Claude Sonnet 4.6 review, pricing and limits

Claude Sonnet 4.6 replaces Claude Opus 4.5 for most agentic coding and computer-use workloads at Sonnet-tier cost. Teams running high-volume coding agents can drop a tier without a real quality loss, since it lands within 1.2 points of Opus 4.6 on SWE-bench Verified. It is the wrong pick for graduate-level scientific reasoning tasks.

Released in February 2026, Claude Sonnet 4.6 sits between Anthropic's lighter and flagship model tiers, below Claude Opus 4.6. It scores 89% on the MATH benchmark and matches Opus-tier results on computer-use automation, while its context spans 1 million tokens with no added per-token cost.

Provider: Anthropic · Family: Claude 4

More about Anthropic on HokAI

Context window: 1,000,000 tokens · Max output: 64,000

Input modalities: text, image, pdf, tool-calls · Output: text, tool-calls

About Claude Sonnet 4.6

Anthropic released Claude Sonnet 4.6 on February 17, 2026 as its mid-tier flagship model. It sits in the Claude 4 family between the budget-oriented Haiku 4.5 and the research-grade Opus 4.6, and it replaced Sonnet 4.5 as the default model for Free and Pro users on claude.ai. The model uses a dense transformer architecture; Anthropic has not disclosed the parameter count, though third-party estimates place it in the 50-100 billion range. Sonnet 4.6 was designed to collapse the performance gap between Anthropic's mid-tier and flagship tiers, delivering near-Opus results on the tasks most development teams care about: agentic coding, long-document analysis, and computer use automation.

On benchmark evaluations, Claude Sonnet 4.6 scores 79.6% on SWE-bench Verified, matching Claude Opus 4.6 (80.8%) within 1.2 points and placing it ahead of Claude Sonnet 4.5 (77.2%) by 2.4 points. On GPQA Diamond, the model scores 74.1%, compared to 91.3% for Opus 4.6 and 73.8% for GPT-5.2, a meaningful gap in graduate-level scientific reasoning tasks that tips decisions toward Opus when deep domain expertise matters. On OSWorld (computer use), Sonnet 4.6 scores 72.5%, nearly matching Opus 4.6 at 72.7% and far ahead of GPT-5.2 at 38.2%. On MATH, the model achieves 89% accuracy, up from 62% on Sonnet 4.5. On ARC-AGI-2, Sonnet 4.6 scores 60.4%.

Claude Sonnet 4.6's context window spans 1 million tokens, a size that became generally available on March 13, 2026, meaning no beta header is required and no long-context premium applies. A 900,000-token request bills at the same per-token rate as a 9,000-token request. The synchronous API supports up to 64,000 output tokens per call. For larger generation tasks, the Message Batches API supports up to 300,000 output tokens per call via the output-300k-2026-03-24 beta header. The model's reliable knowledge cutoff is August 2025; its training data cutoff extends to January 2026. Anthropic has not published a needle-in-haystack recall figure for Sonnet 4.6, but the 1M context window is architecturally shared with Opus 4.6, which showed strong recall in internal evaluations. The model also features adaptive context compaction for extended agentic sessions.

Claude Sonnet 4.6 accepts text, image, and PDF inputs and produces text and tool-call outputs. Vision is live and supports up to 600 images or PDFs per request. Individual images can be up to 8,000x8,000 pixels. There is no native audio input or output. Function calling uses Anthropic's standard tool_use schema with support for parallel tool calls and structured JSON output. The model also supports computer use, including GUI navigation, web form completion, and spreadsheet manipulation. Adaptive thinking (the successor to extended thinking) lets the model decide how much internal reasoning to apply per query based on an effort parameter; the three standard levels are low, medium, and high, plus a max level for intensive tasks. Sonnet 4.6 also inherits Opus-level prompt injection resistance, which is a notable upgrade over Sonnet 4.5.

Pricing for Claude Sonnet 4.6 on the Anthropic API is $3.00 per million input tokens and $15.00 per million output tokens. Prompt caching reduces costs substantially: a 5-minute cache write costs $3.75 per million tokens and a cache read costs $0.30 per million tokens (10% of the base input rate). A 1-hour cache write costs $6.00 per million tokens. The Batch API offers a 50% discount, bringing prices to $1.50 input and $7.50 output per million tokens, with results returned asynchronously and a 300K output token ceiling via beta header. To put the pricing in context: summarizing a 100,000-token research paper costs roughly $0.30; running a daily coding agent generating 1 million input tokens and 200,000 output tokens costs approximately $6.00; processing 1,000 customer support turns at an average 2,000 input and 500 output tokens each costs approximately $9.50.

Claude Sonnet 4.6 is accessible via the Anthropic API (api.anthropic.com), AWS Bedrock (model ID: anthropic.claude-sonnet-4-6), Google Vertex AI (model ID: claude-sonnet-4-6), and Microsoft Foundry. On Bedrock, the model supports global cross-region, geo cross-region (US, EU, AU, JP), and select in-region endpoints. On Vertex, global, multi-region, and regional endpoints are available. Regional and multi-region endpoints on Bedrock and Vertex include a 10% premium over global routing. Authentication uses API keys for the Anthropic API and IAM credentials for Bedrock and Vertex. SDKs are available in Python, TypeScript, Java, Go, and Ruby; Go and Ruby do not support Microsoft Foundry. The model is closed-weights and cannot be self-hosted.

