Claude Opus 5 in Practice: Can Fable 5-Level Performance Become an Everyday Work Model?
Anthropic has announced Claude Opus 5, its newest Claude model. It became available on July 25, 2026, Japan time.
Anthropic describes it as a model that comes close to its flagship Claude Fable 5 while keeping the same price as Opus 4.8. Benchmark numbers alone do not tell us whether a model is useful in everyday work, though. This article separates the official announcement from my own experience after asking Opus 5 to build a website.
Opus 5 is more than a cheaper Fable 5
Claude Opus 5 costs $5 per million input tokens and $25 per million output tokens through the API. Those are the same rates as Opus 4.8, and half the $10/$50 rates of Fable 5.
Anthropic says Opus 5 is state of the art on coding and knowledge-work evaluations. On CursorBench, the company reports that Opus 5 at maximum effort comes within 0.5% of Fable 5’s peak score while costing half as much per task.
These are evaluations published by Anthropic, so they should not be read as proof that Opus 5 beats Fable 5 at every job. Fable 5 remains the model Anthropic positions for the hardest long-running autonomous tasks.
My current interpretation is that Opus 5 is not simply a less expensive model below Fable 5. It is a strong candidate for the model you use every day when a task deserves careful thinking.
What the official specifications tell us
- The Claude API model ID is
claude-opus-5 - Input at $5 and output at $25 per million tokens, unchanged from Opus 4.8
- A 1M-token context window and up to 128k output tokens
- Effort levels from low, medium, high, xhigh, and max
- The default model on Claude Max and the strongest model available on Claude Pro
- A Fast mode that runs around 2.5 times faster at twice the base price
Adjustable effort makes it possible to use less computation for a simple writing task and more for design or debugging. Keeping the same model while changing the amount of thinking to match the job is practical.
There are also migration details to keep in mind. Web Fetch is not available on Opus 5, and Priority Tier is not supported. When using the API, it is important to check tool availability as well as the price.
The ability to read between the lines
Here is what stood out when I used Opus 5 myself.
It is good at inferring intent from a short instruction. Instead of following only the literal wording, it tries to understand what I am ultimately trying to accomplish and returns the questions or suggestions that would help.
It is not just a matter of deeper reasoning. The model is attentive to the person it is working with. It explains points where I might get stuck and lays out options before moving ahead, which makes the pace of the conversation feel surprisingly human.
It is not right every time, and important changes still need review. But when it makes a mistake, it tends to explain its assumptions and propose a correction. That felt like a meaningful improvement over previous Opus models.
What happened when I put it in charge of a website
I happened to have a website I wanted to build, so I asked Opus 5 to handle the initial construction. In about half a day, it reached a state close to publication, and I could continue with small corrections and follow-up instructions.
The most impressive part was that it did not immediately start writing code. It first organized the necessary features and the user’s journey, then created a wireframe and suggested a layout based on how the feature would be used.
Fable 5 and GPT-5.6 SOL can also produce excellent final results. Without detailed instructions about the process, however, they tend to push ahead quite aggressively. Opus 5 naturally inserted a cycle of checking before building, verification during the work, and self-correction when necessary.
That personality works for both Spec-Driven Development, where requirements are settled before implementation, and vibe coding, where an idea takes shape through conversation. Its ability to suggest a direction should be especially useful when the finished form of a site or tool is still unclear.
How I would use it alongside Fable 5 and GPT-5.6 SOL
My current impression is that Fable 5 is for making a difficult job run all the way to completion, while GPT-5.6 SOL is for moving several workstreams forward and bringing them to a deliverable. Opus 5 sits between those roles as a design partner you can consult while it thinks.
One possible division of labor looks like this:
- The requirements are still ambiguous and I want to decide the direction through conversation: Opus 5
- The specification is settled and I want a long task completed end to end: Fable 5 or GPT-5.6 SOL
- I want to find omissions and risks before implementation: Opus 5
- I want to change many files and run tests in one pass: the agent model that best fits the work environment
Model strengths also depend on the tools and instructions around them. Rather than making Opus 5 do everything, it is realistic to use Claude to refine the design and Codex or another agent to implement it.
Start with work that has no single correct answer
To try Opus 5, I would give it a task that requires several decisions rather than a short factual question. A site structure, an improvement to existing code, a long document to organize, or a proposal that compares multiple directions are good candidates.
Instead of saying “make everything good” and handing over the entire job, give it the goal and constraints first. Watch the plan it proposes, then decide what to delegate and what to review. That makes the strengths of Opus 5 easier to see.
Can Fable 5-level performance become an everyday work model?
Claude Opus 5 has arrived with benchmark results that approach Fable 5 while keeping the same price as Opus 4.8.
After using it, what stayed with me was not only the raw capability but the way it moves through a task: checking, proposing, and verifying as it goes. If Fable 5 feels too expensive or excessive for every task, while lighter models feel unreliable, Opus 5 could become the practical default in between.
This is still a first-day impression, so long-term consistency needs more testing. For now, the best experiment is to choose one task from your own work that involves extended thinking and several decisions, then see how much responsibility Opus 5 can carry.
