Claude Opus 5 Checks Its Own Work. Here's When to Use It Over Sonnet.

Ask most AI models to fix something and they'll tell you it's fixed. Whether it actually is remains your problem. Claude Opus 5, which Anthropic released on 24 July 2026, is the first one we've used that reliably goes and looks before it answers.

That sounds like a small thing. It isn't. The gap between "I changed the file" and "I checked the page and the change is live" is where most of the time goes when you work with these tools. Opus 5 closes it, and Anthropic is confident enough about that to tell developers to delete the verification instructions from their existing prompts.

This is what actually changed, what the benchmark numbers mean once you translate them out of benchmark-speak, and where the line falls between using Opus 5 and using the cheaper, faster Sonnet 5. If you want the previous instalment, we covered whether Claude Fable 5 was worth its premium in June.

What actually changed in Claude Opus 5

Opus 5 slots in above Claude Opus 4.8 at exactly the same price. That's the first thing worth registering, because the usual pattern is a better model at a higher number. Here it's a better model at the old number, and Anthropic positions it as roughly Fable 5 intelligence for half of Fable 5 money.

Two API changes will bite if you're upgrading rather than starting fresh. Thinking is now on by default, where Opus 4.8 ran without it unless you asked. Since max_tokens caps thinking and visible response together, a request tuned tightly around its answer on 4.8 can now run out of room and truncate. And disabling thinking is only permitted at effort high or below, so this combination returns a 400:

{
  "model": "claude-opus-5",
  "thinking": { "type": "disabled" },
  "output_config": { "effort": "xhigh" }
}

The rest is upside. The prompt cache minimum dropped from 1,024 tokens to 512, so prompts that were previously too short to cache now cache with no code change. The context window is 1M tokens as both the default and the maximum. And the knowledge cutoff is May 2026, which is four months fresher than Sonnet 5 or Fable 5. For anything touching recent events, that difference shows.

It checks its own work before saying it is done

Anthropic's documentation puts this plainly: Claude Opus 5 "verifies its own work without being told to". What that looks like in practice is the model running the thing it just changed, rather than describing what it changed and stopping.

We've watched it do all three of the behaviours that matter, running as the default model in Claude Code on this site. It runs checks nobody asked for. It catches its own mistakes partway through and corrects them. And it won't call a job finished until it has evidence, which sometimes means telling you the thing still doesn't work.

Notice what the check caught. The file edit worked. The syntax was valid. Both of the obvious signals said success. The change simply wasn't reaching the page, because the function had been written but never called. That's the class of error that survives every check except actually looking at the output, and it's the one that used to cost half an hour of "it's still not working" before anyone thought to question the model's report.

The phrase that stuck with me there is "before handing the work back". I've lost count of the mornings that started with a client message saying the checkout button had vanished on mobile, and the honest answer was that nobody looked at the page on a phone. A model that opens its own work at two viewport widths, unprompted, is doing the step that human teams skip under deadline. It didn't need to be clever to catch that bug. It needed to bother looking, which is exactly the part people stop doing when they're tired.

There's a practical consequence, and it runs the opposite way to normal prompt advice. If your system prompts tell the model to verify, remove that instruction. Anthropic's own guidance is that these now stack on top of behaviour the model performs anyway, and that taking them out "reduces wasted tokens with no loss in quality".

When Claude Opus 5 asks instead of acting

The second behaviour is harder to demo and more useful day to day. Opus 5 stops and asks when the request is ambiguous in a way that matters, and it flags actions that can't easily be undone before it takes them.

On this site that has meant being questioned before a full FTP deploy and a CDN purge, which is exactly the right place to be questioned. The site sits behind an edge cache with a 30-day TTL. Deploy the wrong thing and you don't just have a bad page, you have a bad page pinned at the edge until someone purges it.

The minify catch in that example is the useful one. Nothing about it is dangerous. It's just a step in our build that's easy to skip, and skipping it produces a deploy that looks completely successful while changing nothing a visitor sees. Being asked about it costs three seconds. Finding out a week later costs considerably more.

The more interesting case is when it disagrees with you outright.

We wrote about that exact scenario recently, because a client with domains dating to the early 1990s really does block every crawler in robots.txt while expecting to appear in AI answers. A model that flags the contradiction before writing the rule is worth having.

"It didn't fold when I insisted" is the part I keep coming back to. Every one of these tools has caved the moment I pushed, and I'd stopped noticing how much that costs. If a model agrees with whatever you said last, its agreement carries no information, and you end up doing all the judging yourself while it does the typing. Wang's description is of something better: it held the position, narrowed the disagreement to one specific question rather than restating everything, and then found a version that kept his idea. That's not stubbornness. That's what a good colleague does, and I hadn't realised I'd been missing it until I read someone else describe getting it back.

