Claude Opus 4.7: what's new and what the API changes mean

Claude Opus 4.7 launched April 16, 2026. It is Anthropic's most capable generally available model. Pricing is unchanged from Opus 4.6: $5 per million input tokens and $25 per million output tokens.

The model ID is claude-opus-4-7. It is available on the Claude API, Amazon Bedrock, Google Cloud Vertex AI, and Microsoft Foundry.

Important before you upgrade: Opus 4.7 has API breaking changes compared to Opus 4.6. If you use extended thinking with budget_tokens, set temperature, or use assistant prefills, your existing code will break. See Migrating from Opus 4.6 to Opus 4.7 for the full list.

What improved

Software engineering. Benchmark results show a 13% improvement over Opus 4.6 on a 93-task coding suite, with 3x more successful resolutions on Rakuten-SWE-Bench. Anthropic reports it resolves specific tasks that both Opus 4.6 and Sonnet 4.6 could not solve. For developers using Claude in agentic coding workflows — writing PRs, debugging production issues, scaffolding projects — this is the most practically relevant improvement.

Finance and document reasoning. Scores 0.813 on General Finance modules, up from 0.767 for Opus 4.6. Code review recall improved by over 10% while maintaining precision.

Vision. Opus 4.7 accepts images up to 2,576 pixels on the long edge, which works out to roughly 3.75 megapixels — more than three times the previous limit of 1,568 pixels. This matters for reading dense screenshots, extracting data from charts, and computer use workflows. Full-resolution images consume up to 4,784 tokens each (compared to ~1,600 tokens on prior models), so re-budget accordingly if you send images. Coordinates returned by the model are 1:1 with actual image pixels — no scale-factor conversion is needed anymore.

Long-context work. The 1M token context window is available at standard pricing with no long-context premium, same as Opus 4.6.

New things

xhigh effort. A new effort level between high and max. Anthropic recommends xhigh as the default for most coding and agentic use cases:

client.messages.create(
    model="claude-opus-4-7",
    max_tokens=64000,
    thinking={"type": "adaptive"},
    output_config={"effort": "xhigh"},
    messages=[{"role": "user", "content": "..."}],
)

Effort levels available: low, medium, high, xhigh, max. Effort is more important for Opus 4.7 than for any prior model.

Task budgets (public beta). A new parameter that tells the model how many total tokens it has for an agentic loop — across thinking, tool calls, and output. The model sees a running countdown and uses it to prioritize work. Unlike max_tokens (a hard per-request ceiling the model doesn't see), task budgets are advisory: the model self-moderates.

output_config = {
    "effort": "high",
    "task_budget": {"type": "tokens", "total": 128000},
}

Requires the beta header task-budgets-2026-03-13. Minimum value is 20k tokens. Use for agentic workloads where you want the model to scope itself; don't use where quality matters more than token count.

Better agentic progress updates. Opus 4.7 provides more regular status updates during long agentic runs. If you added scaffolding to force interim messages ("After every 3 tool calls, summarize progress"), try removing it first.

Behavioral changes

These aren't API breaking changes, but they may affect your outputs:

More literal instruction following. Opus 4.7 interprets prompts more literally. It won't generalize an instruction from one item to another or infer requests you didn't make. Good for structured pipelines. May require prompt updates for open-ended tasks.

More direct tone. Less validation-forward phrasing, fewer softeners and emoji than Opus 4.6. If your product relies on a specific voice, re-test style prompts.

Fewer subagents by default. Steerable via prompting — give explicit guidance when subagents are desirable.

Stricter effort calibration at low/medium. At low effort, the model scopes narrowly to what was asked. If you see shallow reasoning on moderately complex tasks, raise effort to high or xhigh.

New tokenizer. Text consumes 1.0–1.35x as many tokens as on Opus 4.6, varying by content. This is a cost consideration, not a capability concern. Re-budget max_tokens and re-test client-side token estimates.

Cybersecurity safeguards

Opus 4.7 adds real-time safeguards that block high-risk cybersecurity content. For legitimate security work — penetration testing, vulnerability research, red-teaming — apply to the Cyber Verification Program for reduced restrictions.

Who should upgrade

Now: If you are building new applications and weren't using the old extended thinking API. The upgrade path is simple: change the model name, omit sampling parameters, adopt adaptive thinking.

After testing: If you have existing Opus 4.6 prompts in production. The behavioral changes (literalism, tone, response length calibration) mean your outputs will shift. Test first.

Migration docs: platform.claude.com/docs/en/about-claude/models/migration-guide

The practical checklist and code examples are in Migrating from Opus 4.6 to Opus 4.7.