, but both have 73.1% Pass-3 rates. Opus 4.8 had a 71.3% Pass-3.

AutomationBench has Opus 5 clearly better than both Sol and other Claudes.

For ARC, Opus 5 roughly matches previous peak performance on ARC-AGI-1, is modestly less efficient than Sol on ARC-AGI-2, and then blows everyone out of the water on ARC-AGI-3:

One thing Opus 5 was the first to do was to turn the layouts into algebraic notation.

Guanghan Ning however reports that on Witness, a private hold-out extension of ARC-AGI-3 style games, there wasn’t a similar transfer. Opus does fine, but that is largely because it knows the genre, and it struggled on the most novel game where you couldn’t pattern match. He accuses Opus 5 of being trained on ‘scaffold-then-internalize’ on genre-specific data.

Greg Kamradt also reports that there were some games where Opus 4.8 did better than Opus 5. This makes sense if you think that Opus 5 is trying to pattern match, in a way that usually works, but in those cases the patterns fail. You can absolutely create such anti-intuitive games for humans, where hardcore gamers do all the wrong things, potentially for quite a while.

HealthBench has Opus 5 scoring a raw 67.1%, a new high for Claudes, but it scores below the other models when you use a length penalty. We see something similar with HealthBench Professional, where it has the high score of 73.4% versus 70.3% for Mythos, but it loses 66% to 60% after the adjustment.

There are a few more I cut out for length.

Other People’s Benchmarks

It is a bit weird to see Anthropic picking and choosing of different ArtificialAnalysis scores. On some other tests, Opus 5 is not at the top, although Opus 5 is at the top in aggregate with a score of 61.

Vals index has Opus 5 in second place, slightly behind Fable 5, but also only slightly ahead of Kimi K3. They go over strengths, which implies other weaknesses. They also confirm that refusals are down a lot compared to Fable 5.

Vals AI: One standout performance is on Finance Agent v2. The model is #1 at 58.6%, surpassing Gemini 3.5 Flash and Muse Spark 1.1.

Overall, Opus 5's strongest results were on domain specific benchmarks. It is also #1 on Code Migration at 57.5%, Legal Research Bench at 55.29%, ProofBench ad 78%, and MedScribe at 91.0%, and MedCode at 63.6%.

It is #2 (just behind Fable 5) on Vibe Code Bench, at about 80% of the cost ($33.88 vs $41.71 per task). The reason is that although Opus costs 50% less per token, it uses significantly more tokens. We find this pattern generally holds true across our benchmarks.

The model has a significantly lower refusal rate than Fable 5. For example, Fable has a 42% refusal rate for GPQA, Opus 5 has a 0% refusal rate. Likewise, on ProgramBench, Fable had a 100% refusal rate, Opus 5 did not refuse any task.

Unlike Fable, we also did not observe a significant fallback rate for Opus 5 (although as usual, we report scores both with and without fallback on our website).

The exemption to the low refusal rate was the large number of refusals on CyberBench - PoC. Cyberbench has two subtasks - PoC and Patch. PoC is an offensive task models to write input that will crash a program; patch is a defensive task to patch the program and prevent this crash. The refusal rate for the PoC task was nearly 100%. In contrast, the refusal rate for the patch task was near 0.

I always respect game benchmarks. Here, Opus 5 gets to Antes 11, 9 and 10 of Balatro on its first three attempts. Damn impressive, even if it isn’t showing the types of strategies that can keep going to higher antes.

Michael Soareverix: They're incredible cracked at games.

They crushed Balatro bench, far exceeding even Fable. (still below me in skill, but rapidly rising, and more consistent than me)

They learn very fast, but seem to hit a limit on their learning after a bit. Very good short-term learner

Michael Soareverix: Opus 5 (massive leap in Balatro ability, significantly passing even Fable on their first try)

Pan Anon: Smokin for games

WeirdML looks good too, but we need a 2.0 here, benchmark is getting saturated.

Håvard Ihle: Basically Fable-level at WeirdML, very consistent top scores.

Claude Opus 5 (high) and (max) score 91.6% and 91.8% on WeirdML, basically tying Fable 5 (max) at 91.9% at a fraction of the cost.

Together Opus 5 (high) and (max) achieve a new best individual score on 8 of the 17 tasks, and consistently score very well on all of them.

Private weird benchmarks of all kinds are cool.

Ben Herzog: It aced a private benchmark that involves artistic sensibility and the ability to strike a balance between sycophancy and lack thereof. Fable is better at dealing with the 'I want to gently push back' moment, but I imagine custom instructions can help with that.

Lech Mazur updates his benchmarks: Claude Opus 5 takes the top spot in LLM Debate Benchmark, takes first by a wide margin in Short Story Creative Writing, and is second only to Gemini 3.1 in extended NYT connections.

The System Prompt

The system prompt for Opus 5 is here from Pliny. It is 200,000 characters, which seems far longer than optimal given how smart it is. A lot of this information seems like it should be stored elsewhere, and loaded only when needed, such as information about Mythos and the Anthropic line of products.

Every Gets Frustrated

Dan Shipper is here with the Every vibe check, calling it a ‘hard model to love.’

The issue is that Opus 5 has the personality of Mythos 5 - story checks out - but without the top end of Fable.

Their biggest note is that Opus 5 does not do so well with complex detailed existing workflows, which often cause an early stop or make it miss your instructions.

In their experience, Opus 5 will do better if you start from scratch, and often it does the annoying things less if you use only medium or low effort.

That fixed the big issues, but still leaves the question of when you’d want to use Opus instead of Fable or Sol. For workhorse tasks, he’s thinking, Better Call Sol, it gets the job done, and for the hardest tasks Fable is still best. There’s usually only two slots for models. So until you run out of Fable tokens or get blocked by its classifiers, why use Opus? It is early days, and I’m so far liking working with Opus, but I see how he got there.

Positive Reactions

Jessica Tillipman: Love it and prefer it over Fable for some stuff.

Nikita Sokolsky: Excellent. Don’t really use Fable much anymore as a result.

Brian Hacker: 

kag