Claude Opus 4.6

Supported Functionality

Item	Specification
Input	Text, Image
Output	Text
Context	200,000 tokens (1,000,000 in beta)
Max Output	128,000 tokens
Vision	✓ Supported
Function Calling	✓ Supported (Agent Teams, Context Compaction, MCP)

Description

Claude Opus 4.6 is Anthropic's frontier flagship model, released on February 5, 2026 and available via the Claude API as claude-opus-4-6. It is the first Opus-class model to ship a 1M-token context window (beta) and the first Claude model with Adaptive Thinking plus four effort levels (low / medium / high / max), producing up to 128K output tokens.

Opus 4.6's defining breakthrough is turning frontier reasoning into sustained, reliable long-horizon execution. It tops Terminal-Bench 2.0 (65.4%), Humanity's Last Exam (53.0% with tools), ARC-AGI-2 (69.17%), ARC-AGI-1 (94.00%), and BrowseComp (84.0%); it finishes Vending-Bench 2's year-long simulation with a final balance of $8,017 (well above Gemini 3 Pro's prior SOTA); and on MRCR v2 8-needle 1M retrieval it scores 76% versus Sonnet 4.5's 18.5%. The release also introduces Agent Teams in Claude Code (parallel multi-agent collaboration), Context Compaction on the API, a major Claude in Excel upgrade, and a Claude in PowerPoint research preview — a full lift for enterprise workflows.

Key Capabilities

• Top-Tier Agentic Coding: #1 on Terminal-Bench 2.0 at 65.4%; SWE-Bench Verified 80.8%; handles cross-file refactors and complex bug diagnosis autonomously.
• Frontier Deep Reasoning: HLE 53.0% (with tools), ARC-AGI-2 69.17%, ARC-AGI-1 94.00% — leads on expert-level multi-disciplinary and novel-problem reasoning.
• Long-Context Retrieval: MRCR v2 8-needle 1M at 76% (Sonnet 4.5: 18.5%) — dramatically reduced context rot, supporting deep reasoning across full repos and large document sets.
• Adaptive Thinking + Effort Control: The model auto-tunes reasoning depth; developers further adjust via low / medium / high / max for precise quality/latency/cost tradeoffs.
• Agent Teams in Claude Code: Multiple Claude instances collaborate in parallel on a single task — ideal for code review and read-heavy large-codebase work.
• Enterprise Knowledge Work: GDPval-AA +190 Elo over Opus 4.5 and +144 over GPT-5.2; BigLaw Bench legal reasoning 90.2%; BrowseComp agentic search 84.0%.
• Life Sciences & Cybersecurity: ~2× improvement over Opus 4.5 on computational/structural biology, organic chemistry, and phylogenetics; industry-leading on CyberGym vulnerability discovery.

Technical Strengths

Feature	Benefit
First 1M Context in the Opus Tier	Frontier-grade reasoning over entire codebases, hundreds of contracts, or full research corpora — no RAG plumbing required
Adaptive Thinking + Effort Levels	Model auto-tunes reasoning depth and developers get four explicit dials for quality/latency/cost trade-offs
Context Compaction (API)	Automatic context summarization lets very long agentic runs continue beyond raw token limits
Agent Teams Orchestration	Multiple Claude agents run in parallel inside Claude Code, accelerating large-scale and read-heavy engineering
Sustained Coherence	Vending-Bench 2 final balance of $8,017 demonstrates strategic planning and self-correction across thousands of decisions
Anthropic Safety & Alignment	Automated audits show low rates of deception, sycophancy, and delusion reinforcement — capability gains without safety regressions

Capability Ratings

Dimension	Rating	Notes
Reasoning	Top-tier	Leads HLE, ARC-AGI-2, and BigLaw Bench — undisputed frontier expert reasoning
Coding	Top-tier	#1 on Terminal-Bench 2.0; SWE-Bench Verified 80.8%; strongest at complex cross-file refactors
Creative Writing	Top-tier	Anthropic's signature voice with strong pacing, restraint, and cultural depth
Multimodal	Strong	Text + image input; no native video/audio (later improved in Opus 4.7)
Response Speed	Moderate	Flagship priced for depth over latency; Low effort mode trades reasoning for speed
Context Window	Huge	200K default and 1M beta; MRCR long-context retrieval is best-in-class among contemporary models

Use Cases

• Complex Codebase Refactoring & Diagnosis: #1 Terminal-Bench plus 1M context enables whole-repo analysis, multi-file refactors, and tough concurrency bug fixes.
• Multi-Agent Engineering Workflows: Claude Code Agent Teams let multiple Opus instances review, write, and test in parallel for large engineering projects.
• Enterprise Research & Deep Retrieval: 84.0% on BrowseComp plus long context make it ideal for multi-step research, cross-document synthesis, and intelligence work.
• Legal / Finance / Life Sciences: BigLaw Bench 90.2%, Real-World Finance SOTA, and near-2× life sciences gains support high-stakes professional agents.
• Long-Horizon Strategic Planning: Vending-Bench 2 results show sustained coherence over thousands of decisions — useful for business simulation and operational planning.
• Office Productivity: Upgraded Claude in Excel and Claude in PowerPoint research preview extend Opus into spreadsheets, documents, and slide decks.
• Highest-Difficulty Research Assistant: SOTA on HLE, GPQA, and ARC-AGI-2 makes it a reliable partner on frontier math, physics, biology, and CS problems.