Claude Opus 5: the same headline price, and two changes that break integrations

What changed

Anthropic released Claude Opus 5 on 24 July 2026 with the API model ID claude-opus-5. Headline pricing is unchanged from Claude Opus 4.8 at 5 USD per million input tokens and 25 USD per million output tokens. The context window is 1M tokens as both default and maximum, with no smaller variant, and the maximum output is 128k tokens. Anthropic's documentation records two changes it describes as breaking for existing integrations: extended thinking is now on by default, and setting thinking to disabled with effort xhigh or max returns a 400 error. The minimum cacheable prompt length drops from 1,024 tokens to 512. Claude Opus 4.8 remains available on every platform Opus 5 launched on.

Why it matters for UK business

The commercial framing of this release is unusually simple, and that is what makes it worth a briefing rather than a note. When a new frontier model arrives at the same price as the one it succeeds, the usual reason to defer a migration disappears. What remains is engineering risk, and Anthropic has documented exactly where it sits.

Take the two breaking changes in order. Thinking being on by default is the one that will surprise a finance team rather than an engineering team. Thinking tokens draw on the same output allowance as response text, so a workload that previously ran with no thinking will now produce more output tokens to do the same job, at the same price per token. That is a bill change arriving through a behaviour change rather than a price change, which is the second time in two months that pattern has appeared in Anthropic's release notes.

The second change is narrower and sharper. If your integration explicitly disables thinking, and you also set effort to xhigh or max, the request returns a 400 error rather than degrading gracefully. That is a hard failure, and it is the kind that shows up in production rather than in a smoke test, because most test suites do not exercise the top two effort levels.

The context window deserves a more careful reading than it usually gets. A 1M token window as standard removes a specific architectural constraint: it makes it possible to put a very large document set in front of the model in one request. It does not remove the reasons retrieval exists. Cost still scales with what you send, latency still scales with what you send, and precision on a focused set of retrieved passages is a different property from capacity. Treat the window as one fewer thing to engineer around, not as a replacement for the retrieval layer.

The change most likely to save money is also the least discussed. Cutting the minimum cacheable prompt from 1,024 tokens to 512 brings shorter system prompts inside the caching threshold for the first time. For firms running a high volume of small, repetitive requests, that is a direct reduction in input cost on work that previously could not be cached at all.

Sitting underneath all of this is a deadline that is not about Opus 5 at all. Claude Opus 4.1 was deprecated on 5 June 2026 and retires from the Claude API on 5 August 2026, and requests to retired models fail rather than falling back. Anthropic names Opus 4.8 as the replacement. If you have an integration pinned to claude-opus-4-1-20250805, that is the urgent item on this page, and it has a date on it.

What to do, and what not to do

Do:

• Check today whether any integration pins claude-opus-4-1-20250805. That is the 5 August item, and it fails hard rather than degrading.
• Grep your codebase for explicit thinking configuration before switching model IDs. The combination of thinking disabled with xhigh or max effort is the documented breaking case.
• Re-measure output token consumption after migrating, not before. Thinking on by default changes the output profile of unchanged prompts, and the price per token will not tell you that.
• Revisit prompt caching if your system prompts sit between 512 and 1,024 tokens. They are now cacheable and were not before.
• Confirm which platform you actually consume Claude through. Anthropic's lifecycle dates apply to Anthropic-operated platforms; Amazon Bedrock and Google Cloud publish their own retirement schedules.

Do not:

• Tell your team that Opus 4.8 is going away. It is not, and conflating it with the Opus 4.1 retirement creates unnecessary migration work in the same fortnight as necessary migration work.
• Say "no price rise" without qualifying it. Headline per-token pricing is unchanged; Fast mode is a research preview on the Claude API at twice base rate.
• Rip out your retrieval layer because the context window is now 1M tokens. Cost and latency still scale with what you send.
• Assume high effort is the default everywhere. Anthropic documents that default for the Claude API and Claude Code specifically.
• Migrate a production integration in the same change as a prompt rewrite. If the output moves, you will not know which change caused it.

Where The AI Consultancy fits

Model migrations are usually small pieces of work that go wrong in predictable places: a pinned model ID nobody owns, a configuration flag set two years ago, and a test suite that never exercised the path that now errors. Working through that on a live integration is what our Claude implementation engagements cover, and the AI App Production Clinic exists for the case where the integration was built quickly and nobody is quite sure what it does now. Our briefing on Claude Opus 4.8 covers the previous generation, and connecting Claude to business systems covers the integration layer this sits on.

All figures verified on 27 July 2026 against anthropic.com and platform.claude.com. Prices are US dollar list rates; Anthropic does not publish a sterling price list. Anthropic's comparison of Claude Opus 5 to Claude Fable 5 is the company's own launch framing and is not an assessment by The AI Consultancy. Model lifecycle dates apply to Anthropic-operated platforms; confirm your own platform's schedule.