Claude Opus 4.7 Guide: The Release That Changed the API

Claude Opus 4.7 is easy to misread as a routine point release. The price did not move and the model ID follows the usual pattern. But for anyone calling the API directly, 4.7 is the release where the request surface changed, and code written for Opus 4.6 can fail on it with a 400 rather than degrade quietly.

Anthropic released it on 16 April 2026 as a direct upgrade to Opus 4.6, aimed at hard, long-running software engineering that needs little supervision. Six months on, three newer Opus models exist. This guide covers what 4.7 introduced, what those changes cost you, and the narrow set of cases where it is still the right model to pin.

What Opus 4.7 changed

Most of what is now normal across the Claude line arrived here first. According to Anthropic's announcement and API documentation:

• Manual thinking budgets were removed. thinking: {type: "enabled", budget_tokens: N} returns a 400. Adaptive thinking plus the effort parameter replaces it.
• Sampling parameters were removed. Setting temperature, top_p or top_k to a non-default value returns a 400 on 4.7 and every later model.
• A new xhigh effort level sits between high and max. Anthropic recommends it for most coding and agentic work, and Claude Code moved its default to it.
• Thinking text is omitted by default. Thinking blocks still arrive, but empty, unless you set display: "summarized". A product that streams reasoning to users suddenly shows a long pause.
• Adaptive thinking is off unless requested. A request with no thinking field runs without thinking, as on 4.6.

None of these are exotic. They are exactly the parameters an older integration is likely to set, which is why the 4.6 to 4.7 move is the one that breaks things. Opus 4.8 later kept this surface unchanged.

The tokenizer change is a hidden price change

The list price stayed at $5 per million input tokens and $25 per million output. The tokenizer did not stay the same. Anthropic says the same input may map to roughly 1.0 to 1.35 times as many tokens depending on content type, and its pricing page describes the newer tokenizer used by Claude 4.7 and later models as producing roughly 30% more tokens for the same text.

The practical effect is that an identical prompt can cost more on 4.7 than on 4.6 at the same per-token rate. It also breaks anything calibrated against 4.6 token counts: client-side estimators, max_tokens ceilings that used to leave headroom, compaction triggers and rate-limit thresholds. Re-run count_tokens against the new model on a sample of real prompts rather than applying a blanket multiplier. Tokenizer differences across models explains why the same text costs different amounts.

High-resolution vision and task budgets

Two additions were genuinely new capability rather than API housekeeping.

Vision. Opus 4.7 accepts images up to 2,576 pixels on the long edge, which Anthropic describes as more than three times the resolution of prior Claude models. Coordinates the model returns map one-to-one to image pixels, which removes scale-factor arithmetic from computer-use and screenshot pipelines. The cost is more image tokens per full-resolution image, so downsample when you do not need the fidelity.

Task budgets. A public beta that tells the model how many tokens it has for a whole agentic loop. It sees a running countdown and paces itself. This is advisory, not a hard cap; max_tokens remains the enforced limit. Agent token budgets covers how to use a budget without starving a hard task.

Anthropic also reported that 4.7 follows instructions more literally than earlier models, so prompts written for 4.6 may need retuning. Secondary coverage of its launch chart put SWE-bench Pro at 64.3%, against 53.4% for Opus 4.6; treat that as a vendor-run figure.

Is there still a reason to run 4.7?

Mostly no, and the reason is price. Anthropic's pricing page lists Opus 4.7, Opus 4.8 and Opus 5 at the same $5 / $25. Opus 5.5 is cheaper again at $4 / $20. A newer model at the same or lower price, with the same tokenizer family, removes the usual argument for staying behind.

The 4.7 to 4.8 move in particular carries no new breaking changes: it is a model-ID swap plus prompt re-tuning. If you are on 4.7 today, that is the cheapest upgrade you will get. The Opus 4.8 guide covers what shifts in its behaviour.

The cases for pinning 4.7 are narrow:

• You validated a regulated or audited workflow on 4.7 and re-validation costs more than the model difference is worth.
• You have an eval suite that shows 4.7 does better on your specific task, which happens more than leaderboards suggest.
• You need fixed behaviour while you migrate something else. Anthropic lists its retirement as not sooner than 16 April 2027, so there is runway.

Note that fast mode is not one of them: Anthropic has removed it from 4.7 and it now exists only on Opus 4.8 and later Opus models.

Variants and the batch endpoint

The catalogue lists two endpoints for this model. The plain ID is the standard synchronous model. The :batch variant is the same model on Anthropic's asynchronous Batch API, which Anthropic prices at half the standard rate — $2.50 input and $12.50 output per million tokens. Results arrive within hours rather than seconds, so it only suits work nobody is waiting on.

OpenRouter lists a 1M-token context and 128K maximum output for 4.7, which matches Anthropic's documentation.

Testing a move off 4.6

If you are still on Opus 4.6 and deciding where to land, skip 4.7 and target 4.8 or later directly. The breaking changes are the same either way. A sensible order:

1. Strip budget_tokens, sampling parameters and any assistant-turn prefill from your requests.
2. Set thinking: {type: "adaptive"} explicitly, and display: "summarized" if users see reasoning.
3. Re-baseline token counts with count_tokens and raise max_tokens headroom.
4. Sweep effort at medium, high and xhigh on twenty real tasks and record cost per completed task, not cost per request.

The last step matters most. Effort moves both quality and spend, and the right level depends on the work. How to benchmark LLMs on your own work has a harness for doing it properly.