Claude Opus 5: What Actually Changed, and Who It's For

Anthropic shipped Claude Opus 5 on July 24, 2026. The price doesn't move a cent from Opus 4.8, and the software engineering score more than doubles. Here's what that covers in practice, what the model does worse, and the two things to check before you switch.

What Anthropic announced

The model is live in the Claude app, in Claude Code and through the developer API as claude-opus-5. It becomes the default model for Max subscribers and the most capable model available to Pro subscribers.

One note on the pricing unit before the numbers: models bill per token — the chunk of text they work with. Reckon on roughly 750,000 English words per million tokens; the count is less favorable in most other languages.

Figures announced by Anthropic on July 24, 2026. These are the vendor's own measurements, published as charts rather than as numeric scores in the text. Nobody has reproduced them so far.

Anthropic sums up the positioning this way: the model comes close to the frontier intelligence of its top-end model, Claude Fable 5, at half the price.

The sticker price didn't drop — the cost per task did

This is the most misread part of the release, and it's worth slowing down on. The listed price is identical to Opus 4.8's: nothing dropped there. What changed is the price of getting a piece of work finished.

Look at how the announcement is phrased: Anthropic almost never compares its models per million tokens, but per completed task. On CursorBench, Opus 5 lands within half a point of Fable 5's score "at half the cost per task." On OSWorld, it beats Fable 5's best result "at just over a third of the cost." On Zapier's automation benchmark, it passes roughly 1.5× as many tasks "at the same cost."

The mechanism is simple: a model that fails less retries less. A task that lands first time is structurally cheaper than one run three times, even when every token bills the same. It's also why the sticker price tells you very little about the actual bill.

Two caveats. The model writes longer responses than its predecessors, which pushes the other way on the final number. And fast mode, which runs about 2.5× faster, bills at double — $10 and $50 per million tokens — and only through Anthropic's own API, not Bedrock, Google Cloud or Microsoft Foundry.

What it does better

Long-horizon agentic coding. This is where the gap is widest: multi-file tasks, larger refactors, features carried end to end. The model finishes the job instead of leaving stubs behind. On small single-file edits, by contrast, the gap with the previous generation is much narrower — don't expect a miracle on short tasks.

Code review. A high rate of real bugs found, with few false positives. One useful quirk: it stays accurate even at reduced effort, which makes a cheap fast pass before a thorough one a practical pattern.

Images. It reads charts, documents and diagrams better, and reproduces an interface from a screenshot more faithfully. It works up to 2,576 pixels on the long edge, and the coordinates it returns map to the image's real pixels, with no scale math to do.

Long context. One million tokens, as both the default and the maximum. Instruction following and tool calling stay solid across the full window.

Office work. Multi-sheet Excel files with non-trivial formulas, decks that follow slide-design conventions. It can stick to a template or a house style when one is imposed.

Coordinating multiple agents. Few cases of agents overwriting each other's work, and effective use of writer-verifier patterns.

What it does worse

Anthropic is explicit on two counts, which is worth crediting.

The model stays behind Mythos 5 on cybersecurity, particularly on exploiting vulnerabilities. It also shows important limitations on long-running autonomous research tasks — precisely the ground you'd be tempted to send it to.

Two behaviors also need tuning, and they run against habits picked up on earlier generations:

• It verifies its own work. Instructions like "double-check before answering" or "add a verification step," which helped on previous models, now produce over-verification. They need deleting, not rewriting. That inverts a prompt-writing rule that's usually sound.
• It expands scope. It sometimes adds steps nobody asked for, or applies its own judgment about what the task should be without saying so. An explicit instruction to deliver what was asked, at the scope intended, fixes it.

Unlike Opus 4.8, which was reluctant to delegate, this one reaches for subagents readily — enough that a cap is worth setting if cost matters.

Two gotchas before switching from Opus 4.8

These are the only changes that break existing code.

Thinking is on by default. On Opus 4.8, a request that asked for nothing didn't think. On Opus 5, it does. Since the max output length covers thinking and final text together, a request sized tightly on the old generation can now get cut off mid-answer. Worth revisiting on every route that never configured it.

Turning thinking off isn't allowed everywhere. Combining "thinking disabled" with the highest effort levels is rejected outright. Either leave thinking on, or step effort down a notch. And since the model is unusually strong at low and medium effort, the second option is often the better one: that's the main cost and latency lever of this generation.

Two pleasant details in passing. The minimum for caching a prompt drops from 1,024 to 512 tokens, which makes prompts cacheable that previously weren't. And Opus 5's rate limits sit in a separate bucket from the Opus 4.x pool: switching neither frees the old quota nor inherits it.

Who it's a real step up for

If you run coding agents on long tasks, this is the most interesting release in a while, and the gap widens exactly where previous models ran out of steam.

If you were paying for Fable 5 on work Opus 5 can handle, halving the cost is worth a serious test — keeping Fable 5 for what genuinely needs the top tier.

If you're on Opus 4.8 with short, well-scoped tasks, the upgrade is nearly transparent: same price, same calls, two gotchas to check. The difference in results will be modest.

And if your subject is cybersecurity or long-running autonomous research, the announcement itself says this isn't the right tool.