What it is
Claude Opus 4.6 is Anthropic's flagship large language model released in March 2026, succeeding Claude Opus 4.5 in the Opus 4.x line. It is the first Opus model to ship with a 1M-token context window (in beta), adaptive thinking with developer-controlled effort levels, and a suite of agentic features — agent teams in Claude Code, context compaction for long-running tasks, and a research preview of Claude in PowerPoint — designed for sustained, multi-step autonomous work.
Benchmark position at launch
Anthropic claimed top scores on four major evaluations at release: Terminal-Bench 2.0, Humanity's Last Exam, GDPval-AA (by 144 Elo over GPT-5.2), and BrowseComp. These results extended the gains established by Claude Opus 4 (72.5% SWE-bench, 43.2% Terminal-bench) and Opus 4.5, which had already positioned the Opus line as the leading tier for software engineering tasks. Pricing was held at $5/$25 per million input/output tokens — the same level introduced with Opus 4.5, a significant reduction from Opus 4's $15/$75.
Post-launch, independent research reinforced the agentic framing: the AutoLab benchmark — 36 expert-curated tasks measuring iterative persistence under wall-clock budgets rather than single-turn quality — identified Opus 4.6 as the strongest performer across 17 frontier models evaluated, with most competitors either terminating early or exhausting budgets with minimal progress.
Real-world cybersecurity capability
The most consequential demonstration of Opus 4.6's capabilities came from a two-week partnership with Mozilla in February 2026. The model scanned nearly 6,000 C++ files in the Firefox codebase, submitted 112 unique vulnerability reports, and identified 22 confirmed vulnerabilities — 14 classified as high-severity, representing nearly a fifth of all high-severity Firefox vulnerabilities remediated in 2025. Most issues were patched in Firefox 148.0. 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 directly motivated Project Glasswing: when Anthropic subsequently published the Claude Mythos Preview model card — benchmarking the new model substantially above Opus 4.6 on CyberGym (83.1%), Terminal-Bench 2.0 (82%), GPQA Diamond (94.5%), and HLE (64.7%) — it assembled a 40+ organization consortium to proactively patch vulnerabilities before broader deployment.
Architecture and capabilities
The events bundle does not disclose internal architecture. Externally observable: Opus 4.6 supports hybrid near-instant and extended thinking modes with developer-adjustable effort, parallel tool execution, local file access for persistent memory across sessions, and context compaction — a mechanism for gracefully handling tasks that approach or exceed the 1M-token window by compressing earlier context. These features are designed for long-horizon agentic runs where the model must maintain coherent state across hours of autonomous execution.
Ecosystem and deployment
Opus 4.6 is available via the Anthropic API, Amazon Bedrock, Google Cloud Vertex AI, and Microsoft Foundry. The Microsoft partnership — announced in November 2025 — committed Anthropic to $30B of Azure compute and up to 1 GW of NVIDIA Grace Blackwell/Vera Rubin capacity, with Claude models (including Sonnet 4.5, Opus 4.1, and Haiku 4.5) available across the Copilot product family. Amazon remains Anthropic's primary cloud and training partner. Claude Code, which became generally available with the Opus 4 launch, uses Opus 4.6 as its primary model for autonomous coding sessions and integrates with GitHub Actions, VS Code, and JetBrains.
Position in the broader Opus 4.x lineage
Opus 4.6 occupied the frontier for roughly seven weeks before being succeeded by Claude Opus 4.7 (released May 4, 2026), which added cybersecurity classifiers and a Cyber Verification Program as the first model to receive safety mechanisms developed in response to Project Glasswing. Opus 4.7 was explicitly positioned below Claude Mythos Preview in overall capability, serving as a testbed for safety mechanisms before broader Mythos-class deployment. Opus 4.8 followed, adding always-on adaptive reasoning across five effort levels, parallel subagent execution, and topping the Artificial Analysis Intelligence Index.
In research contexts, Opus 4.6 continued to appear as a reference model after its commercial succession: it was used as the "untrusted model" in AI control research on coding scaffolds (BashArena benchmark), and a 32B open-weight model trained with the AutoMem memory framework was benchmarked against Opus 4.5 and Gemini 3.1 Pro Thinking as frontier comparators for long-horizon agentic tasks.
Where it fits now
As of mid-2026, Opus 4.6 has been superseded by Opus 4.8 as the current mid-tier Opus model and by the Mythos/Fable 5 line at the frontier. Its significance is durable in two respects: it established the 1M-token + adaptive-effort pattern that carried through the Opus 4.x line, and its Firefox vulnerability work provided the empirical foundation for Anthropic's most consequential safety initiative to date.
