Claude Opus 5 Brings Near-Frontier Coding Down to Team Budgets
Anthropic holds pricing flat, roughly doubles agentic performance, and makes self-verification the headline feature.
On 24 July 2026 Anthropic released Claude Opus 5, the newest model in its Opus family, available the same day on the Claude API, Claude.ai, Claude Code and Claude Cowork. Benchmark tables take up most of the announcement, but for a small development team two other details matter more: the price did not move, and the company's central claim is about the model checking its own work before reporting back.
The price is the easy part. Opus 5 costs $5 per million input tokens and $25 per million output tokens, which Anthropic states is the same as Opus 4.8, its predecessor. A capability increase that arrives without a matching rise in the invoice is worth stating plainly: every cost estimate a team built around the previous model still holds.
The performance figures Anthropic published are large. On Frontier-Bench v0.1 the company reported Opus 5 more than doubling Opus 4.8's performance at lower cost. On CursorBench 3.2 it landed within 0.5% of Fable 5's peak score at half the cost, and on the OSWorld 2.0 computer-use benchmark it passed Fable 5's best result at just over a third of the cost. On ARC-AGI 3 Anthropic reported a score three times as high as the next-best model. All of those numbers are the vendor's own, on the vendor's choice of benchmark.
The feature underneath the numbers
Anthropic's framing is that the model is much better at checking what it has produced and reworking it until the task is genuinely done. That reads like marketing language until you look at the worked examples. In one, the model was asked to write code rebuilding a machine part as a 3D FreeCAD model, but was deliberately given no way to look at the drawing it was working from; according to Anthropic it responded by writing a computer vision pipeline of its own to recover the geometry from raw pixels. In another, with no live feed available to validate against, it assembled a test harness to confirm its code parsed an exchange's data correctly.
This matters because the real cost of machine-written code has never been the tokens. It is the review. A model that produces plausible work and then stops leaves a person to find out whether it is correct, and on anything non-trivial that check costs more than the generation did. A model that runs the tests, reads the failure and fixes it before handing over moves where the human hours go.
The testimonials Anthropic published describe the same shift in different words. Denis Shiryaev of JetBrains put the difference down to judgment, saying the model catches its own logical faults while planning rather than afterwards, and reasons about why an answer is right rather than only whether it runs. Richard Pham, an evals and product lead, reported that on hard financial-modeling tasks the model averaged nine percentage points higher accuracy than Opus 4.8 across effort levels, with a third fewer turns and tool calls. A staff engineer, Cristian Rivera, described handing it a chief-of-staff role over his development environments for a weekend and being pulled in only for the judgment calls.
What this means if you are building something
Give the work a way to check itself. The advantage only appears when the model has something objective to test against: a test suite, a script that reproduces the bug, a staging environment it can call. A task whose only judge is a person reading the output gets none of it. Before handing over a piece of work, ask what would tell the model it had failed. If there is no answer, that is the thing to build first.
Restructure around longer runs. Work that used to be worth a few minutes of supervised prompting may now be worth queueing as a single instruction: migrate the schema and make the tests pass, fix the failing build, port this component. Batch those and review the result rather than the process.
Use the effort setting deliberately. Anthropic documents an effort setting that lets customers optimize for intelligence or conserve tokens for cheaper, faster results. Opus 5 is also offered in Fast mode, at around 2.5 times the default speed for twice the base price, the same arrangement that applied to Opus 4.8. Fast mode is for interactive work where somebody is waiting. For an unattended overnight job, paying double for speed nobody is watching is money thrown away.
Do not drop the review. Anthropic's automated behavioral audit scored the model at 2.3 for overall misaligned behavior, which it calls the lowest of its recent models, though it does not explain the scale. On cybersecurity it puts the model close to Mythos 5 at identifying vulnerabilities and considerably less successful at developing exploits. A low score is not a zero, and none of this is an external audit. Anything touching payments, authentication or personal data still gets read by a person before it ships.
Our reading: flat pricing is the headline that gets the attention, but self-verification is the part that changes how a small team plans its week. Treat the benchmark figures as claims to test on your own codebase, not settled results. The first change worth making is not switching models; it is making sure every task you hand over carries its own definition of done.
Sources
