Anthropic Claude Opus 4.6
by Anthropic
The February 2026 Claude flagship that introduced adaptive thinking and effort control
Claude Opus 4.6 is Anthropic's February 2026 frontier model, still active on the API as a pinned snapshot at $5 per million input and $25 per million output tokens. It was the first Claude release with adaptive thinking and a four-level effort control, pairing a 1M-token context with 128K output for long-running agentic coding, research and professional knowledge work.
Claude Opus 4.6 is Anthropic's February 2026 frontier release, announced on 5 February 2026 and still listed as an Active model on the Claude API under the ID `claude-opus-4-6`, with a tentative retirement date no sooner than 5 February 2027. It was the first Claude model to ship adaptive thinking — where the model reads contextual cues about how much extended reasoning a request warrants — alongside an explicit `effort` parameter offering low, medium, high and max levels, and it pairs a 1M-token context window with up to 128,000 output tokens, rising to 300,000 on the Batch API behind the `output-300k-2026-03-24` beta header. Anthropic reported state-of-the-art results at launch on Terminal-Bench 2.0, Humanity's Last Exam, GDPval-AA and BrowseComp, citing 81.42% on SWE-bench Verified with prompt modification, 53.0% on Humanity's Last Exam with tools, and roughly a 144-Elo lead over GPT-5.2 on GDPval-AA. Independent aggregation is more measured: BenchLM records 80.8% SWE-bench Verified, 91.3% GPQA and 82% MMLU-Pro, ranks it #20 of 214 models overall and #8 of 55 on knowledge, but clocks throughput at about 40 output tokens per second against a field median of 108. Pricing is $5 per million input tokens and $25 per million output — unchanged from Opus 4.5 and identical to today's Opus 5 — and the full 1M context now bills at standard rates rather than the launch-era long-context premium. It runs on the Claude API, Amazon Bedrock, Google Cloud, Microsoft Foundry and Claude Platform on AWS. Anthropic has since shipped Opus 4.7, Opus 4.8 and Opus 5 above it, so 4.6 now serves as a price-stable pinned snapshot rather than the recommended default.
Engineering and research teams that pinned an agent or evaluation harness to a specific early-2026 frontier snapshot and need it to keep behaving identically — plus anyone who still relies on the temperature, top_p and top_k sampling controls that Claude 4.7 and later reject outright.
Frontier-grade agentic coding and long-horizon reasoning across a 1M-token context at $5/$25 per million tokens, on a snapshot Anthropic has committed not to retire before 5 February 2027.
At a Glance
- Category
- AI Models & APIs
- Pricing
- Usage-based, Contact for pricing
- Target Market
- CTOs, CIOs, Enterprise Developers, Data Scientists, AI Research Engineers
- Deployment
- API-based, Cloud-only, Multi-cloud
- Founded
- 2021
- Headquarters
- San Francisco, United States
- Team Size
- 500+
Key Features
- ✓Adaptive thinking
The model infers from context how much extended reasoning a request needs, instead of requiring a manually set thinking budget on every call
- ✓Four-level effort control
Low, medium, high and max settings let you trade latency and token spend against depth explicitly per request
- ✓1M-token context window
Holds roughly 750,000 words in a single request, now billed at standard per-token rates across the whole window
- ✓128K output, 300K in batch
Emits up to 128,000 tokens synchronously and 300,000 via the Batch API extended-output beta header
- ✓Prompt caching
Cache reads bill at $0.50 per million tokens, a tenth of base input price, which matters most for long system prompts
- ✓Data residency via inference_geo
US-only inference is selectable at a 1.1x pricing multiplier, a control unavailable on models older than the 4.6 generation
- ✓Multi-cloud availability
Ships on the Claude API, Amazon Bedrock, Google Cloud, Microsoft Foundry and Claude Platform on AWS with regional endpoints
Capabilities
Use Cases
- •Autonomous agentic coding
Runs multi-hour terminal and repository tasks end to end, the workload Anthropic tuned it for and measured on Terminal-Bench 2.0
- •Financial analysis and research
Anthropic published a dedicated finance-focused Opus 4.6 programme; long context suits filings, transcripts and model documentation in one pass
- •Deep web research
Anthropic reported category-leading BrowseComp results, so multi-hop browsing agents that must chase citations across sites are a natural fit
- •Frozen evaluation baselines
Research teams comparing models over time need a pinned snapshot whose scores stay reproducible for the length of a study
- •Long-document professional knowledge work
Contract review, technical due diligence and specification analysis where the 1M context avoids chunking and retrieval error
Ideal For
Best For
- ✓Long-running autonomous coding agents that operate across a whole workflow rather than a single file — METR estimated a 50%-success time horizon of around 14.5 hours
- ✓Whole-repository and whole-corpus reasoning where the 1M-token context removes the retrieval layer entirely
- ✓Reasoning-heavy analysis where the effort parameter can be pushed to max with a large thinking budget, as Anthropic did for its ARC-AGI-2 runs
- ✓Regression-sensitive production pipelines that need a pinned model ID whose behaviour will not shift under them until at least February 2027
- ✓Codebases still passing temperature, top_p or top_k, which return a 400 error on Claude Opus 4.7 and later
Not Ideal For
- ✗Latency-sensitive interactive products — BenchLM measures about 40 output tokens per second against a 108 field median, and unlike Opus 5 and Opus 4.8 it cannot use Fast mode, which runs at standard speed and standard rates if requested
