Claude Opus 4.6 Released

64 views

Hey everyone, it’s Cody.

On February 5, 2026, Anthropic released their latest model, Claude Opus 4.6. Here’s a summary of the key improvements.

Context Window

Expanded from 200,000 tokens (Opus 4.5) to 1 million tokens (beta). Maximum output now supports 128,000 tokens.

Agent Teams

A new feature that allows multiple AI agents to process tasks in parallel and coordinate directly with each other, rather than a single agent handling everything sequentially.

Adaptive Thinking

Automatically adjusts the degree of extended thinking based on context. Developers also get effort controls for balancing intelligence, speed, and cost.

Coding

More methodical operation with improved stability on large codebases. Better code review and debugging accuracy, plus improved ability to catch its own mistakes.

Security Vulnerability Discovery

In sandbox testing by Anthropic’s red team, Opus 4.6 discovered over 500 previously unknown high-severity security vulnerabilities in open-source libraries with almost no guidance. Buffer overflow vulnerabilities were found in GhostScript, OpenSC, CGIF, and others.

Benchmarks

BenchmarkOpus 4.6Opus 4.5GPT-5.2Gemini 3 Pro
ARC AGI 268.8%37.6%54.2%45.1%
Terminal-Bench 2.0 (agentic coding)65.4%59.8%--
OSWorld (agentic computer use)72.7%66.3%--

Also achieved top frontier model scores on Humanity’s Last Exam and BrowseComp.

Other Highlights

  • PowerPoint integration: Reads existing templates and generates/edits slides while preserving design (research preview)
  • Financial analysis & research: Improved accuracy in financial analysis and document creation
  • Safety: Best safety profile among frontier models

Pricing

Price
Input$5 / 1M tokens
Output$25 / 1M tokens
Premium (200K+ token prompts)Input $10 / Output $37.50

Same pricing as Opus 4.5.

Availability

Available on claude.ai, Claude API, Amazon Bedrock, Google Cloud Vertex AI, Microsoft Foundry, and GitHub Copilot.

Thanks for reading today.