Claude Opus 4.6 Explained: 1M Context, Benchmark Claims, Pricing and Who Should Use It

Bottom line: Claude Opus 4.6, released on February 5, 2026, is Anthropic’s flagship model upgrade focused on longer agentic tasks, larger codebases and better self-checking. It is the first Opus-class model with a 1M-token context window (in beta, on the Claude Developer Platform only), and Anthropic kept list pricing at $5 per million input tokens and $25 per million output tokens. The benchmark results below are Anthropic’s own; treat them as a guide to where the model is strong, then confirm on your own tasks.

This overview is based on Anthropic’s announcement and its system card. It does not include first-hand test results.

What changed

According to Anthropic, Opus 4.6 “plans more carefully, sustains agentic tasks for longer, can operate more reliably in larger codebases, and has better code review and debugging skills to catch its own mistakes.” Beyond coding, Anthropic positions it for financial analysis, research, and working with documents, spreadsheets and presentations, including in Cowork, Claude in Excel and the Claude in PowerPoint research preview.

Key developer-facing changes:

• 1M-token context (beta). Prompts above 200K tokens are billed at a premium rate of $10 input and $37.50 output per million tokens.
• Up to 128K output tokens, so larger outputs do not need to be split across requests.
• Adaptive thinking and effort levels. Instead of turning extended thinking on or off, the model decides when to think more; developers choose an effort level (low, medium, high by default, or max).
• Context compaction (beta), which summarises older context automatically so long-running agents do not hit the limit.
• US-only inference at 1.1x token pricing for workloads that must stay in the United States.

Reported benchmark results

Anthropic reports that Opus 4.6:

• achieves the highest score on Terminal-Bench 2.0, an agentic coding evaluation;
• leads other frontier models on Humanity’s Last Exam, a multidisciplinary reasoning test;
• outperforms OpenAI’s GPT-5.2 by around 144 Elo points on GDPval-AA, an independent evaluation of economically valuable knowledge work run by Artificial Analysis, and its predecessor Opus 4.5 by 190 points;
• scores 76% on the 8-needle 1M variant of MRCR v2, a long-context retrieval test, versus 18.5% for Sonnet 4.5.

Methodology matters: Anthropic’s footnotes describe specific harnesses, effort settings and tool access for each result, and it corrected one Humanity’s Last Exam score after improving cheating detection. Benchmarks run under different conditions are not directly comparable.

Long context and “context rot”

A recurring problem with long conversations is what practitioners call “context rot”: performance degrades as the number of tokens grows, and details buried early in a long input are missed. Anthropic says Opus 4.6 is much better at retrieving relevant information from large document sets, tracks information over hundreds of thousands of tokens with less drift, and performs markedly better than its predecessors on the 8-needle, 1M-token variant of a needle-in-a-haystack test (Anthropic). For workloads such as contract review, due diligence or large-codebase analysis, that is more important than the size of the window itself.

What early-access customers reported

Anthropic’s announcement quotes early-access partners. These are vendor-selected testimonials, but a few include specific numbers:

• Harvey reported a 90.2% BigLaw Bench score, the highest of any Claude model on that legal-reasoning benchmark.
• Box reported that, on its multi-source analysis evaluation across legal, financial and technical content, performance rose to 68% from a 58% baseline.
• One partner reported that, across 40 cybersecurity investigations ranked blind, Opus 4.6 produced the best result 38 times out of 40 compared with Claude 4.5 models.

Treat these as signals of where the model is strong rather than as guarantees for your workload.

Safety

Anthropic says Opus 4.6 has a low rate of misaligned behaviour in its automated audit, is as well aligned as Opus 4.5, and has the lowest over-refusal rate of recent Claude models. Because the model is stronger at cybersecurity tasks, Anthropic added new detection probes for misuse. Details are in the system card.

Pricing and practical cost

List pricing is unchanged from Opus 4.5 at $5/$25 per million tokens; see Claude pricing. Anthropic also notes the model “often thinks more deeply,” which improves hard problems but can add cost and latency on simple ones, and recommends lowering effort from high to medium if it overthinks. In practice, the effort setting is the main lever for controlling spend.

Who should use it

• Good fit: large-codebase maintenance, long agentic coding sessions, document-heavy analysis, and tasks where catching its own mistakes matters more than speed.
• Probably overkill: short, simple questions and high-volume classification or extraction, where a smaller Claude model is cheaper and faster.

How to evaluate it for your team

1. Pick a handful of real tasks your team does weekly, including one long, multi-step task.
2. Run them at medium and high effort and compare quality, token usage and latency.
3. If you plan to use very long context, budget for the premium rate above 200K tokens.
4. Compare against your current model on the same prompts before switching.