- ✗Anyone starting fresh today: Opus 4.7, Opus 4.8 and Opus 5 all sit above it at the identical $5/$25 price, so choosing 4.6 for new work buys older capability for the same money
- ✗High-volume or cost-sensitive workloads — at $25 per million output tokens it is five times Haiku 4.5, and thinking tokens bill as output
- ✗Work that depends on recent events: its reliable knowledge cutoff is May 2025, and Simon Willison's system-prompt analysis shows Anthropic had to hard-code current-affairs corrections into the prompt to compensate
Integrations
Deployment
Market Analysis
Pros
- ✓Frontier-tier scores that still hold up months later — 80.8% SWE-bench Verified, 91.3% GPQA and 82% MMLU-Pro on BenchLM's independent leaderboard, ranking #20 of 214 models
- ✓1M-token context with 128K output (300K in batch) removes the retrieval layer for whole-repository and whole-corpus work
- ✓Adaptive thinking plus a four-level effort dial gives per-request control over the accuracy-versus-cost trade rather than a single global setting
- ✓Anthropic has committed not to retire it before 5 February 2027 and gives at least 60 days' notice, which is what a pinned production dependency needs
Cons
- ✗Slow: BenchLM measures about 40 output tokens per second against a 108 field median, and Fast mode — available on Opus 5 and 4.8 — is explicitly unsupported here
- ✗Superseded three times over. Opus 4.7, Opus 4.8 and Opus 5 all sit above it at the exact same $5/$25 price, so there is no economic reason to pick it for new work
- ✗Reliable knowledge cutoff of May 2025 is the weakest of the 4.6-and-later generation; Anthropic's own system prompt had to hard-code corrections about current events to compensate
- ✗The explicit extended-thinking parameter is marked deprecated on this model, so code written against it is already on a migration path
- ✗Reliability wobbles are on the public record — Anthropic's status page logged an 'Elevated errors on Claude Opus 4.6' incident, and practitioners posted a BridgeBench run showing hallucination accuracy dropping from 83% to 68%
- ✗Price pressure from below is real: HN threads point to MiniMax M2.5 claiming comparable SWE-bench performance at roughly 20x lower cost
Pricing
Claude API (pay-as-you-go)
From $5/M input tokens
- ✓$5 per million input tokens
- ✓$25 per million output tokens
- ✓5-minute cache writes $6.25/M, cache reads $0.50/M
- ✓Full 1M context at standard rates
Batch API
From $2.50/M input tokens
- ✓50% discount on both input and output
- ✓$2.50/M input, $12.50/M output
- ✓Supports the 300K extended-output beta header
- ✓Not compatible with Fast mode
Cloud marketplaces
Contact for pricing
- ✓Amazon Bedrock and Google Cloud billed by the provider
- ✓Claude Platform on AWS and Microsoft Foundry bill in Claude Consumption Units at $0.01 per CCU
- ✓US-only inference at a 1.1x multiplier via inference_geo
- ✓Private offers and negotiated volume discounts
Purely token-metered at $5 input and $25 output per million, with no seat licence — the same rate as Opus 4.5, Opus 4.7, Opus 4.8 and Opus 5, which is the key commercial fact: the whole Opus line is priced identically, so 4.6 offers no discount for being a generation behind. The launch-era long-context premium is gone; Anthropic now bills the full 1M window at standard rates, and prompt caching drops reads to $0.50 per million. Batch halves both sides to $2.50/$12.50. Fast mode is not available on 4.6 — requests with `speed: "fast"` simply run at standard speed and standard rates rather than erroring. US-only inference via `inference_geo` adds a 1.1x multiplier across every token category, and regional endpoints on Bedrock or Google Cloud add 10% over global. Nothing is gated behind an Enterprise SKU; custom rate limits and volume discounts are negotiated with sales.
Security & Compliance
Connect
Sources
This page was written from 9 sources, 8 on domains other than anthropic.com.
- 1.anthropic.com — claude opus 4 6vendor
- 2.platform.claude.com — pricing
- 3.platform.claude.com — model deprecations
- 4.platform.claude.com — overview
- 5.benchlm.ai — claude opus 4 6
- 6.openrouter.ai — claude opus 4.6
- 7.hn.algolia.com — hn.algolia.com
- 8.simonwillison.net — opus system prompt
- 9.en.wikipedia.org — Anthropic
Stay Ahead of the Curve
Weekly enterprise AI insights for technology leaders. No spam, no vendor pitches—unsubscribe anytime.
SubscribeRelated Products
Xiaomi MiMo-V2.6
MIT-licensed, 1M-context multimodal MoE models that took the top open-weights score on Artificial Analysis at well under $1 per million tokens
TypeSafe Jev
A 'System One' decision model that returns typed choices, scores and calibrated probabilities instead of text, for agent routing and classification
PrismML Bonsai
Open-weight 1-bit and ternary LLMs that run 27B-class AI on laptops and phones
Sakana AI Fugu Max
Multi-agent orchestration behind one API — frontier-grade results at $2 per million input tokens
Mentioned In
AWS vs GCP vs Azure ML: The Real Costs Nobody Tells You
Enterprise AI analysis: AWS vs GCP vs Azure. Strategic insights, ROI considerations, and implementation guidance for technical and business leaders evaluatin...
March 15, 2026AnthropicAnthropic Leaks Claude Mythos in CMS Failure: The Most Capable AI Model—And Biggest Cyber Risk—They've Ever Built
Anthropic's CMS misconfiguration exposed Claude Mythos—a new AI model tier called Capybara that's 'far ahead of any other AI model in cyber capabilities.' The leak revealed draft launch plans, cybersecurity risk assessments, and an invite-only CEO summit. Enterprise security teams need to understand what ' unprecedented cybersecurity risks' means for their threat models.
March 29, 2026