Claude Opus 5: what it means for the software you run

Last updated: 27 July 2026

Anthropic released Claude Opus 5 on 24 July 2026, at the same price as the model it replaces and installed as the default on its own top consumer tier [1]. Coverage has concentrated on benchmarks and pricing. The change likely to reach your codebase first sits in the developer documentation rather than the launch post: Anthropic has reversed its guidance on how hard to make the model work on your code. This article continues our industry news coverage: practical takes on what a release means for the software you run.

What actually changed

Anthropic states that Claude Opus 5 is "available today on all platforms, priced at $5 per million input tokens and $25 per million output tokens" [1]. That is the same rate as Claude Opus 4.8 [2]. It is also "the new default model on Claude Max, and the strongest model on Claude Pro" [1], and developers reach it as claude-opus-5 on the Claude API [1].

One capability result deserves separating from the rest, because an independent body produced it. The ARC Prize Foundation, which designs and administers the ARC-AGI series, reports that "Claude Opus 5 (High) is the highest-performing model on ARC-AGI-3, scoring 30.2%", and marks the result verified [3]. Hold on to the "(High)" inside that model name. It matters more than the score.

The dial Anthropic has just told you to turn down

Claude models expose a setting called effort that governs how many tokens the model spends on a response. It is not new to this release: Anthropic's documentation lists it as supported on ten models going back to Claude Opus 4.5, among them Sonnet 5 when it landed four weeks ago [4]. Opus 5 accepts five values, low, medium, high, xhigh and max, and defaults to high [4].

What is new is the advice. For Claude Opus 4.7 and 4.8, Anthropic told engineers to push the dial up: "Start with xhigh for coding and agentic use cases" [4]. For Claude Opus 5, the same page says close to the opposite. "Start with high, the default", and use low and medium "liberally as your primary control for token cost and response time wherever your evals show quality holds" [4].

The page adds a warning worth more attention than it has had: "If you carried effort settings over from an earlier model, run a fresh effort sweep on your evals rather than reusing them" [4]. In plain terms, a given setting does not mean the same thing on this model as it did on the last. Any team that upgraded by changing a model string in a config file has inherited a behaviour change nobody decided to make.

Anthropic documents what moves when the dial moves. Lower effort levels "combine multiple operations into fewer tool calls", "make fewer tool calls" and "proceed directly to action without preamble". Higher levels "explain the plan before taking action", "provide detailed summaries of changes" and "include more comprehensive code comments" [4].

Why "we built it with Opus 5" tells a buyer nothing

The clearest evidence of what the dial is worth comes from an evaluation Anthropic did not run. Artificial Analysis states that "all evaluations are conducted independently by Artificial Analysis"; its GDPval-AA benchmark puts anonymised outputs into blind pairwise comparisons and aggregates them into an Elo rating anchored to a human baseline of 1,000 [5].

Claude Opus 5 appears on that leaderboard five times, once per effort level. Max effort ranks first at 1,861. Xhigh ranks second at 1,827. High, the default, ranks fourth at 1,741. Medium ranks eighth at 1,632, and low ranks twenty-fifth at 1,454 [5]. One model, one benchmark, twenty-four places between its best and worst configuration.

That spread is set by a developer, per request, and it does not announce itself in the output. Anthropic is explicit: "Effort controls thinking volume, not visible response length: on Claude Opus 5, changing effort does not reliably shorten responses" [4]. So a reviewer cannot look at a function and infer how hard the model worked on it. Even the independent verifier ran into this constraint from the other direction. ARC Prize notes that "due to the short testing window, ARC-AGI-3 was evaluated only at High reasoning effort" [3], which means the verified 30.2% describes one rung of a five-rung ladder.

Nor is higher simply better. CodeRabbit, which sells an AI code-review product and has an interest in the conclusion, published a comparison on release day reporting that "more reasoning did not consistently produce a better review" [6]. Against its production baseline, the x-high configuration caught fewer known issues, 55.2% against 61.1%, flagged what it did find more precisely, 39.3% actionable precision against 35.2%, and produced roughly four times as many low-value comments, 92 against 23 [6]. Turning the dial up traded recall for precision rather than improving the review.

This is what makes provenance harder than it looks. When a founder, an acquirer or a new engineering lead is told a product was "built with Claude Opus 5", they have learned the model name and nothing about the care taken. The same sentence covers a codebase generated at max, with plans explained and changes summarised, and one generated at low, with fewer tool calls and no preamble. In our audits the artefact is the only honest witness, which is why a review examines what the code does rather than how it was produced, why technical due diligence on an AI-built product cannot rest on a vendor name in a data room, and why the defects that survive a clean test run repay a senior engineer's time.

An alignment claim is not a code-quality claim

Anthropic makes a strong safety claim for this model. It says Opus 5 "adheres to Claude's Constitution better than Opus 4.8, Sonnet 5, or Fable 5; exhibits the lowest rates of deceptive behavior; and is the least susceptible to being tricked into misuse" [1]. The supporting evidence is its own: "during pre-deployment testing, our automated behavioral audit found Opus 5 to be our most aligned model to date", scoring "2.3 on overall misaligned behavior, the lowest of our recent models" [1].

