Anthropic has released Claude Opus 4.6, describing it as an upgrade to its “smartest model” with stronger coding performance and longer-running agentic work. The company says Claude Opus 4.6 is available on claude.ai, through its API, and on major cloud platforms, and Amazon Web Services says the model is also available in Amazon Bedrock.
Anthropic says Opus 4.6 improves on its predecessor’s coding skills by planning more carefully, sustaining agentic tasks for longer, operating more reliably in larger codebases, and improving code review and debugging to catch its own mistakes. AWS characterizes Claude Opus 4.6 as Anthropic’s most powerful model yet for autonomous coding, enterprise agents, and professional work.
New tools for agent workflows
In Claude Code, Anthropic says users can assemble “agent teams” so multiple agents can work in parallel on tasks and coordinate autonomously, which it frames as best for work that splits into independent, read-heavy tasks like codebase reviews. Anthropic describes agent teams in Claude Code as a research preview and says users can take over subagents directly using Shift+Up/Down or tmux.
On the API, Anthropic says Claude can use “compaction” to summarize its own context so it can run longer tasks without running into limits. Anthropic also says it is introducing “adaptive thinking,” where the model decides when deeper reasoning is useful, along with “effort” controls so developers can choose between low, medium, high (default), and max to balance intelligence, speed, and cost.
Longer context and bigger outputs
Anthropic says Opus 4.6 is the first Opus-class model to offer a 1M token context window in beta, and it says this 1M context is available in beta on the Claude Developer Platform only. AWS says Opus 4.6 supports both 200K and 1M context tokens, with 1M listed as preview.
Anthropic says Opus 4.6 supports outputs of up to 128K tokens, enabling larger-output tasks without splitting them into multiple requests. A separate developer-oriented summary notes Opus 4.6 doubles maximum output tokens to 128K from 64K and describes a Compaction API (beta) for server-side context summarization to support long conversations.
Anthropic also says US-only inference is available for workloads that need to run in the United States, priced at 1.1× token pricing. The developer summary similarly describes data residency controls via an inference_geo parameter, including a “us” option priced at 1.1× on Opus 4.6 and newer models.
Performance claims and safety testing
Anthropic says the model is state-of-the-art on several evaluations, including the highest score on Terminal-Bench 2.0 and a lead among frontier models on Humanity’s Last Exam. Anthropic also says Opus 4.6 outperforms the next-best model on GDPval-AA by around 144 Elo points and outperforms its predecessor Claude Opus 4.5 by 190 points.
For long-context performance, Anthropic says Opus 4.6 improves at retrieving relevant information from large document sets and performs better as conversations grow, addressing what it calls “context rot.” As one example, Anthropic reports that on the 8-needle 1M variant of MRCR v2, Opus 4.6 scores 76% while Sonnet 4.5 scores 18.5%.
Anthropic says its system card shows Opus 4.6 has an overall safety profile as good as, or better than, other frontier models, with low rates of misaligned behavior in its safety evaluations. It also says Opus 4.6 is as well-aligned as Opus 4.5 and has the lowest rate of over-refusals of any recent Claude model, while adding it ran its most comprehensive set of safety evaluations for this release.
Availability, pricing, and Bedrock access
Anthropic says developers can use the claude-opus-4-6 model via the Claude API and that pricing remains $5/$25 per million tokens. For prompts exceeding 200K tokens, Anthropic says premium pricing applies at $10/$37.50 per million input/output tokens, available only on the Claude Developer Platform.
AWS says adding Claude Opus 4.6 to Amazon Bedrock expands customer model choice and is intended to simplify building applications with enterprise-grade security and responsible AI controls. In the same post, AWS includes a quote from Anthropic’s Kate Jensen: “Claude models consistently set new standards in coding, advanced reasoning, and multi-step workflows while understanding full business contexts and delivering precise results,” and adds, “The real breakthrough is freeing your talent for strategic work while Claude handles the heavy lifting.”
