Claude Opus 4.6: Features, Benchmarks, and Pricing Guide

What's New in Claude Opus 4.6

At launch, Opus 4.6 was Anthropic's most capable model, positioned as the successor to Opus 4.5, released in November 2025. The update focuses on three areas: reasoning depth through adaptive thinking, context capacity with the 1M token beta, and agentic task execution, where it led most of the evaluations in Anthropic’s launch table.

Opus 4.5 vs Opus 4.6 at a Glance

The pricing parity with Opus 4.5 is notable — Anthropic delivered substantially more capability at the same cost, except for the 1M context window, which carried a premium tier above 200K tokens until March 13, 2026.

Adaptive Thinking Mode Explained

Adaptive thinking replaces extended thinking as the recommended reasoning mode for Opus 4.6. Instead of manually setting a budget_tokens parameter, Claude now dynamically decides when and how much to reason based on the complexity of each request.

Four Effort Levels

Migration from Extended Thinking

// Before (Opus 4.5 — deprecated on 4.6)
const response = await anthropic.messages.create({
  model: "claude-opus-4-5-20251101",
  max_tokens: 16000,
  thinking: { type: "enabled", budget_tokens: 10000 },
  messages: [{ role: "user", content: "Solve this problem..." }]
});

// After (Opus 4.6 — recommended)
const response = await anthropic.messages.create({
  model: "claude-opus-4-6",
  max_tokens: 16000,
  thinking: { type: "adaptive" },
  // Optional: control depth with the effort parameter
  // output_config: { effort: "max" },  // low | medium | high (default) | max
  messages: [{ role: "user", content: "Solve this problem..." }]
});

Adaptive thinking also automatically enables interleaved thinking, which means Claude can reason between tool calls without the previously required interleaved-thinking-2025-05-14 beta header. If you’re still passing that header, it is no longer needed and should be removed.

Benchmark Performance Analysis

Opus 4.6 leads on several key benchmarks, particularly in agentic coding and economically valuable knowledge work. For a broader competitive landscape, see our Claude vs GPT-5.2 vs Gemini 3 Pro comparison.

Agentic and Coding Benchmarks

Source: the benchmark table in Anthropic’s Claude Opus 4.6 announcement (February 5, 2026) and the Opus 4.6 system card. Like Anthropic’s table, the GPT-5.2 and Gemini 3 Pro columns show the best reported variant of each; “self-reported” marks the vendor’s own figure where Anthropic’s run differed. BrowseComp reflects Anthropic’s March 2026 system card revision (first published as 84.0%). Anthropic labels Gemini’s 59.2% BrowseComp score as Gemini Deep Research; Google’s own evaluation table lists the same 59.2% for Gemini 3 Pro.

Reasoning and Knowledge Benchmarks

Source: Anthropic’s announcement table and system card. GDPval-AA was run independently by Artificial Analysis. HLE with tools reflects Anthropic’s February 2026 revision (first published as 53.1%). GPT-5.2 Pro figures match OpenAI’s GPT-5.2 announcement, and Gemini 3 Pro’s HLE, GPQA Diamond and MMMLU figures match Google’s Gemini evaluation table. Google reports 31.1% on ARC-AGI-2 for Gemini 3 Pro; the 45.1% that Anthropic’s table showed in the Gemini column is Gemini 3 Deep Think. BigLaw Bench is Harvey’s result, quoted in Anthropic’s announcement, with no comparable figure for other vendors’ models.

A notable result: Opus 4.5 slightly edges out Opus 4.6 on SWE-bench Verified (80.9% vs 80.8%), while GPT-5.2 with Codex CLI reaches 64.7% on Terminal-Bench 2.0 — just 0.7 percentage points behind Opus 4.6. The ARC AGI 2 result is particularly striking, with Opus 4.6 nearly doubling Opus 4.5's score (68.8% vs 37.6%).

1M Context Window and Compaction API

The 1M token context window is the headline expansion for Opus 4.6. In beta at launch and available only on the Claude Developer Platform, it represents a 5x increase over the standard 200K window.

Long-Context Retrieval Performance

On the MRCR v2 benchmark (8-needle), Opus 4.6 scores 93% at 256K context and 76% at 1M context, according to Anthropic’s system card. For comparison, Sonnet 4.5 manages only 10.8% at 256K and 18.5% at 1M — making Opus 4.6 roughly 4-9x more reliable at retrieving information from deep in long contexts. Anthropic describes this as a qualitative shift in reducing context rot.

Compaction API for Infinite Conversations

The Compaction API (beta) provides server-side context summarization. When a conversation reaches a configurable token threshold, the API automatically summarizes older parts of the conversation. This enables effectively infinite conversations without manual context management, sliding window hacks, or truncation strategies.

For agent-based workflows that involve many tool calls and long chains of reasoning, compaction can significantly reduce the overhead of maintaining conversation state. The summarization happens server-side, so there are no additional API calls required from your application.

Pricing and Availability

Opus 4.6 maintains the same base pricing as Opus 4.5, making it a direct upgrade with no cost increase for existing workloads under 200K tokens.

Source: Anthropic’s launch announcement and API pricing docs. Prices as of February 2026; the long-context premium was dropped on March 13, 2026.