We applied the same test to OpenAI's launch figures three weeks ago, and it applies here. This is a vendor grading itself on an evaluation it designed. Zvi Mowshowitz, writing independently the day after release, drew the distinction precisely: "All we can say for sure is that it has the highest scores on the automated alignment tests. That is a very different fact about the world" [7].

There is a narrower point that matters more for engineering teams. Even taken entirely at face value, an alignment score describes the model's disposition, not the quality of any particular run. A model that is harder to trick into misuse is not thereby a model that wrote correct authorisation logic in your repository at medium effort on a Tuesday. Those are different claims about different things, and only one of them can be settled by reading your code.

More teams on the top tier, sooner

The economics point at adoption rather than value. Anthropic's top tier did not get more expensive: it costs $5 and $25 per million tokens, exactly what Opus 4.8 cost [1] [2], and it is now the default on Claude Max and the strongest model on Claude Pro [1].

A frontier model at last month's price, switched on by default, is one more teams will use without a procurement conversation or a deliberate decision. More of the code reaching your repository this quarter will come from it than from its predecessor, written at whatever effort level somebody set, or did not set. That is a shift in volume and in variance at once, and it does not answer the question a capable model never answers: whether what it produced is safe to put in front of your customers.

Frequently asked questions

Is Claude Opus 5 good for coding?

The strongest independent result is from the ARC Prize Foundation, which verified Claude Opus 5 at 30.2% on ARC-AGI-3, the highest of any model it has tested [3]. That figure was measured at High effort only, because of a short testing window [3]. Anthropic's own coding benchmark figures are self-reported and not independently administered.

What are Claude Opus 5's effort levels?

Claude Opus 5 supports five: low, medium, high, xhigh and max. The API default is high, and Anthropic states that setting effort to high "produces exactly the same behavior as omitting the effort parameter entirely" [4]. Effort is a per-request setting, so it can differ between one call and the next.

Does the effort setting change the quality of the code?

Yes, measurably. On Artificial Analysis's independently run GDPval-AA leaderboard, the five effort levels of Claude Opus 5 occupy ranks one, two, four, eight and twenty-five [5]. Higher is not uniformly better: CodeRabbit found x-high traded review coverage for precision [6]. The setting is recorded nowhere in the resulting code.

How is Claude Opus 5 different from Claude Opus 4.8?

Both cost $5 and $25 per million tokens [2], so the change is capability and guidance rather than price. Anthropic's coding advice reversed: 4.8 told engineers to start at xhigh, while Opus 5 tells them to start at high and use lower levels liberally where evals show quality holds [4].

Is Claude Opus 5 safe to use for production code?

A more capable model is not a self-certifying one. Anthropic's "most aligned model to date" claim rests on its own automated audit [1], and describes the model's disposition rather than the correctness of a given run. Treat its output as work from a fast, unfamiliar contractor: useful, and read by a senior engineer before it ships.

How much does Claude Opus 5 cost?

Claude Opus 5 is "priced at $5 per million input tokens and $25 per million output tokens" [1], unchanged from Claude Opus 4.8 [2]. Because effort governs how many tokens the model spends, the same task can cost materially different amounts at different effort levels, which is the trade-off Anthropic's documentation asks teams to measure on their own evals [4].

Get a senior read on what your effort setting actually shipped

A dial that moves a model twenty-four places on an independent leaderboard, leaves no trace in the output, and now carries the opposite vendor advice from the model you upgraded from is not an argument against Claude Opus 5. It is a genuinely strong engineering tool. It is an argument for knowing what has landed in your codebase rather than inferring it from the model name on the invoice. Our Vibe Code Audit puts a named senior engineer against your software and tells you what is safe to ship and what needs work first. Book an audit.

Sources

1. Anthropic, "Introducing Claude Opus 5", 24 July 2026. https://www.anthropic.com/news/claude-opus-5
2. Anthropic, "Pricing", Claude developer documentation, accessed 27 July 2026. https://platform.claude.com/docs/en/about-claude/pricing
3. ARC Prize Foundation, "Claude Opus 5 results", 24 July 2026. https://arcprize.org/results/anthropic-claude-opus-5
4. Anthropic, "Effort", Claude developer documentation, accessed 27 July 2026. https://platform.claude.com/docs/en/build-with-claude/effort
5. Artificial Analysis, "GDPval-AA", accessed 27 July 2026. https://artificialanalysis.ai/evaluations/gdpval-aa
6. CodeRabbit, "Opus 5 model review", 24 July 2026. https://www.coderabbit.ai/blog/opus-5-model-review
7. Zvi Mowshowitz, "Claude Opus 5: The System Card", Don't Worry About the Vase, 25 July 2026. https://thezvi.substack.com/p/claude-opus-5-the-system-card