Claude Sonnet 4.6 review, pricing and limits

Claude Sonnet 4.6 is Anthropic's February 2026 mid-tier model, now one generation behind Sonnet 5.5 and Sonnet 5. It still handles coding agents and computer use well, and Anthropic lists it as Active until at least February 2027. New projects should start on the current Sonnet instead; keep 4.6 only where prompts and evaluations are already tuned to it.

Claude Sonnet 4.6, released in February 2026, is a mid-tier model from Anthropic's Claude 4 family, positioned below Claude Opus 4.6. It accepts text, images and PDFs, and holds a 1 million token context window billed at the standard rate. Anthropic still lists it as an active model, though two newer Sonnet generations now sit above it.

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 model, and made it the default on the Free and Pro plans of claude.ai in place of Sonnet 4.5. It is now a prior-generation model. Claude Sonnet 5.5 and Claude Sonnet 5 have since taken the current Sonnet slot at Anthropic. Anthropic's deprecations page still shows claude-sonnet-4-6 as Active, though its older sibling Sonnet 4.5 is already deprecated.

The launch figures put Sonnet 4.6 close to the flagship of its day. NxCode's write-up of Anthropic's launch table reports 79.6% on SWE-bench Verified against 80.8% for Claude Opus 4.6, and 72.5% on OSWorld-Verified (computer use) against 72.7%. Anthropic's own footnote says the SWE-bench score was averaged over ten trials and reached 80.2% with a prompt modification. The same table gives 74.1% on GPQA Diamond, a graduate-level science test where Opus 4.6 scores 91.3%, and Anthropic states that the 60.4% ARC-AGI-2 result was measured at high effort, not max effort. Among rivals at the time, Gemini 3.1 Pro posts 80.6% on SWE-bench Verified in Hokai's record.

Anthropic's launch post says early testers in Claude Code preferred Sonnet 4.6 to Sonnet 4.5 about 70% of the time, and to the November frontier model Opus 4.5 59% of the time, citing less overengineering and fewer false claims of success. It reports prompt injection resistance close to Opus 4.6, which matters for computer use agents that read untrusted web pages. On OfficeQA, which tests reading charts, PDFs and tables from enterprise documents, the launch post says it matches Opus 4.6.

The model accepts text, images and PDFs and returns text and tool calls, with no audio in or out. It supports adaptive thinking, extended thinking and context compaction (in beta at launch), which summarizes older context as a long conversation nears its limit. The context window is 1 million tokens, and Anthropic's pricing page now states that Claude 4.6 and later models bill the full window at the standard rate, so a 900,000-token request costs the same per token as a 9,000-token one. One detail changes cost comparisons: Sonnet 4.6 uses the earlier tokenizer, while Claude 4.7 and later models produce roughly 30% more tokens for the same text, so per-token prices across generations do not compare like for like.

Distribution spans Anthropic's own API and the three big cloud marketplaces, while the closed weights rule out self-hosting. Pricing detail sits in the notes below, and the headline is that Sonnet 5.5 now lists at a lower input and output rate than this model.

The case against starting new work on Sonnet 4.6 is simple. The current Sonnet is cheaper per token, Anthropic's migration guidance points developers forward, and retirement of this model, while not yet scheduled, will come with notice. It also trails Opus on deep science questions, lacks voice, and cannot run on-device or air-gapped. Teams should stay on it only where a prompt set and evaluation suite are already validated on this exact model, or where an enterprise cloud contract pins it. Everyone else should weigh Claude Opus 4.7, Claude Opus 4.8 and the newer Sonnet line first. Hokai's guide on which Claude model to use in 2026 walks through that choice, and a separate guide covers the 1M token context window in Sonnet 4.6.

Pricing

As of 23 October 2026, Claude Sonnet 4.6 costs $3 per million input tokens and $15 per million output tokens through the Claude API. Cache reads cost $0.30, 5-minute cache writes $3.75 and 1-hour cache writes $6. The Batch API halves token prices to $1.50 input and $7.50 output, and the full 1M context window bills at the standard rate.

US-only inference through the inference_geo setting carries a 1.1x multiplier on the Claude API, and Claude in Microsoft Foundry applies the same multiplier to US Data Zone Standard deployments, while Bedrock and Google-managed Vertex AI set their own regional prices. The newer Claude Sonnet 5.5 lists lower, at $2 input and $10 output per million tokens.

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 and Extended Thinking: Anthropic's launch post says Sonnet 4.6 supports both adaptive thinking, where the model decides how much to reason, and extended thinking, and that it performs well at any effort setting, even with extended thinking off.
• 1M Token Context Window: Billed at the standard per-token rate on the Claude API with no long-context surcharge, per Anthropic's pricing page. Context compaction, in beta at launch, summarizes older turns as a conversation nears the limit.
• Computer Use: Navigates GUIs, spreadsheets and multi-step web forms. Scores 72.5% on OSWorld-Verified in the launch table, and Anthropic reports prompt injection resistance similar to Opus 4.6.
• Prompt Caching: Cache reads cost a tenth of the base input rate, and a 1-hour cache write costs $6 per million tokens. Caching suits long-running agent sessions that resend the same context every turn.
• US-Only Inference Option: Setting inference_geo to us on the Claude API keeps processing in the United States at a 1.1x price multiplier, and Claude in Foundry charges the same 1.1x on its US Data Zone Standard deployment type.

Pros

• Close to Opus 4.6 on coding and computer-use results in the launch table, at a fifth of the Opus price per output token.
• Prompt injection resistance that Anthropic describes as similar to Opus 4.6, a real step up from the previous Sonnet for agents that browse.
• Full 1M token context at the standard rate, so long documents and repositories need no chunking or surcharge.
• Anthropic has set no retirement before February 2027, so pinned production apps have runway.

Cons

• It is a prior generation: Sonnet 5.5 lists at a lower rate and is the model Anthropic now points developers toward.
• GPQA Diamond of 74.1% trails Opus 4.6 by a wide margin, so chemistry, biology and physics reasoning still needs a flagship model.
• No native audio in or out, so voice products must add separate transcription and speech layers.
• Closed weights rule out self-hosting, fine-tuning on your own hardware, and air-gapped use.

Benchmarks

• MATH: 89% independent · 23 Oct 2026 — Competition maths problems, % solved.
• ARC-AGI 2: 60.4% vendor-reported · 23 Oct 2026 — Abstract visual puzzles built to resist memorisation, % solved.
• GPQA Diamond: 74.1% independent · 23 Oct 2026 — PhD-level science questions that are hard to search for, % correct.
• SWE-bench Verified: 79.6% independent · 23 Oct 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