Anthropic today unveiled Claude Opus 4.6, its flagship model, just three months after Opus 4.5. Now available on claude.ai, Anthropic API, Amazon Bedrock, and Google Cloud Vertex AI, pricing remains unchanged at $5 per million input tokens and $25 per million output tokens.
1M Token Context Window (Beta)
Opus 4.6 introduces a 1 million token context window in a flagship model for the first time — five times the 200k limit of Opus 4.5. This equates to roughly 1,500 pages of text, 30,000 lines of code, or over an hour of video, eliminating the need for chunking or summarization. In the MRCR v2 long-context retrieval test with a million-token variant, Opus 4.6 scored 76% accuracy, up from the prior Sonnet 4.5's 18.5%.
Adaptive Thinking and 128k Output
The new adaptive thinking mechanism lets the model decide when to deep-reason or respond quickly based on task complexity. Developers can tune the effort parameter across low, medium, high, and maximum levels to balance intelligence, speed, and cost. Single response output tops out at 128,000 tokens, enabling generation of complete code modules or long reports in one go.
Agent Teams: Multiple Claudes Collaborate
The most disruptive feature is Agent Teams — multiple Claude agents working in parallel on subtasks, syncing via a shared task directory. Anthropic researcher Nicholas Carlini used the system to have multiple Claudes collaboratively build a C compiler from scratch. The result: a 100,000-line Rust codebase that successfully compiled the Linux 6.9 kernel, supports x86, ARM, and RISC-V architectures, passes 99% of standard tests, and compiles QEMU, FFmpeg, SQLite, and Postgres. The experiment consumed roughly 2,000 Claude Code sessions and $20,000 in API costs — proving autonomous large-scale development is now an engineering reality.
500 Zero-Day Vulnerabilities Discovered Autonomously
Before release, Anthropic placed Opus 4.6 in a sandbox with Python, a debugger, and fuzzing tools — no specific instructions. The model autonomously uncovered over 500 previously unknown security flaws, including a crash bug in GhostScript, buffer overflows in OpenSC, and a buffer overflow in CGIF for which it wrote a proof-of-concept exploit. Each vulnerability was verified by internal or external researchers. Anthropic added real-time detection systems to Opus 4.6 to prevent potential misuse.
Benchmark Leadership
In the independent GDPval-AA evaluation, Opus 4.6 outperformed OpenAI GPT-5.2 by approximately 144 Elo points and its predecessor Opus 4.5 by 190 Elo points. It also set the highest score in the agentic coding benchmark Terminal-Bench 2.0. Anthropic describes the model as “more deliberate in planning, capable of sustained longer agentic tasks, more reliable in large codebases, and stronger at code review and debugging.”
Product Integrations: Excel Upgraded, PowerPoint Added
Claude in Excel gets an enhanced version supporting long-running tasks and multi-step changes — users can now refactor entire spreadsheets. Claude in PowerPoint (research preview) reads design systems, maintains brand consistency, and generates full presentations. Excel integration and Cowork are available to all paid tiers; PowerPoint integration is prioritized for Max, Team, and Enterprise plans.