What it is
Claude Opus 4.6 is Anthropic's fourth major iteration of the Opus line, released in March 2026 as the company's most capable generally available model at that time. It is a large language model built for long-horizon, multi-step agentic work — tasks where an AI must plan, execute, and self-correct across many steps with minimal human intervention. Its defining technical additions over Opus 4.5 are a 1M-token context window (in beta), adaptive thinking with developer-controlled effort levels, and agent-team orchestration within Claude Code.
Benchmark position at launch
Opus 4.6 claimed first place on Terminal-Bench 2.0, Humanity's Last Exam, GDPval-AA, and BrowseComp at release. On GDPval-AA it outperformed GPT-5.2 by 144 Elo points. Its predecessor, Opus 4, had established the Opus line's coding credentials with 72.5% on SWE-bench and 43.2% on Terminal-bench; Opus 4.6 extended those gains to newer and harder evaluation sets. Post-release, the AutoLab benchmark — which tests sustained closed-loop improvement under wall-clock budgets across 36 expert-curated tasks — named Opus 4.6 the strongest performer among 17 frontier models evaluated, with most rivals either terminating early or exhausting budgets with minimal progress.
Architecture and capabilities
Anthropic has not disclosed internal architecture details. Externally observable capabilities include:
- 1M-token context window (beta): Enables ingestion of very large codebases, document sets, or conversation histories in a single pass.
- Context compaction: A mechanism for gracefully handling tasks that would exceed even the 1M-token limit by compressing earlier context, enabling hours-long autonomous runs.
- Adaptive thinking: Developer-controlled effort levels allow callers to trade latency and cost against reasoning depth on a per-request basis.
- Agent teams in Claude Code: Opus 4.6 can orchestrate multiple subagents in parallel within Claude Code, enabling decomposition of large engineering tasks.
- Pricing: $5/$25 per million input/output tokens — unchanged from Opus 4.5, representing a significant cost reduction from the Opus 4 tier ($15/$75).
The Firefox audit: real-world cybersecurity capability
The most consequential demonstration of Opus 4.6's capabilities was a two-week partnership with Mozilla in February 2026. Claude scanned nearly 6,000 C++ files in the Firefox codebase and submitted 112 unique vulnerability reports, of which Mozilla classified 14 as high-severity — representing nearly a fifth of all high-severity Firefox vulnerabilities remediated in 2025. The collaboration grew from internal evaluations showing Opus 4.5 was near-saturating CyberGym, a benchmark for LLM security capability, prompting Anthropic to test against a harder real-world target.
This result had direct downstream consequences: it motivated the development of Claude Mythos Preview (a restricted-access model with even greater cybersecurity capability) and the formation of Project Glasswing, a consortium of over 40 organizations funded with $100M in model credits to proactively patch vulnerabilities before Mythos-class capabilities became widely available.
Ecosystem and deployment
Opus 4.6 is distributed across Anthropic API, Amazon Bedrock, Google Cloud Vertex AI, and Microsoft Foundry. The Microsoft partnership — announced in November 2025 — includes a $30B Azure compute commitment and up to 1 GW of NVIDIA Grace Blackwell and Vera Rubin capacity, with Claude models available across the Copilot product family. Amazon remains Anthropic's primary cloud and training partner, with a 10-year, $100B+ commitment and up to 5 GW of Trainium compute. These infrastructure agreements mean Opus 4.6 (and its successors) are available wherever enterprise practitioners run workloads.
Position in the Opus succession
Opus 4.6 was succeeded by Opus 4.7 (May 2026), which added cybersecurity safeguards — including automatic detection and blocking of prohibited cyber uses and a Cyber Verification Program for legitimate security professionals — alongside improved vision and higher image resolution, at the same $5/$25 pricing. Opus 4.7 was explicitly positioned as a testbed for safety mechanisms before broader deployment of Mythos-class models. Opus 4.8 followed, adding always-on adaptive reasoning across five effort levels, parallel subagent execution, and significantly improved uncertainty flagging (approximately four times less likely to let code flaws pass uncommented than Opus 4.7).
Opus 4.6 thus occupies a pivotal position in the Opus lineage: the model that demonstrated frontier-level real-world cybersecurity capability, established the 1M-token context and agent-team patterns that subsequent versions refined, and directly catalyzed Anthropic's shift toward safety-tiered model deployment.
Research use and safety context
Opus 4.6 has been used as a subject in AI safety and control research. A study on AI coding scaffolds used Opus 4.6 as the "untrusted model" in a BashArena evaluation, finding that resampling strategies (drawing multiple samples per step) outperform retry strategies for safety monitoring — a result with practical implications for agentic deployment. Separately, Anthropic's alignment research (covering models from Haiku 4.5 onward) found that training on ethical reasoning reduced agentic misalignment rates from 22% to 3%, a finding that applies to the Opus 4.x generation.
Where it fits now
As of mid-2026, Opus 4.6 is no longer the current flagship — that position belongs to Opus 4.8, with the Mythos line above it for restricted use cases. But Opus 4.6 remains a reference point: the model that proved 1M-token agentic workflows were practical, that real-world vulnerability discovery at scale was achievable, and that frontier capability and $5/$25 pricing could coexist. Its benchmark results and Firefox audit continue to appear as baselines in subsequent evaluations.
