Anthropic announced Claude Opus 4.6 on February 5, 2026. The release upgrades the Opus-class models with stronger coding abilities, longer agentic task endurance, improved long-context retrieval and reasoning, and an expanded safety evaluation suite. Opus 4.6 is the first Opus-class model to offer a 1 million token context window in beta.
Capabilities
- Enhanced coding and debugging: Opus 4.6 plans more carefully, sustains agentic workflows for longer, navigates larger codebases more reliably, and has improved code review and self-debugging skills.
- Long-context handling: the model demonstrates better retrieval and multi-step reasoning over very large contexts and is designed to reduce "context rot." The 1M-token context window is currently available in beta on the Claude Developer Platform.
- Knowledge work: Opus 4.6 can be applied to financial analysis, research, and document, spreadsheet and presentation workflows. In Anthropic’s Cowork environment, where Claude can multitask autonomously, these capabilities combine to support extended agentic work.
Benchmark performance and specific results
Anthropic reports that Opus 4.6 achieves state-of-the-art results on several evaluations. Highlights include leading scores on Terminal-Bench 2.0 (agentic coding) and top performance on Humanity’s Last Exam (a complex multidisciplinary reasoning test).
Notable quantitative results provided by Anthropic:
- On GDPval-AA (evaluating economically valuable knowledge work in finance, legal and related domains), Opus 4.6 outperforms OpenAI’s GPT-5.2 by about 144 Elo points and its predecessor Claude Opus 4.5 by 190 points.
- MRCR v2 (8-needle 1M variant): Opus 4.6 scored 76% versus Sonnet 4.5’s 18.5% on a needle-in-a-haystack retrieval task.
- BigLaw Bench: Opus 4.6 scored 90.2%, with 40% perfect scores and 84% above 0.8, indicating strong legal reasoning capability.
- Vending-Bench 2: the model earned on average $3,050.53 more than Opus 4.5.
- Domain expertise: Anthropic reports nearly 2× improvement over Opus 4.5 on computational biology, structural biology, organic chemistry and phylogenetics tests.
Anthropic has published a detailed system card with methodology and more benchmark comparisons.
Safety testing and mitigations
Anthropic says the intelligence gains do not compromise safety. In their automated behavioral audit, Opus 4.6 showed low rates of misaligned behaviors (deception, sycophancy, encouraging delusions, cooperation with misuse) and matched or improved upon Opus 4.5’s alignment. It also registered the lowest rate of over-refusals among recent Claude models.
The company ran an extensive set of safety evaluations — some newly developed — including tests for user wellbeing, richer refusal behavior evaluations, and probes for covert harmful behavior. Anthropic also used interpretability techniques to better understand internal model behavior and to catch issues standard tests might miss.
Because Opus 4.6 is strong at cybersecurity tasks, Anthropic introduced six new cybersecurity probes to detect potential misuse and is accelerating defensive applications (finding and patching vulnerabilities in open-source software). They note they will continue to update safeguards and may add real-time interventions to block abuse.
Product and API updates
Anthropic introduced several API and platform features to support long-running agents and let developers control trade-offs between intelligence, speed and cost:
- Adaptive thinking: instead of a binary choice for extended thinking, Claude can now decide when deeper reasoning is helpful. At the default effort level (high) the model uses extended thinking selectively; developers can tune this with the effort parameter.
- Effort levels: four levels are available — low, medium, high (default), and max — to balance cost, latency and reasoning depth.
- Context compaction (beta): automatically summarizes and replaces older context as conversations approach a configurable threshold, enabling longer tasks without hitting the context window.
- 1M-token context (beta): Opus 4.6 supports a 1 million token context window on the Claude Developer Platform; premium pricing applies for prompts exceeding 200k tokens.
- 128k output tokens: the model supports outputs up to 128,000 tokens to avoid splitting large outputs into multiple requests.
- US-only inference: an inference option restricted to run in the United States is available at 1.1× token pricing.
The model is exposed on the Claude API as claude-opus-4-6. Base pricing remains at $5/$25 per million tokens; premium pricing for prompts over 200k tokens is $10/$37.50 per million input/output tokens.
Platform integrations and developer tools
Anthropic also added product features across Claude and Claude Code to help knowledge workers and developers tackle more complex tasks:
- Agent teams (research preview): spin up multiple subagents that work in parallel and coordinate autonomously, useful for tasks that split into independent, read-heavy subtasks such as codebase reviews.
- Claude in Excel: improved handling of long-running tasks, the ability to infer structure from unstructured data, plan before acting, and perform multi-step changes in one pass.
- Claude in PowerPoint: now in research preview for Max, Team and Enterprise plans; the model reads layouts, fonts and slide masters to keep output on brand and can generate full decks from descriptions.
Availability and next steps
Claude Opus 4.6 is available on claude.ai, via the Claude API and on major cloud platforms. Developers can call the model using claude-opus-4-6 on the API. Anthropic’s system card and accompanying documentation provide detailed descriptions of capability and safety evaluations, benchmark methodologies, and implementation notes.
Anthropic’s announcement also includes multiple early access partner testimonials describing improved autonomy, planning, and real-world efficacy across engineering, legal, security and design workflows. The company emphasizes it will continue iterating on safeguards and product controls as usage and threat models evolve.
