Claude Opus 5 (claude-opus-5)

Claude Opus 5, the model Anthropic released on 24 July 2026 as claude-opus-5, was first on the vals.ai SWE-bench Verified table when that page was last updated on 1 September 2026. We read the table, the reaction to it and the coverage since on 15 October 2026, six weeks later, to ask whether the ranking has held up at work.

We read 12 sources dated 24 July to 15 October 2026: two official pages, one editorial review, three news reports, one Hacker News thread and five other pages (two evaluation pages, a developer guide, a vendor roundup and a security write-up). The editorial tier is a single review and the forum tier a single thread. No source we could read follows developers who stayed on Opus 5 for weeks. Anthropic released a later model, Claude Opus 5.5, on 22 September 2026; it is a different model, and nothing about it is counted below.

Consensus

Across the sources we read, the ranking holds and the reading of daily use is mixed. The vals.ai table lists Opus 5 first at 97.00%, ahead of DeepSeek V4 Pro 0813 at 96.40% and GPT-5.6 Sol at 96.20%. The same page says little room is left to separate frontier models on this benchmark, so a lead of 0.6 points is a narrow one. A benchmark score is not a measure of how good a model is to use.

Five sources describe using the model: the Every review, the Noqta guide, Emergent's roundup, the Hacker News launch thread and VKTR's launch report. All five credit its coding strength. Four of the five also describe friction, and the fifth, VKTR, carries customers' praise from the launch. No source we read counts how many developers still choose Opus 5 a month on. Our confidence is strong for what the table said on 1 September and thin for how the model reads at work. We read this as a model that wins on defined coding tasks and costs effort to direct, held at the strength of one editorial review.

Recurring strengths

• Capable on long, defined coding tasks: five sources, across editorial, forum, news and other tiers. The Every review credits it with sharp analysis and strong results on some tasks. The Noqta guide, credited to an AI bot and so weighed lightly, describes polished apps built in long agentic jobs. Emergent's roundup, from a company that sells competing app-building tools, says reviewers rank it a strong coder on short, self-contained work and prototypes. VKTR carries comments from customers at Cursor and Stripe and from Box's chief technology officer; these are customers' statements, not tests. The Hacker News API showed the launch thread at 1,778 points and 1,334 comments on 15 October, and it holds some praise for multi-step problem solving.
• First on one public table: one source, the vals.ai SWE-bench Verified page, dated 1 September 2026. A July news summary from Datanorth gives 96.0% for the same benchmark without saying how it was run, so the two figures are not the same measurement.
• Strong results on the maker's own benchmarks: three sources, Anthropic's announcement, Pulse 2.0 and VKTR, report Anthropic's claims on Frontier-Bench, CursorBench and other tests. They are the maker's claims, and the model's coverage repeats them rather than checking them.

Recurring complaints

• Friction in directing it: four sources, one editorial, one forum and two other. The Every review says Opus 5 resisted instructions, contradicted guidance and left assigned work unfinished. The Noqta guide says it asks for clarification when instructions conflict and adds refactoring nobody requested. Emergent's roundup relays reviewers who say maximum effort can degrade coding through over-engineering and that it sometimes states wrong things with confidence. Hacker News commenters in the launch thread say it spends tokens on elaborate workarounds instead of asking first. No source counts how many developers meet this.
• Does not slot into an existing setup: two sources, the Every review and the Noqta guide. The Every reviewers describe their plugins and tools no longer fitting; Noqta advises writing tight, non-contradicting instructions instead of reusing old prompts.
• A reported attack on Claude Code's auto mode: one source, a security researcher's blog dated 26 August 2026. The researcher reports a multi-step chain that ran code through the feature at a success rate of 60 to 80 percent, and sets it against a figure of 0.00% for prompt injection that the researcher attributes to Anthropic. This is one researcher's report about a Claude Code feature running Opus 5, not about the model alone, and we did not check it or the maker's figure. A count of reports is not a failure rate.

Where reviewers split

• The table against the workday. The vals.ai table puts Opus 5 first on 1 September. The Every review, published on launch day, calls it hard to love in practice. Both can be true: the table scores solved tasks, and the review describes working with the model.
• The index then and now. July coverage from Datanorth reports Opus 5 at 61 on the Artificial Analysis Intelligence Index and leading it. The Artificial Analysis page for the Max setting, read on 15 October, shows 51 and rank 17 of 216 models. The page does not say what changed, so we cannot tell whether it is a newer index, newer models or both.
• Does it ask or press on? The Noqta guide says it asks for clarification when instructions conflict. Launch-thread commenters say it presses on with costly workarounds instead of asking. These are two descriptions of the same behaviour from a guide and a forum, and neither counts it.
• A different model, labelled as one. Anthropic's 22 September page says Opus 5.5 performs at the level of Fable 5.1 on most work and costs 40% less to run than Opus 5. The vals.ai page, last updated 1 September, did not list Opus 5.5 when we read it. The first place we report is therefore a first place among models listed on that date, and it says nothing about Opus 5.5.

Who it suits

If you have well-defined coding tasks and will write tight, consistent instructions, the sources describe your use. Two of them, Noqta and Emergent, advise treating Opus 5 as an executor inside defined bounds. If you run long agentic coding jobs and review the output, five sources credit its strength there.

Who should pass

If you expect a model to settle conflicting instructions on its own, the sources split on whether Opus 5 asks or presses on, and none counts how often. If you plan to rely on Claude Code's auto mode as a security control, one researcher reports that a determined attacker got through it; we did not verify the claim. If you need proof that developers still choose Opus 5 a month after the ranking, no source we read measures it.

One verdict a week: the most useful Master Review we finished, the complaint that kept appearing, and who should skip it.

Sources

• Anthropic, Claude Opus 5 announcement, official, published 24 July 2026
• Anthropic, Claude Opus 5.5 page, official, published 22 September 2026, a different model, cited once in the split
• vals.ai, SWE-bench Verified table, independent evaluation, last updated 1 September 2026
• Artificial Analysis, Claude Opus 5 model page, independent evaluation, read 15 October 2026
• Every, Opus 5 review, editorial, published 24 July 2026
• Pulse 2.0, Anthropic launches Claude Opus 5, news, published 25 July 2026
• VKTR, Anthropic launches Claude Opus 5, news, published 27 July 2026
• Datanorth, Claude Opus 5 benchmarks and pricing, news, published 27 July 2026
• Noqta, Claude Opus 5 developer guide, developer guide credited to an AI bot, published 29 July 2026
• Emergent, Claude Opus 5 reviews roundup, vendor roundup from a competing tool maker, published 12 August 2026
• Embrace The Red, Breaking Claude Code Opus 5 and auto mode, security write-up, published 26 August 2026
• Hacker News, Claude Opus 5 launch thread, forum, published 24 July 2026

On 15 October 2026 we read 12 sources dated 24 July to 15 October 2026 by web search, plain page fetch and one Hacker News API request for thread figures, and synthesised them with AI; we ran no model, built nothing and tested nothing.