Opus 4.7 migration, what changed for API builders

Opus 4.7 went GA on April 16, 2026. At first glance it looks like a routine refinement release, the pricing is unchanged (5 dollars input, 25 dollars output per million tokens, 1 million context stays), and the model is still called claude-opus-4-7. But the moment you see the first 400 error in your production code, you know the migration is not entirely trivial.

This playbook is for you if you have a codebase that makes Anthropic API calls with Opus 4.6 or earlier, and you want to make sure nothing breaks when the model is bumped to 4.7. Plus a short assessment of whether the new xhigh effort level makes sense for your use case.

Three breaking API changes, hard 400 errors

Anthropic removed three parameters from the messages.create() endpoint that were still accepted in 4.6. If you send them to Opus 4.7, you get a 400 Bad Request back.

First, thinking.budget_tokens is gone. In 4.6 you could give the model a hard token budget for the reasoning phase (thinking: { type: "enabled", budget_tokens: 5000 }). In 4.7 the concept is replaced by Adaptive Reasoning, the model decides for itself how much it thinks. Migration required, remove the budget_tokens field entirely from the request. If you need cost control (see below), use Task Budgets in the public beta.

Second, temperature and top_p are gone. Both were deprecated with a warning in 4.6, in 4.7 they are removed hard. If you send temperature: 0.7 along, 400 error. Migration required, remove both parameters with no replacement. Anthropic trained the model to calibrate the "right" sampling strategy itself. In practice that means you only give the API the conversation and the tool schema, no more sampling tuning.

Third, top_k is gone. Same logic as temperature and top_p.

In sum, the sampling configuration is no longer part of the API contract with Opus 4.7. That is a deliberate design statement by Anthropic, they want fewer knobs on the builder side.

Reasoning traces are hidden by default

A second point that stands out during migration, in 4.6 you got the full reasoning tokens back as a response stream by default with thinking.type: "enabled". In 4.7 this is changed to "summarized" as the default. That means you get a short summary of what the model thought, not the raw trace.

If you want the raw trace (debugging, eval infrastructure, or because you want to monitor the reasoning style), set thinking.display: "raw" explicitly. That is not more expensive, just explicitly requested.

Tokenizer update, 1.0 to 1.35x more tokens

Opus 4.7 uses a new tokenizer. Anthropic officially states that the same text input produces 1.0 to 1.35 times as many tokens as in 4.6 depending on language and content type.

In practice that means, code about 1.10x slightly more expensive, English-only text at 1.0 to 1.05x practically the same, multilingual or non-Western inputs up to 1.35x noticeably more expensive. If your API bills run in German or Spanish, plan for roughly 15 to 20 percent more token cost. If you build for Japanese- or Korean-speaking users, it can be up to 35 percent more.

CloudZero and the independent Karoz Substack confirmed this with test corpora. Anthropic's statement is found in the help-center release notes document for Opus 4.7.

Extended Thinking is gone entirely

In 4.6 you had the switch thinking.type: "enabled" | "disabled" plus the separate mode "Extended Thinking" for longer agentic reasoning phases. In 4.7 Extended Thinking is no longer its own mode. Adaptive Reasoning is the only mode, the model decides per turn whether and how much it thinks.

Migration note, if you have Extended Thinking explicitly enabled in your codebase (e.g. via beta header extended-thinking-2026-01-15), remove the header. Nothing dramatic happens if you leave it in, but it has no effect anymore.

New effort level "xhigh" plus Claude Code default switch

Anthropic introduced a new effort level between high and max, xhigh. That is relevant if you use the Claude API directly and set the effort_level parameter, or if you use Claude Code (CLI).

Important, Claude Code defaults to xhigh for all plans (Pro, Max, Team, Enterprise) since 2.1.110. That is a deliberate decision by Anthropic, the reasoning is in the NxCode developer guide for Opus 4.7 and in the Apiyi post on the xhigh mode. In practice that means, if you use Claude Code, you now think more often and longer per turn than before the update. Output quality gets noticeably better as a result, but token consumption goes up.

If you need cost control in Claude Code, you can lower the effort level per session to high with the /effort high command or via setting. Tradeoff, less reasoning, faster and cheaper answers, but in agentic workflows less clean tool-use behaviour.

Task Budgets as public beta

In parallel with 4.7, Anthropic released the concept of "Task Budgets" as a public beta. That is the replacement for thinking.budget_tokens, but at a higher level, namely the whole agentic loop (reasoning plus tool calls plus output).

Activation via beta header task-budgets-2026-03-13. You give the API a token limit for the whole task, and the model respects it. When the limit is reached, the model stops cleanly instead of grinding through. That is the relevant lever against cost explosion, especially for Claude Code and agentic frameworks.

Implementation note, the beta header goes into the extra_headers slot of the SDK. Plus max_tokens_for_task as a body parameter. You check the exact spec in the Verdent guide or in the Anthropic Help Center.

Migration checklist, seven points

If you want to migrate today, go through these seven points.

One, search your codebase for temperature:, top_p:, top_k: and thinking.budget_tokens. Remove all occurrences. Fastest pattern, grep -rn "temperature\|top_p\|top_k\|budget_tokens" src/ and manual cleanup.

Two, if you had cost control via budget_tokens, plan Task Budgets instead. Beta header and body parameter.

Three, if you parse reasoning tokens in your codebase (e.g. for your own eval pipelines), set thinking.display: "raw" explicitly.

Four, plan for a token-cost increase of roughly 15 to 20 percent for German-language workloads. Work updated cost estimates into your pricing pages or internal dashboards if needed.

Five, remove the Extended Thinking beta header (if used). No functional impact, but less dead code.

Six, if you use Claude Code and do not want an implicit default update, set the effort level explicitly instead of relying on the default.

Seven, run your existing eval tests once against 4.7 and compare with 4.6. Anthropic reports a score jump from 53 to 57 (Artificial Analysis Intelligence Index), that should be reflected in your real-world tests. If not, you have a prompt that was specifically optimized for 4.6 behaviour. Re-tune or stay on 4.6 deliberately (claude-opus-4-6 is still available).

Should you migrate at all

Anthropic set 4.7 as the default for new API calls. 4.6 stays explicitly addressable as claude-opus-4-6, but is no longer the default in the docs or in Claude Code. Practical recommendation, if your codebase is actively developed and you get fresh releases, migrate in the next 4 to 8 weeks. If your codebase is frozen and you only pay API bills, you can stay on 4.6 until Anthropic announces a sunset deadline.

As of May 2026 Anthropic has published no official sunset deadline for 4.6, but the pattern from the 4.5 sunset was roughly 6 to 9 months after the successor's GA. That means, plan for 4.6 being deprecated by Q4 2026 at the latest.

What comes next

If you are migrating Claude Code 2.1+ in parallel, the "claude-code-update-strategie" playbook explains the update ritual and which releases are stable. If you want to take token cost control via Task Budgets seriously, "claude-code-cost-controls" has the matching workflow. For API cost monitoring at the provider level, "claude-code-headless-in-ci-cd" shows how you measure token consumption per run.

Main sources if you want to dig deeper, Anthropic news post on Opus 4.7 (anthropic.com/news/claude-opus-4-7), help-center release notes (support.claude.com/en/articles/12138966-release-notes), Artificial Analysis pre-release review, and the NxCode developer guide for the migration-relevant API details.