Anthropic deployed Claude Sonnet 4.6 under AI Safety Level 3 (ASL-3), the same standard as Opus 4.6. The system card notes that automated safety evaluations placed Sonnet 4.6 at or below the capability level of Claude Opus 4.6, meaning it does not push the capability frontier beyond what was already managed under ASL-3 safeguards. The model did not cross the threshold for ASL-4 classification on biological domain uplift tasks. On cyber capability evaluations across more than 1,500 CyberGym tasks, Sonnet 4.6 found security flaws 65% of the time (versus 67% for Opus 4.6 and 83% for the experimental Claude Mythos Preview). The alignment method combines Constitutional AI with RLHF. Safety evaluations included agentic and computer use scenarios, prompt injection resistance, and alignment under unusual and extreme conditions. On some alignment measures, Sonnet 4.6 showed the strongest results Anthropic has recorded for any Claude model.

Claude Sonnet 4.6 is the right choice for teams that need near-flagship coding and computer use performance at a lower cost per token. The 79.6% SWE-bench score and 72.5% OSWorld score place it well ahead of older Opus-class models. Teams running high-volume agentic coding loops that previously required Opus 4.5 can often downgrade to Sonnet 4.6 without a measurable quality drop. The model is also the right tool for long-document analysis tasks, given the 1M token context window at no premium. It is a poor fit for tasks requiring deep graduate-level scientific reasoning (GPQA 74.1% versus Opus 4.6's 91.3%), for real-time voice applications (no audio I/O), and for teams that need on-device deployment or air-gapped inference (closed weights, API-only). Gemini 3.1 Pro (80.6% SWE-bench) and GPT-5.2 (73.8% GPQA) are the primary alternatives to evaluate on those specific benchmarks.

Training data for Claude Sonnet 4.6 includes a curated mix of public web text, licensed datasets, and synthetic reasoning traces. Anthropic does not train on API inputs by default; inputs are retained for up to 30 days for safety and abuse monitoring and then deleted unless flagged. Enterprise customers can request a zero-retention arrangement. The model is deployed under Anthropic's Responsible Scaling Policy. SOC 2 Type II, HIPAA-eligible, and GDPR-compliant configurations are available through the Anthropic API and via AWS Bedrock's data governance controls. Anthropic's trust center URL is anthropic.com/transparency.

Claude Sonnet 4.6 replaced Claude Sonnet 4.5 as the default model on claude.ai for Free and Pro users on its launch date. The older Claude Sonnet 4 model (claude-sonnet-4-20250514) and Claude Opus 4 (claude-opus-4-20250514) are deprecated and will be retired on June 15, 2026; Anthropic's official migration target for Sonnet 4 users is Sonnet 4.6. Claude Sonnet 3.7 and Haiku 3.5 have already been retired. The next Sonnet-class model is expected to be Sonnet 4.8, with no announced release date. Sonnet 4.6 is the current production standard for mid-tier Claude deployments as of May 2026.

Pricing

$3.00 per 1M input tokens, $15.00 per 1M output tokens. Prompt caching: 5-minute cache write $3.75/MTok, 1-hour cache write $6.00/MTok. Batch API: $1.50 input / $7.50 output per 1M tokens for asynchronous jobs. Regional endpoints add 10% premium.

What a real job costs

Budgets: 20-page PDF = 30k in / 1k out · Support reply = 2k in / 300 out · Coding agent run = 200k in / 20k out. Computed from the vendor's per-token prices at render time; cached-input discounts are not applied.

Key Features

• Adaptive Thinking: Replaces fixed extended thinking with a dynamic system that allocates reasoning tokens based on task complexity and the configured effort level (low/medium/high/max). The model averages 246 thinking tokens per question on standard tasks, reducing overhead on simple queries while scaling up for hard ones.
• 1M Token Context Window: Generally available as of March 13, 2026, at standard per-token pricing. No beta header or long-context premium required. Accepts up to 600 images or PDFs per request alongside text.
• Computer Use: Supports GUI navigation, spreadsheet manipulation, and multi-step web form completion. Scores 72.5% on OSWorld-Verified, matching Opus 4.6 and more than doubling GPT-5.2's 38.2% score.
• Prompt Caching: Caches system prompts, documents, and conversation history. Cache reads cost $0.30 per million tokens, a 90% reduction from the base input rate. The 1-hour TTL supports long-running agent sessions without re-processing context on every turn.
• 300K Batch Output: Raises the synchronous 64K output ceiling to 300,000 tokens per call for asynchronous jobs, gated behind the output-300k-2026-03-24 beta header and paired with a 50% price cut. Suited for large code generation, documentation, and report synthesis tasks.

Pros

• Near-Opus coding and computer-use performance at Sonnet-tier pricing, making it the default choice for high-volume agentic workloads.
• Matches Opus 4.6 on computer-use automation while carrying Opus-level prompt injection resistance, the first time that safety bar reached a Sonnet-class model.
• 1M token context window at no surcharge, allowing full-codebase or multi-document analysis without chunking or vector retrieval.

Cons

• GPQA Diamond 74.1% trails Opus 4.6 by 17 points; scientific reasoning tasks in chemistry, biology, or advanced physics still require the flagship model.
• No native audio I/O; voice products must add a separate transcription and synthesis layer.
• Closed weights; teams requiring on-device, air-gapped, or fine-tuned deployment cannot use this model.

Benchmarks

• MATH: 89% vendor-reported · 17 Feb 2026 — Competition maths problems, % solved.
• ARC-AGI 2: 60.4% vendor-reported · 17 Feb 2026 — Abstract visual puzzles built to resist memorisation, % solved.
• GPQA Diamond: 74.1% vendor-reported · 17 Feb 2026 — PhD-level science questions that are hard to search for, % correct.
• SWE-bench Verified: 79.6% vendor-reported · 17 Feb 2026 — Real GitHub issues fixed end to end, % solved.

A benchmark is an exam, not the job. Scores transfer unevenly between tasks, so weigh the one closest to your workload and read every figure with its source.

Visit Claude Sonnet 4.6 Official Page