Claude Opus 5 Matches Fable 5's Peak at Half the Cost
In the early hours of July 25 Beijing time, Anthropic officially released Claude Opus 5.
Image source: Anthropic official release
Image source: Anthropic official release
It's now available on Claude.ai and Claude Code. Claude Max has set it as the default model, and Claude Pro users can also access the current strongest Opus.
Image source: Claude.ai
I skimmed through the official article and System Card. Benchmarks one after another, dazzling to the eye.
But in the end, you only need to remember one thing.
Claude Opus 5's capability is close to Fable 5, but at half the price.
It's a case of a small horse pulling a big cart.
Image source: Anthropic official release
Opus 5's API pricing is $5 per million input tokens and $25 per million output tokens, exactly the same as Opus 4.8.
Fable 5 is $10 and $50. Exactly double the price.
With the price halved, how much capability remains?
On CursorBench 3.2, Opus 5 at max effort is within 0.5% of Fable 5's peak, at half the cost per task.
Image source: Anthropic official release
Frontier-Bench v0.1 is even more impressive. Opus 5 achieves over twice the score of Opus 4.8 at a lower per-task cost.
Image source: Anthropic official release
Knowledge-based tasks are similar.
ARC-AGI 3 tests a model's problem-solving ability when facing new problems. Opus 5's score is three times that of the second-place model.
Image source: Anthropic official release
Zapier's AutomationBench tests whether a model can complete business tasks end-to-end. At the same task cost, Opus 5's pass rate is about 1.5 times that of the second-place model. Even at minimum effort, it completes more tasks than other models.
Image source: Anthropic official release
On OSWorld 2.0, Opus 5 also outperforms other models. It surpasses Fable 5's best score at slightly over one-third of the cost.
Image source: Anthropic official release
Of course, these are mostly official evaluations, and Frontier-Bench is Anthropic's internal test. How it actually performs in real projects remains to be seen through practical use, community reviews, or pressure from other vendors.
One thing I quite like about Opus 5 this time is that it figures out how to verify results on its own.
Official examples were given.
Frontier-Bench asked it to write code to reconstruct a 3D FreeCAD model based on a mechanical part drawing. The evaluation environment didn't give it a tool to directly view the image. Opus 5 wrote its own computer vision pipeline, extracted geometric shapes from raw pixels, and reconstructed the part. Under the same setup, it succeeded consecutively; other models failed all five attempts.
Another example involved a real bug in an open-source package manager. The community patch missed an edge case. Opus 5 found the root cause and patched it along the way. Another model only fixed the surface issue and then claimed the problem was solved.
Another example came from a trading firm. An engineer asked it to create a market data feed for a new exchange, but there was no real-time data available for verification on-site. Opus 5 simply built its own testing framework to check whether the code correctly parsed the exchange data.
It's no longer rare for models to run long-duration tasks. The fear is that after writing the code, it pats itself on the back and tells you the task is done, hallucinates, or just does a sloppy job and calls it a day.
But my impression of Opus 5 this time is that it knows how to clean up after itself.
Moreover, Anthropic specifically stated in the latest prompt documentation that previous instructions added for older models, like doing a final verification or finding a SubAgent to review, can be directly deleted.
Because Opus 5 checks itself. If you write it again, it might start checking endlessly...
Officially, two visual demos were also released, both well done.
The first is an interactive wind tunnel. You can adjust wind speed, rotate the car body, and see how airflow moves around different objects.
Image source: Anthropic official demo
The second is a 3D interactive animal cell. Clicking different structures shows explanations on the right, and you can slice it open to see the interior.
Image source: Anthropic official demo
Both demos are directly interactive, showing a fairly high level of completion.
Everyone has been competing on front-end generation recently, and Opus 5 has kept up this time.
Next, some parameters beyond price.
Opus 5 has a context window of 1M tokens and a maximum output of 128k tokens. thinking is enabled by default, with effort levels from low, medium, high, xhigh, up to max.
Use low and medium for ordinary tasks, and xhigh or max for difficult problems.
It also offers Fast mode, which is about 2.5 times the speed of the default mode, and the price is exactly doubled.
2.5 times the speed sounds great, but double the price is indeed expensive.
When developers call it via API, the model name is claude-opus-5.
Image source: Anthropic official release
Claude Code lead Boris Cherny also specifically mentioned: Opus 5 is currently the Claude model least susceptible to prompt injection.
Image source: Boris Cherny
Some netizens also ran a visual effect comparison between Opus 5, Fable 5, GPT-5.6, and Kimi 3.
From this image, Opus 5's effect is the most realistic. Of course, this is just a netizen's test, so take it with a grain of salt.
Image source: Netizen test
Anthropic also mentioned safety.
In automated behavior audits, Opus 5's misaligned behavior score is 2.3, the lowest among recent Claude models.
Simply put, this version of Opus is more rule-abiding, exhibits less deceptive behavior, and is harder to induce into high-risk operations.
Image source: Anthropic official release
Image source: Anthropic official release
Regarding cybersecurity, Anthropic clearly stated they did not specifically train Opus 5 on related tasks.
However, Opus 5's vulnerability-finding capability is already close to Mythos 5. But when it comes to finding development vulnerabilities or weaponizing them into exploits, the gap is still large.
Image source: Anthropic official release
Anyway, the model is online now. No amount of hype matters. I'm considering whether to give Claude Code another try.
Fable 5 was released just over a month ago, and Anthropic has already produced a half-price Opus 5, directly making it the default model in Claude Max.
Now the pressure is on OpenAI.
By the way, should GPT-5.6 Sol also get a reset?
Godfather Tibo, where are you?
Sources: