Anthropic releases Claude Opus 4.7 with 1.0-1.35x tokenizer overhead

• The linked launch post says Claude Opus 4.7 is generally available across Claude, the API, Amazon Bedrock, Google Cloud Vertex AI, and Microsoft Foundry at the same list price as 4.6, $5 per million input tokens and $25 per million output tokens.
• According to the linked launch post, the migration catch is not price but token accounting: Anthropic says the updated tokenizer can turn the same input into roughly 1.0 to 1.35 times more tokens.
• The HN discussion roundup and Anthropic's migration guide both point to a second cost lever, deeper default effort on harder agentic runs, including xhigh as the new Claude Code default discussed by commenters.
• The HN discussion roundup also surfaced an API behavior change that is easy to miss in the launch copy: Opus 4.7 uses adaptive thinking, and human-readable reasoning summaries are omitted unless you request display: "summarized".
• In the linked launch post, Anthropic also says Opus 4.7 shipped with new cyber safeguards and is intentionally less broadly capable than the limited-release Claude Mythos Preview.

Anthropic's announcement bundles the obvious stuff, availability and benchmark gains, with a few migration gotchas hidden lower in the page. The migration guide spells out that manual thinking budgets are gone, while the main HN thread quickly zeroed in on xhigh, missing reasoning summaries, and higher effective token spend.

Anthropic positions Opus 4.7 as a direct Opus 4.6 upgrade for advanced software engineering, stricter instruction following, and longer agentic runs in the official announcement. The model slug is claude-opus-4-7, and day-one availability spans Claude surfaces plus Bedrock, Vertex AI, and Foundry.

The launch post also adds two product notes that got less attention than the coding claims: higher-resolution vision, and Anthropic's statement that Opus 4.7 is still below Mythos Preview in broad capability. That makes this a flagship GA model, but not Anthropic's top internal system.

Anthropic's migration section says two changes affect real token usage even though list pricing stayed flat in the announcement:

• Updated tokenizer: the same input may map to about 1.0 to 1.35 times more tokens.
• Higher effort on hard agentic work: Opus 4.7 spends more thinking at higher effort levels, especially on later turns.
• New control surface: the docs shift developers from manual budget_tokens style control to an effort setting in output_config, as described in the migration guide.

The HN discussion roundup adds the practical read from early users: Claude Code appears to default to xhigh in the terminal workflow, which compounds the tokenizer change for long coding sessions.

The biggest API migration gotcha is that Opus 4.7 treats adaptive thinking as the only supported thinking mode. Anthropic's migration guide says manual extended thinking with thinking: {type: "enabled", budget_tokens: N} now returns a 400 error, budget_tokens has no direct replacement, and thinking.display defaults to "omitted".

The HN discussion roundup captured the first developer reaction faster than the launch copy did: reasoning summaries that were previously visible now disappear unless requests opt into display: "summarized". The same discussion also flagged stricter instruction following, which lines up with Anthropic's note that prompts tuned for earlier models may behave differently.

Buried below the benchmark section, Anthropic says Opus 4.7 is the first model it is shipping with new automatic detection and blocking for prohibited or high-risk cybersecurity requests. The company also says it deliberately reduced some cyber capabilities during training and is using Opus 4.7, not Mythos Preview, as the public testbed for those safeguards in the official announcement.

That section also introduces a new Cyber Verification Program for security professionals who want legitimate access for vulnerability research, penetration testing, and red-teaming. It is a rare case where the capability limit is part of the product story, not just the safety appendix.