Feb 5 2026, Anthropic just launched Claude Opus 4.6 — and the benchmarks are brutal for competitors.
The new model outperforms OpenAI’s GPT-5.2 by 144 Elo points on real-world work tasks, scores highest in the industry on agentic coding, and becomes the first Opus-class model with a 1 million token context window.
It’s live now on claude.ai and the API.
The Numbers That Matter
- 144 Elo points ahead of GPT-5.2 on GDPval-AA (knowledge work tasks in finance, legal, and other domains)
- 190 Elo points ahead of its own predecessor, Claude Opus 4.5
- #1 score on Terminal-Bench 2.0 (agentic coding evaluation)
- #1 score on Humanity’s Last Exam (complex reasoning)
- #1 score on BrowseComp (finding hard-to-find information online)
- 76% vs 18.5% — Opus 4.6 vs Sonnet 4.5 on long-context retrieval (MRCR v2)
- 90.2% on BigLaw Bench (legal reasoning) — highest of any Claude model
- 38 out of 40 blind cybersecurity investigations won against Claude 4.5
What’s Actually New
1M Token Context Window (Beta)
Opus 4.6 is the first Opus-class model that can handle a million tokens. Premium pricing kicks in above 200k tokens ($10/$37.50 per million input/output).
Agent Teams in Claude Code
You can now spin up multiple AI agents that work in parallel, coordinate autonomously, and tackle tasks together. Best for codebase reviews and read-heavy work that splits into independent chunks.
Adaptive Thinking
Previously, extended thinking was either on or off. Now the model decides when deeper reasoning would actually help — and developers can tune how selective it is.
Effort Controls
Four levels: low, medium, high (default), max. Dial down if the model is overthinking simple tasks. Dial up for complex problems.
Context Compaction
Long-running tasks used to hit the context window and stall. Now Claude can automatically summarize older context and keep going.
128K Output Tokens
Opus 4.6 can generate outputs up to 128,000 tokens in a single response — no need to break large tasks into multiple requests.
Claude in PowerPoint (Research Preview)
New. Claude can now build presentations from scratch, read your layouts and fonts, and stay on brand. Pairs with the upgraded Claude in Excel for full document workflows.
What Partners Are Saying
The early access feedback is unusually strong:
“Claude Opus 4.6 is the biggest leap I’ve seen in months. I’m more comfortable giving it a sequence of tasks across the stack and letting it run.” — Austin Ray, Staff Software Engineer, Ramp
“Opus 4.6 handled a multi-million-line codebase migration like a senior engineer. It planned up front, adapted its strategy as it learned, and finished in half the time.” — Gregor Stewart, Chief AI Officer, SentinelOne
“Claude Opus 4.6 autonomously closed 13 issues and assigned 12 issues to the right team members in a single day, managing a ~50-person organization across 6 repositories.” — Yusuke Kaji, General Manager AI, Rakuten
“Both hands-on testing and evals show Claude Opus 4.6 is a meaningful improvement for design systems and large codebases. It also one-shotted a fully functional physics engine.” — Eric Simons, CEO, Bolt.new
The Safety Angle
Anthropic says Opus 4.6 shows “an overall safety profile as good as, or better than, any other frontier model in the industry.”
Key points:
- Lowest rate of over-refusals (refusing harmless requests) of any recent Claude model
- Low rates of deception, sycophancy, and misuse cooperation in automated audits
- Six new cybersecurity probes to detect misuse of the model’s enhanced hacking capabilities
- Full system card published with comprehensive testing details
Pricing
Unchanged: $5 per million input tokens / $25 per million output tokens
Premium pricing for 1M context (above 200k tokens): $10/$37.50 per million.
US-only inference available at 1.1× token pricing for compliance-sensitive workloads.
The Bottom Line
Claude Opus 4.6 isn’t a minor version bump. It’s a significant capability jump that:
- Dominates benchmarks across coding, reasoning, search, and professional work tasks
- Finally gives Opus users the million-token context window Sonnet already had
- Ships genuinely new features (agent teams, adaptive thinking, context compaction)
- Maintains or improves safety alignment despite the intelligence gains
Available now on claude.ai, the API, Amazon Bedrock, and Google Cloud Vertex AI.
Model string for developers: claude-opus-4-6