Platform Availability

• Anthropic API: Available now with model ID claude-opus-4-6
• AWS Bedrock: Available at launch with regional and global endpoints
• Google Vertex AI: Available at launch
• Microsoft Foundry: Available at launch
• claude.ai: Available at launch in Anthropic’s Claude apps

API Changes and Migration Guide

Opus 4.6 introduces both deprecations and a notable breaking change. If you're upgrading from Opus 4.5 or Sonnet 4, review these carefully before switching your model ID.

Breaking: Prefilling Disabled

Prefilling — where you start the assistant's response with specific text like {"role": "assistant", "content": "Here is the JSON:"} — is no longer supported. Alternatives include:

• Structured outputs: Use output_config.format with json_schema for guaranteed JSON output
• System prompt instructions: Guide response style and format through the system message
• JSON output mode: Use output_config.format for general JSON responses

Deprecations

• Extended thinking (type: "enabled" + budget_tokens): Still functional but deprecated. Migrate to adaptive thinking.
• Interleaved thinking beta header: No longer needed, because adaptive thinking enables interleaved thinking automatically. Remove from requests.
• output_format parameter: Moved to output_config.format. Old parameter still works but is deprecated.

New Features

• Effort parameter (GA): No longer requires a beta header. Combine with adaptive thinking for cost-quality tradeoffs.
• Fine-grained tool streaming (GA): Now generally available on all models and platforms without a beta header.
• inference_geo parameter: Request US-only inference with a 1.1x pricing multiplier for data residency compliance.
• 128K output tokens: Double the previous 64K limit. SDKs require streaming for large max_tokens to avoid HTTP timeouts.

Claude Code and Agent Teams

Alongside the model release, Anthropic launched several product updates that leverage Opus 4.6's improved agentic capabilities. For a deeper look at Claude Code's impact on development workflows, see our Claude Code development guide.

Agent Teams (Research Preview)

Claude Code now supports agent teams — multiple autonomous agents that coordinate in parallel on complex tasks. This enables workflows where one agent handles frontend changes, another tackles backend logic, and a third manages tests, all working simultaneously under supervisory control.

For organizations exploring how Anthropic's enterprise tools fit into larger workflows, our Cowork plugins enterprise guide covers the broader plugin ecosystem.

Office Integrations

Enterprise and Safety Features

Opus 4.6 introduces data residency controls and maintains what Anthropic describes as the lowest over-refusal rates among recent Claude models. For industries with strict compliance requirements, these updates address key adoption barriers.

Data Residency Controls

The new inference_geo parameter lets you specify where model inference runs on a per-request basis. Setting it to "us" guarantees US-only processing at a 1.1x pricing multiplier. The default "global" routing uses standard pricing.

Safety Profile

• Low misalignment score: On Anthropic’s automated behavioral audit, Opus 4.6 scores approximately 1.8 on a 1–10 overall misaligned-behavior scale, the lowest bar on Anthropic’s chart, with error bars overlapping Opus 4.5’s (~1.9), ahead of Haiku 4.5 (~2.2), Sonnet 4.5 (~2.75), and Opus 4.1 (~4.3). Anthropic describes it as just as well-aligned as Opus 4.5
• Lowest over-refusal rates: Among recent Claude models, Opus 4.6 is less likely to refuse legitimate requests while maintaining appropriate safety boundaries
• Cybersecurity: Anthropic developed six new cybersecurity probes to detect potential misuse. Separately, early-access customer NBIM (Norges Bank Investment Management) reported that Opus 4.6 produced the best result in 38 of 40 cybersecurity investigations in a blind ranking against Claude 4.5 models

For legal and compliance teams, Harvey’s report that Opus 4.6 scored 90.2% on its BigLaw Bench, the highest of any Claude model, is particularly relevant. For a detailed look at Claude's legal capabilities, see our Claude legal plugin analysis.

Opus 4.6 vs the Competition

Here's how Opus 4.6 compares to its closest competitors across key criteria as of February 2026:

Sources: Anthropic’s announcement and docs; OpenAI’s GPT-5.2 model page (400K context, 128K output, $1.75 / $14); Google’s Gemini 3 developer announcement and Gemini 3 Pro model page (1M input, 64K output, $2 / $12 up to 200K tokens). Google shut down Gemini 3 Pro Preview on March 9, 2026 and moved developers to Gemini 3.1 Pro Preview. Benchmark rows as in the tables above.

Opus 4.6 leads in knowledge work (GDPval-AA), agentic search (BrowseComp), and multidisciplinary reasoning (Humanity’s Last Exam). GPT-5.2 edges ahead on graduate-level reasoning (GPQA Diamond), is within 0.7 points on Terminal-Bench 2.0, and costs far less per token. Gemini 3 Pro offers a 1M-token window as standard, where Opus 4.6’s was still a beta at launch, and leads on visual reasoning (MMMU Pro, without tools). The right model depends on your specific use case, latency requirements, and budget.

What This Means for Your AI Strategy

Claude Opus 4.6 represents a meaningful step forward in reasoning depth, context handling, and agentic task execution. The adaptive thinking mode simplifies the developer experience, the compaction API opens the door to persistent agent conversations, and the benchmark results suggest Anthropic is pulling ahead in coding and knowledge work tasks.

The most immediate action item for existing Claude API users is the prefilling removal — check your codebase for assistant message prefills and migrate to structured outputs before switching to the new model ID. For teams evaluating Claude for the first time, the pricing parity with Opus 4.5 means there's no cost reason to start with the older model.

Frequently Asked Questions