Skip to main content

What's new in Claude Opus 4.7

Claude Opus 4.7 was announced by Anthropic on April 16, 2026 as the company's latest generally available flagship Opus model. The release is positioned as a practical upgrade over Opus 4.6 rather than a brand-new family: same headline API pricing, same 1M-token context window, but materially better performance on hard coding, long-running agent tasks, multimodal work, and professional document generation.

For teams already using Claude in production, the important story is not only "the model got better." Anthropic also changed several API behaviors in ways that affect migration, token budgeting, and prompt design.

The short version​

  • Release date: April 16, 2026
  • Model ID: claude-opus-4-7
  • Availability: Claude apps, Anthropic API, Amazon Bedrock, Google Cloud Vertex AI, and Microsoft Foundry
  • Context window: 1M tokens
  • Max output: 128k tokens
  • Pricing: $5 per million input tokens, $25 per million output tokens
  • Knowledge cutoff: January 2026

Source: Anthropic announcement, Models overview

What improved in Opus 4.7​

Anthropic's launch notes emphasize four upgrades over Opus 4.6:

1. Better at difficult engineering work​

Anthropic frames Opus 4.7 as a stronger model for advanced software engineering, especially on long-horizon tasks that previously needed close supervision. The company says the model is more rigorous on multi-step work, follows instructions more precisely, and is more likely to verify its own output before returning an answer.

That positioning is backed by partner feedback in the launch post:

  • Cursor reported its internal CursorBench score rising from 58% to 70%
  • Rakuten said it resolved 3x more production tasks than Opus 4.6 on Rakuten-SWE-Bench
  • CodeRabbit reported a 10%+ recall improvement on code review workloads

These are vendor-supplied evaluations, not neutral public benchmarks, but they all point in the same direction: Opus 4.7 is tuned for serious coding and agent workflows.

2. Stronger autonomy and long-running agent behavior​

This is one of the clearest themes across Anthropic's materials. Opus 4.7 is described as more capable on long-horizon agentic work, with better persistence, fewer tool errors, and less tendency to stop early or require re-prompting.

The official migration guide also hints at a behavior shift here: Opus 4.7 uses reasoning more selectively, makes fewer tool calls by default, and calibrates response length to the task instead of using a fixed verbosity pattern. In practice, that means old harnesses may still work, but they should be re-tested rather than assumed stable.

3. Much better vision​

Opus 4.7 is Anthropic's first Claude model with high-resolution image support. The maximum image size increased:

  • From 1568px / 1.15MP
  • To 2576px / 3.75MP

That matters for:

  • browser and computer-use agents
  • dense screenshots
  • slide and document editing
  • diagram and chart analysis
  • UI generation and visual QA

Anthropic also notes a more direct coordinate mapping: the model now works with image coordinates at a 1:1 pixel scale, which simplifies image-based tooling.

4. Better output quality for professional artifacts​

Anthropic explicitly calls out higher-quality interfaces, slides, and documents. This is not just a benchmark claim; it is part of how the release is marketed. The company is pushing Opus 4.7 as a model that is not only more accurate, but also more polished and more tasteful when generating work products people actually ship or present.

New platform features around Opus 4.7​

Beyond raw model quality, Anthropic shipped several platform-level changes.

xhigh effort​

Opus 4.7 adds a new xhigh effort level between high and max. Anthropic recommends starting with xhigh for coding and agentic work, and at least high for most intelligence-sensitive use cases.

This matters because Opus 4.7 relies more heavily on the effort setting than earlier Opus releases. If quality, latency, or cost shift after migration, this is one of the first levers to retune.

Source: What's new in Claude Opus 4.7, Effort docs

Task budgets beta​

Anthropic also introduced task budgets for Opus 4.7. A task budget is an advisory token budget for the full agent loop, including thinking, tool calls, tool results, and final output.

This is different from max_tokens:

  • max_tokens is a hard per-request cap
  • task_budget is a model-visible budget hint for the whole workflow

For teams building autonomous agents, this is one of the more meaningful release additions because it gives the model explicit awareness of a total token allowance.

Source: Task budgets docs

Breaking changes and migration risks​

If you are moving from Opus 4.6 to Opus 4.7, the upgrade is not a blind string replacement.

1. Old extended thinking budgets are gone​

The old pattern:

{"thinking": {"type": "enabled", "budget_tokens": 32000}}

is no longer supported on Opus 4.7. Anthropic says to use:

{"thinking": {"type": "adaptive"}}

and control depth with the effort setting instead.

2. Non-default sampling parameters now error out​

Starting with Opus 4.7, setting temperature, top_p, or top_k to any non-default value returns a 400 error. Anthropic's recommended migration path is to remove those parameters and steer behavior through prompting and effort selection.

3. Thinking is hidden by default​

Thinking content is now omitted by default unless you explicitly opt in to summarized display. If your product streams reasoning steps to users, this can look like a stall unless you update your request config.

4. Token counts may increase​

Opus 4.7 uses a new tokenizer. Anthropic says the same text may consume roughly 1.0x to 1.35x as many tokens as earlier models, depending on content type. That does not automatically mean worse economics, but it does mean you should re-benchmark:

  • request cost
  • latency
  • max_tokens headroom
  • compaction triggers
  • any client-side token estimation

Source: Migration guide, What's new in Claude Opus 4.7

Safety and product positioning​

Anthropic is also using Opus 4.7 as an early production deployment for new cybersecurity safeguards. In the announcement, the company says Opus 4.7 is the first model released with automatic blocking for prohibited or high-risk cybersecurity requests, while the more capable Claude Mythos Preview remains more restricted.

That positioning matters because it clarifies Anthropic's lineup:

  • Opus 4.7 is the top generally available Opus model
  • Mythos Preview is positioned above it in raw capability, but not broadly released

Source: Anthropic announcement

Bottom line​

Claude Opus 4.7 looks like a meaningful production upgrade, especially for teams building coding agents, research agents, document workflows, or visual automation. The strongest reasons to care are better long-running reliability, much better vision resolution, and a more controllable reasoning stack through adaptive thinking, xhigh effort, and task budgets.

The catch is that migration discipline matters. Prompt behavior is more literal, tokenization changed, and several old API knobs no longer work. If you treat Opus 4.7 as a drop-in replacement for Opus 4.6, you may miss both the gains and the new failure modes.

Sources​