Claude Opus 4.7: Pricing, Context Window, Benchmarks, API

What Is Claude Opus 4.7?

Last updated August 2026.

Claude Opus 4.7 is an Anthropic frontier model released April 16, 2026. The model ID is claude-opus-4-7. It has a 1M-token context window, 128K max output on the Messages API, and costs $5 per million input tokens and $25 per million output tokens, the same rate Opus 4.6 carried.

The rate held. The bills did not. Opus 4.7 introduced a new tokenizer, and the same text produces roughly 30% more tokens on it than on any earlier Claude model. Teams that migrated from 4.6 and compared invoices saw costs rise on an unchanged price sheet. That tokenizer now carries forward into Opus 4.8, Claude Opus 5, and Claude Fable 5.

Anthropic's framing for the release was rigor over raw score: Opus 4.7 "handles complex, long-running tasks with rigor and consistency, pays precise attention to instructions, and devises ways to verify its own outputs before reporting back." It also shipped substantially better vision and the xhigh effort level.

Pricing and the Tokenizer Change

Opus 4.7 costs $5 per million input tokens and $25 per million output tokens. Every Opus-tier model from 4.5 forward carries that same rate.

The interesting cost story is not the rate, it is the tokenizer. Anthropic's docs note that Opus 4.7 introduced a new tokenizer and that, compared with models released before it, the same text produces roughly 30% more tokens. The exact increase depends on the content. Code and non-English text move differently from English prose.

The consequences are mechanical. A 1M-token context window holds less text than the same window on Opus 4.6. A max_tokens value tuned on 4.6 can truncate equivalent output on 4.7. Compaction triggers keyed to measured token counts fire earlier. Client-side estimators calibrated against an older model under-report.

Context Window and Output Limit

Opus 4.7 has a 1M-token context window, which Anthropic annotates as roughly 555,000 words. Max output is 128K tokens on the synchronous Messages API.

On the Message Batches API, Opus 4.7 supports up to 300K output tokens using the output-300k-2026-03-24 beta header, alongside Opus 5, Opus 4.8, Opus 4.6, Sonnet 5, and Sonnet 4.6.

Note the interaction with the tokenizer. The window is 1M tokens on both 4.6 and 4.7, but 4.7 fits meaningfully less source text into it. If you sized a retrieval budget against 4.6 and moved the model without moving the budget, you are closer to the ceiling than your configuration suggests. See context rot for what that does to agent quality.

Claude Opus 4.7 Benchmarks

Anthropic names SWE-bench Verified, SWE-bench Pro, SWE-bench Multilingual, Terminal-Bench 2.0, GPQA Diamond, OSWorld, and ARC-AGI-2 in the Opus 4.7 launch material, but publishes those figures inside chart images rather than as text. The scores stated in the body text are these:

The 93-task result carries the detail worth keeping: Opus 4.7 solved four tasks that neither Opus 4.6 nor Sonnet 4.6 could solve. A 13% aggregate lift on a saturating benchmark is a smaller signal than a category of previously unsolvable tasks becoming solvable.

Vision

Opus 4.7 was the release where Claude's image resolution jumped. It accepts images up to 2,576 pixels on the long edge, about 3.75 megapixels, which Anthropic describes as "more than three times as many as prior Claude models." Model coordinates map one-to-one to image pixels, so scale-factor math written for earlier models should be removed. The cost side: a full-resolution image can consume up to roughly three times the image tokens it did before.

Opus 4.7 vs Opus 4.8 vs Opus 5

All three cost $5 input and $25 output per million tokens, with a 1M context window and 128K max output. The differences are capability, knowledge cutoff, and how much work a migration takes.

Opus 4.7 to Opus 4.8 introduces no new breaking API changes. The same requests that work on 4.7 work on 4.8, so the migration is a model-ID swap plus prompt re-tuning. Opus 4.8 narrates more between tool calls, is more conservative about reaching for search and subagents, and asks more often on minor decisions.

Opus 4.8 to Claude Opus 5 adds two real breaking changes. Thinking is on by default, so a request that omits the thinking parameter now thinks and spends against the same max_tokens ceiling as the response text. And thinking: {type: "disabled"} is only accepted at effort high or lower; pairing it with xhigh or max returns a 400.

