Claude Opus 4.6: The Long-Context Release, Now Legacy

What Claude Opus 4.6 is

Claude Opus 4.6 was released on 5 February 2026 and is the model that made Anthropic's 1 million token context window actually usable rather than merely advertised. On LMArena's text leaderboard Claude Opus 4.6 has 1496 Elo (rank 9); by category, 1546 Elo in Coding (rank 7), 1503 Elo in Math (rank 15), 1533 Elo in Expert (rank 8) and 1477 Elo in Creative Writing (rank 14). It takes text and image input, returns up to 128,000 output tokens, and is priced at $5 per million input tokens and $25 per million output.

The context-rot result is the one that mattered

Most long-context claims are about how much text a model will accept. The useful question is how much of it the model can still use at the far end, a failure mode usually called context rot. Anthropic's published figure here is the sharpest in the release: on the 8-needle 1M variant of MRCR v2, a retrieval test that hides information in very large inputs, Opus 4.6 scored 76% where Sonnet 4.5 scored 18.5%. It is Anthropic's own evaluation, run on their harness.

Vendor benchmark claims, and how they were run

Anthropic reported state-of-the-art results across several evaluations at release, and the conditions attached to them are worth as much as the scores. On SWE-bench Verified the stored figure is 79%, which Anthropic averaged over 25 trials; they note a prompt modification produced 81.42%. On BrowseComp, a test of finding hard-to-locate information online, Claude models were run with web search, web fetch, programmatic tool calling, context compaction triggered at 50k tokens up to 10M total, maximum reasoning effort and no thinking enabled — and Anthropic add that a multi-agent harness lifted the score to 86.8%, which is a different system rather than a different model. On GDPval-AA, an evaluation of economically valuable knowledge work, Anthropic put Opus 4.6 around 144 Elo points ahead of GPT-5.2, translating to winning roughly 70% of the time. Harvey reported the highest BigLaw Bench score of any Claude model at 90.2%. Anthropic also claimed the top Terminal-Bench 2.0 result and a lead on Humanity's Last Exam. Independently, SimpleBench records 68 and FrontierMath 38.

Status

Anthropic list Opus 4.6 under legacy models; it remains available. Opus 5.5 costs $4 and $20 per million tokens against this model's $5 and $25, and Anthropic claim it finishes comparable work in fewer turns. Code written against Opus 4.6 is not affected by the thinking and tool-use breaking changes introduced with the 5.5 generation.

Release details and platform IDs

Opus 4.6 was the first Opus-class model with a 1M token context window, introduced in beta. The release added adaptive thinking, effort controls with a default of high, context compaction on the API, and agent teams in Claude Code. Its model ID is claude-opus-4-6 on the Claude API, Google Cloud, Microsoft Foundry and Claude Platform on AWS, and anthropic.claude-opus-4-6-v1 on Amazon Bedrock through the InvokeModel integration. Anthropic's documentation gives a reliable knowledge cutoff of May 2025, a training data cutoff of August 2025, a 300K token maximum output on the Batch API in beta, and retirement not sooner than 5 February 2027. Cache reads cost $0.50 per million tokens; 5-minute cache writes $6.25 and 1-hour writes $10.

Official Resources