Claude Opus 4.7: The Computer-Use Jump Over Opus 4.6

What Claude Opus 4.7 is

Claude Opus 4.7 was released on 16 April 2026, ten weeks after Opus 4.6, and it is the more interesting of the two if your work involves a model looking at a screen. Anthropic kept the specification steady — 1 million token context, 128,000 maximum output tokens, $5 per million input and $25 per million output, text and image in — and spent the release on capability rather than economics. On LMArena's text leaderboard Claude Opus 4.7 has 1494 Elo (rank 12); by category, 1546 Elo in Coding (rank 8), 1491 Elo in Math (rank 29), 1533 Elo in Expert (rank 9) and 1483 Elo in Creative Writing (rank 12). Reading the release notes, the through-line is agentic reliability: doing multi-step work without losing the thread, and seeing an interface accurately enough to act on it.

The visual acuity result

The single largest reported gain in this release did not come from Anthropic's own suite. XBOW, who run autonomous penetration testing built on computer use, reported Opus 4.7 scoring 98.5% on their visual-acuity benchmark against 54.5% for Opus 4.6. They described their biggest pain point with the Opus line as effectively disappearing, unlocking work they had previously been unable to use the model for. It is a customer's benchmark on a customer's harness, so it is a claim rather than an independent result.

Coding and legal work, per the vendor

On Anthropic's internal 93-task coding benchmark, Opus 4.7 lifted resolution by 13% over Opus 4.6, including four tasks that neither Opus 4.6 nor Sonnet 4.6 could solve at all. Anthropic pair that with faster median latency and stricter instruction following, and argue the combination matters most for long-running coding workflows where each failed step costs a retry. Harvey reported 90.9% on BigLaw Bench at high effort, noting better reasoning calibration on review tables and correct handling of a distinction — assignment provisions versus change-of-control provisions — that they say has historically defeated frontier models. Two caveats Anthropic publish themselves: the Terminal-Bench 2.0 figures used the Terminus-2 harness with thinking disabled, averaged over five attempts per task; and the SWE-bench Multimodal numbers used an internal implementation for both 4.6 and 4.7, so they are not directly comparable to published leaderboards.

LMArena: Opus 4.7 and Opus 4.6

On LMArena, Opus 4.7 has 1494 Elo (rank 12) and Opus 4.6 has 1496 Elo (rank 9). By category: coding 1546 Elo in Coding (rank 8) against 1546 Elo in Coding (rank 7); math 1491 Elo in Math (rank 29) against 1503 Elo in Math (rank 15).

Price and compatibility

The price is identical to Opus 4.6. Opus 5.5 lists at $4 and $20 with cheaper cache reads, and Anthropic claim it completes comparable work in fewer turns. Opus 4.7 predates the Claude 5 breaking changes around thinking and forced tool use.

Vision, safeguards and platform IDs

Anthropic state that Opus 4.7 sees images at greater resolution than Opus 4.6, and that it is the first model released with safeguards that automatically detect and block prohibited or high-risk cybersecurity requests; security professionals can apply to the Cyber Verification Program for legitimate uses. CursorBench, run by Cursor, records 70% against Opus 4.6's 58%. The model ID is claude-opus-4-7 on the Claude API, Google Cloud, Microsoft Foundry and Claude Platform on AWS, and anthropic.claude-opus-4-7 on Amazon Bedrock. The documentation gives a reliable knowledge cutoff of January 2026, a 300K token Batch API output limit in beta, and retirement not sooner than 16 April 2027.

Official Resources