If it asks more often than you'd like, that's tunable. Anthropic publishes a scope instruction that tells the model to make routine calls itself and check in only where different readings would produce materially different work. Drop that in and the interruptions fall away while the caution on irreversible actions stays.

The Claude Opus 5 benchmarks, in plain English

Anthropic published a wall of benchmark results with the launch. Most of them mean nothing without translation, so here's what each one actually tests and whether you should care.

ARC-AGI 3 is the one worth dwelling on, partly because the jump is enormous and partly because the number underneath it is humbling.

Twenty times better than the model it replaces sounds like the story. The more honest story is the green bar. A puzzle that any human solves without effort is one these systems fail roughly seven times in ten. Both facts are true at once, and anyone selling you the first without the second is selling something.

One number from the launch is more useful than any of the benchmark scores. Niko Grupen, Head of Applied Research at one of Anthropic's early-access customers, reported similar quality from Opus 5 while generating "26% fewer tokens on average compared to Opus 4.8 at max reasoning". Per-token price is not per-task price. A model that reaches the same answer with a quarter fewer tokens is cheaper in practice even at identical rates.

Claude Opus 5 vs Sonnet 5: when to use which

Sonnet 5 costs $3 and $15 per million tokens, currently discounted to $2 and $10 until 31 August 2026. That's about 40% below Opus 5 either way, and it's faster. For a lot of work it's the right answer, and reaching for Opus by reflex is how people run up bills they don't need.

There's a second dial most people ignore. Opus 5 exposes five effort levels from low to max, defaulting to high, and Anthropic is explicit that low and medium now produce strong quality at a fraction of the tokens. Before you conclude Opus is too expensive, try it a rung or two down in Claude Code. That's usually a bigger saving than switching model, and you keep the checking behaviour you came for.

If you're choosing between vendors rather than between Claude tiers, our sister site's Claude vs GPT vs Gemini comparison covers that ground, and the pattern it describes still holds: each vendor leads on a different axis.

What Claude Opus 5 does not fix

Self-verification is not the same as being right, and independent testing shows where the edges are.

CodeRabbit ran Opus 5 against roughly 100 verified error patterns from real open-source pull requests. Opus 5 at xhigh effort caught 55.2% of known issues where their production baseline caught 61.1%. Its comments were cleaner when they landed, with actionable precision of 39.3% against 35.2%. But it produced 92 nitpicks where the baseline produced 23, and it used noticeably more tokens. Their conclusion was that Opus 5 "may become a useful precision-oriented lane inside a routed ensemble", which is a long way from "best at everything".

That result has a documented cause worth knowing. If your review prompt says "only report high-severity issues" or "be conservative", Opus 5 follows that instruction literally and reports less. Measured recall falls even though the model found the bugs. The fix is to ask it to report everything with a confidence and severity rating, then filter in a separate pass.

Three other rough edges. Default responses run longer than on Opus 4.8, and lowering effort doesn't reliably shorten them, so you have to ask for brevity in the prompt. It narrates more during agentic work. And it delegates to subagents more eagerly than it should, which multiplies cost on tasks it could have finished itself, so cap that if your harness supports subagents.

None of this is fatal. All of it is worth knowing before you rewire a pipeline around the new model.

What a verifying model means for your website

Here's the part that connects a coding model to the work this site normally covers.

A model that verifies before it answers behaves differently towards your website than one that doesn't. The old failure mode was an AI system reading one page, forming a confident impression, and repeating it. The new behaviour is closer to cross-checking: reading what you say about yourself in one place, then testing it against what it can find elsewhere. That's the same instinct as opening the page at two viewport widths.

Consistency stops being a nicety at that point. If your homepage says one thing, your llms.txt says something slightly different, and your structured data says a third thing, a checking model notices. It won't necessarily conclude you're lying. It may simply conclude it can't safely cite you, which produces the same outcome: you're left out of the answer. We set out why the inputs matter more than the outputs in the AI Visibility Checking definition, and the argument gets stronger as the models get more careful.

The knowledge cutoff makes the same point from another angle. Opus 5 knows the world up to May 2026 and nothing after. Everything since has to be fetched from your site, in whatever form your site provides it. That's what files like identity.json and llms.txt exist to supply: a stated, machine-readable answer to who you are, rather than an inference drawn from your navigation menu. Our quick-start guide covers the full set in an afternoon, and the directory lists sites that have done it.

We measured the other half of this last quarter. AI Discovery File adoption reached 9.4% of top websites in Q3 2026, which means roughly nine sites in ten are still asking a model that now checks its sources to take an unverified guess about them. As the models get better at checking, ambiguity stops being neutral. It becomes the reason you don't get cited.

Frequently asked questions