API Access and Breaking Changes

Pass claude-opus-4-7 as the model to the Messages API. Opus 4.7 is available on the Claude API, on Amazon Bedrock as anthropic.claude-opus-4-7, on Google Cloud as claude-opus-4-7, and on Microsoft Foundry.

Four changes catch code written against Opus 4.6 or earlier. The first two hard-fail; the second two fail quietly.

The two quiet failures are the ones that reach production. Thinking blocks still appear in the response stream on 4.7, but the thinking field is empty unless you opt in with display: "summarized". A UI that streams reasoning shows a long pause instead of an error.

Effort and the xhigh level

Opus 4.7 added xhigh, an effort level between high and max. Effort is set through output_config.effort and accepts low, medium, high, xhigh, and max. Anthropic recommends xhigh for coding and agentic work and a minimum of high for intelligence-sensitive workloads. At xhigh or max, give max_tokens real headroom: the model needs room to think and act across tool calls.

Opus 4.7 also shipped task budgets in public beta, which give the model a token ceiling for a whole agentic loop so it paces itself and finishes gracefully rather than being cut off. That is a different mechanism from max_tokens, which is an enforced per-response cap the model never sees.

For per-request pricing on the Anthropic API including batch discounts and prompt caching, see our Anthropic API pricing guide.

Pros and Cons

Open-Weight Alternatives on Morph

Morph does not serve Claude models. It serves open-weight models on custom kernels through an OpenAI-compatible API. For coding workloads, several cost a fraction of Opus 4.7's $5 and $25, and because the weights are public, a model you prototype on the API is one you can self-host later.

GLM-5.2 is the closest substitute for agentic coding: 753B MoE under an MIT license with a 1M-token context, at 22% of Opus 4.7's input rate and 16% of its output rate. DeepSeek V4 Flash sits two orders of magnitude below on output for high-volume pipelines.

See GLM-5.2, Morph open source models, and pricing.

Frequently Asked Questions

What is Claude Opus 4.7?

An Anthropic frontier model released April 16, 2026. Model ID claude-opus-4-7, 1M-token context, 128K max output, $5/M input and $25/M output. It introduced the xhigh effort level, 2,576px vision, task budgets in public beta, and a new tokenizer. Anthropic now lists it under legacy models.

How much does Opus 4.7 cost?

$5 per million input tokens and $25 per million output tokens, identical to Opus 4.6, Opus 4.8, and Claude Opus 5. The rate held across all four, but the 4.7 tokenizer means the same text costs roughly 30% more tokens than on 4.6.

What is Opus 4.7's context window?

1M tokens, roughly 555,000 words. Max output is 128K on the Messages API and up to 300K on the Message Batches API with the output-300k-2026-03-24 beta header.

What did Anthropic publish for Opus 4.7 benchmarks?

In text: a 13% resolution lift over Opus 4.6 on its 93-task internal coding benchmark (including four tasks neither Opus 4.6 nor Sonnet 4.6 solved), 70% on CursorBench against 4.6's 58%, 90.9% on BigLaw Bench at high effort, and 3x more production tasks resolved on Rakuten-SWE-Bench. SWE-bench and Terminal-Bench scores appear only in chart images, not in body text.

Should I move off Opus 4.7?

Opus 4.8 and Claude Opus 5 cost the same $5/$25 with the same 1M context and 128K max output, so there is no price reason to stay. 4.7 to 4.8 has no new breaking changes, only prompt re-tuning. 4.8 to Opus 5 adds two: thinking on by default, and disabled thinking capped at high effort.

Why did my costs rise after moving to Opus 4.7?

The tokenizer. Opus 4.7 introduced a new one, and the same text produces roughly 30% more tokens than on earlier Claude models at an unchanged per-token rate. Re-run count_tokens against claude-opus-4-7 on your own prompts rather than applying a flat multiplier; the increase varies by content type.

Why are my Opus 4.7 thinking blocks empty?

Thinking content is omitted by default on 4.7, a silent change from Opus 4.6 where summarized thinking was returned. Set thinking: {type: "adaptive", display: "summarized"} to restore it. The block field name is unchanged.

Can I self-host Claude Opus 4.7?

No. Claude models are closed-weight and API-only. If self-hosting or fine-tuning is a requirement, an open-weight model such as GLM-5.2 (MIT license, 1M context) is the substitute.