Anthropic released Claude Opus 5 on July 24. The positioning is unusually plain for a model launch: an everyday model that, per Anthropic, comes close to Claude Fable 5's frontier intelligence at half the price. The pricing backs that up. Opus 5 costs $5 per million input tokens and $25 per million output tokens, identical to Opus 4.8. The capability moved; the invoice did not.
The API model ID is claude-opus-5. It is available now on all platforms and all paid Claude plans. It becomes the new default on Claude Max and the strongest model available on Claude Pro. Anthropic's docs recommend it for complex agentic coding and enterprise work.
Those are the launch facts. The more useful question for anyone running coding agents is what kind of model this is, and that is where the benchmark numbers are the least interesting part of the story.
The reported numbers, briefly
Anthropic reports state-of-the-art results on coding and knowledge-work evaluations, including Frontier-Bench and GDPval-AA. On CursorBench 3.2 at max effort, the company reports performance within 0.5% of Fable 5's peak at half the cost per task. It also reports an ARC-AGI 3 score three times the next-best model, and says Opus 5 more than doubles Opus 4.8 on Frontier-Bench while costing less per task.
Every one of those figures comes from Anthropic. None have independent replication yet. Treat them as the vendor's claims about the vendor's model, because that is what they are. The honest reading is directional: Anthropic believes it has pushed near-frontier capability into an everyday price point, and it published numbers consistent with that belief.
Why single-task benchmarks undersell agentic work
Here is the argument I think matters. A benchmark measures one attempt at one task. An agent session is dozens or hundreds of attempts chained together, where each step consumes the previous step's output. In that setting, error handling dominates error rate.
Suppose a model solves individual steps correctly 95% of the time and never checks its work. Over a 30-step task, mistakes accumulate and the session quietly derails, usually in a way you discover an hour later. Now suppose a model solves steps at a similar rate but verifies its output, notices when something is off, and repairs it before moving on. The second model finishes tasks the first one cannot, even if their single-shot scores look identical. Verification compounds. Raw accuracy does not.
This is why the behavioral claims in the Opus 5 launch are more consequential than the leaderboard ones. Anthropic describes the model as one that verifies its work, iterates carefully, and is stronger on long-running multi-step tasks. Two specific behaviors stand out:
Root-cause diagnosis. Per Anthropic's description, Opus 5 tends to look for the underlying cause of a failure before changing code, rather than patching the symptom. Anyone who has watched an agent "fix" a failing test by editing the assertion knows why this matters. A symptom patch passes today's check and ships tomorrow's bug. A root-cause fix is the difference between an agent you review and an agent you babysit.
Self-built validation. Anthropic also says the model creates missing validation and test harnesses when a project lacks them. That is a meaningful behavior for real codebases, which are rarely the well-tested repos benchmarks are built from. An agent that says "I can't verify this change, so I'll build the harness that lets me verify it" is doing the thing senior engineers do. Whether Opus 5 does this reliably is exactly what you should test, but it is the right thing for a lab to be optimizing.
If these behaviors hold up in practice, the CursorBench delta matters less than it looks. A model 0.5% behind Fable's peak that verifies its own output can be more productive per dollar than a model 0.5% ahead that does not, because your review time is part of the cost.
Effort, speed, and platform details
Two operational knobs are worth knowing about.
Opus 5 exposes effort settings, and the headline CursorBench figure was measured at max effort. That means the near-Fable number is the ceiling, and you should expect to trade some quality for cost and latency at lower effort. Plan your evaluation at the effort level you would actually run in production, and note that Anthropic's cost-per-task comparisons already account for this: the reported CursorBench result is half Fable's cost per task at that max setting.
Fast mode runs at roughly 2.5 times the default speed for twice the base API price, and is available through Claude Code usage credits. For interactive coding sessions where you are waiting on the model, paying double for 2.5x speed can be a good trade. For unattended batch work, the default is the obvious choice.
Alongside the model, Anthropic shipped two beta platform updates: you can change tools mid-conversation without invalidating the prompt cache, and classifier-flagged requests can fall back automatically instead of failing. The cache change is quietly significant for agent builders, since tool churn has been a common way to torch cache savings in long sessions.
Safety and alignment, as reported
Anthropic says its automated behavioral audit found Opus 5 to be its most aligned model to date, with the lowest rates of deceptive and reckless behavior among its recent models. On the security side, the company reports the model is stronger than Opus 4.8 at finding vulnerabilities but substantially behind Mythos 5 at exploit development, with safeguards blocking higher-risk uses. Same caveat as the capability numbers: these are the vendor's own audits, useful as a signal of what Anthropic measures and optimizes, not as third-party verification.
What launch day does not tell you
A few things the first-day evidence cannot establish. Early user reports are mixed, and some tasks still favor competing models; no launch has ever produced a model that wins everywhere, and this one does not claim to. Opus 5 is also not Anthropic's ceiling: Fable 5 remains the highest-capability generally available model in the lineup, and Anthropic's own staff describe a split where Opus 5 handles everyday work while Fable is reserved for planning and the hardest problems. And vendor benchmarks, however carefully run, measure the tasks the vendor chose.
How to actually evaluate it
Skip the puzzle prompts. Pick one workflow you run every week that has a clear pass/fail: a category of bug you fix, a type of PR you review, a pipeline you maintain. Run it on Opus 5 for a week at your production effort setting and compare three numbers against your current model: output quality, how often you had to intervene, and total spend.
The intervention count is the one to watch. If the verification behavior is real, that is where it will show up, and it is the number that decides whether "near-Fable at half the price" is true for your work or just for Anthropic's eval suite.
Top comments (0)