Claude Sonnet 4.6

Claude Sonnet 4.6 is a mid-tier general-purpose model released by Anthropic on February 17, 2026, in the Claude 4.x family. Anthropic describes it as a hybrid reasoning model offering "fast, capable intelligence for real-time agents and high-volume work," with a 1M-token context window available in beta on the API (Source: anthropic.com). Its system card describes it as "substantially improved" over Claude Sonnet 4.5 and, on several evaluations, approaching or matching the higher-capability Opus 4.6 (Claude Sonnet 4.6 System Card). It is deployed under the ASL-3 Deployment and Security Standard and introduces thinking modes and an effort parameter.

Lineage and positioning

Sonnet 4.6 is the latest entry in Anthropic's Sonnet line, which Anthropic positions as the mid-tier between the smaller Haiku models and the higher-capability Opus models. The Sonnet line preceding it includes Claude Sonnet 3.7 (February 24, 2025), Claude Sonnet 4 (May 22, 2025), and Claude Sonnet 4.5 (September 29, 2025) (Source: anthropic.com). Anthropic describes Sonnet 4.6 as its "most capable Sonnet model yet" and a "full upgrade" across coding, computer use, long-context reasoning, agent planning, knowledge work, and design (Source: anthropic.com).

The system card describes Sonnet 4.6 as "substantially improved" over Sonnet 4.5 (documented in the Sonnet 4.5 system card) across coding, agentic tasks, reasoning, multimodal capabilities, computer use, and mathematics, and states that on several evaluations it approached or matched Claude Opus 4.6, Anthropic's frontier model at the time of release (Claude Sonnet 4.6 System Card). Anthropic reports that developers with early access preferred Sonnet 4.6 to its predecessor "by a wide margin" and "often even prefer it to" Claude Opus 4.5, its frontier model from November 2025 (Source: anthropic.com). Anthropic continues to position Opus 4.6 as the strongest option for tasks "that demand the deepest reasoning, such as codebase refactoring, coordinating multiple agents in a workflow, and problems where getting it just right is paramount" (Source: anthropic.com). In Mollick's guide to the agentic era, Sonnet 4.6 is described as "also powerful" alongside Opus 4.6 (Source: A Guide to Which AI to Use in the Agentic Era).

Architecture and training

Anthropic does not disclose Sonnet 4.6's parameter count, training-compute budget, or detailed architecture. It is a hybrid reasoning model: it can produce near-instant responses or extended, step-by-step reasoning, and Anthropic states it "is both a standard model and a hybrid reasoning model in one" (Source: anthropic.com). On the Claude Platform, Sonnet 4.6 supports both adaptive thinking and extended thinking, plus context compaction in beta, which automatically summarizes older context as conversations approach limits to increase effective context length (Source: anthropic.com).

Sonnet 4.6 introduces thinking modes and an effort parameter, documented in section 1.1.2 of the system card, which allow the model to vary its reasoning depth on demand (Claude Sonnet 4.6 System Card). API users have fine-grained control over the model's thinking effort, and Anthropic reports that the model "offers strong performance at any thinking effort, even with extended thinking off," recommending that developers migrating from Sonnet 4.5 explore the effort spectrum to balance speed and reliability (Source: anthropic.com). The model accepts text and image input; Anthropic frames its OfficeQA and document-comprehension results in terms of reading enterprise documents such as charts, PDFs, and tables (Source: anthropic.com).

The model supports a context window of up to 1M tokens, with the 1M window available in beta on the API only; Anthropic describes this as sufficient to hold "entire codebases, lengthy contracts, or dozens of research papers in a single request" and states that the model "reasons effectively across all that context" (Source: anthropic.com).

Capabilities and benchmarks

The system card reports evaluations across software engineering, agentic, reasoning, computer-use, domain, and long-context categories. Software engineering was measured on SWE-bench (Verified and Multilingual). Agentic terminal performance was measured on Terminal-Bench 2.0 and OpenRCA, with complex multi-step reasoning on τ2-bench and agentic business tasks on Vending-Bench 2 and MCP-Atlas. Computer-use and GUI automation were measured on OSWorld-Verified, novel reasoning on ARC-AGI, economic value added on GDPval-AA, and graduate-level science on GPQA Diamond. Domain evaluations covered the financial domain (Finance Agent and Real-World Finance) and cybersecurity (CyberGym). Long-context performance was measured on OpenAI MRCR v2 and GraphWalks (Claude Sonnet 4.6 System Card).

On BrowseComp, Sonnet 4.6 recorded a highest single-agent score of 74.01% and a multi-agent score of 82.07%; both figures reflect a March 6, 2026 correction described under safety and evaluations below (Claude Sonnet 4.6 System Card).

The following figures are reported by Anthropic in the announcement and its footnotes; benchmark scores are self-reported unless attributed to a third party.

For the announcement's headline comparison table and OSWorld chart, Anthropic states that for GPT-5.2 and Gemini 3 Pro it compared against "the best reported model version available via API" (Source: anthropic.com). On Terminal-Bench 2.0, Anthropic reports both scores reproduced on its own infrastructure and published scores from other labs, using the Terminus-2 harness (except OpenAI's Codex CLI), with the Sonnet 4.6 score reported with thinking turned off (Source: anthropic.com).

Anthropic's computer-use claims center on OSWorld-Verified, the standard benchmark for AI computer use, which presents hundreds of tasks across real software (Chrome, LibreOffice, VS Code, and others) on a simulated computer with no special APIs or connectors. Anthropic reports steady Sonnet gains on OSWorld across sixteen months and describes Sonnet 4.6 as showing "a major improvement in computer use skills compared to prior Sonnet models" (Source: anthropic.com). Scores prior to Sonnet 4.5 were measured on the original OSWorld; scores from Sonnet 4.5 onward use OSWorld-Verified, an in-place upgrade released in July 2025 (Source: anthropic.com). In the Vending-Bench Arena evaluation, which tests how well a model runs a simulated business over time in competition with other models, Anthropic reports that Sonnet 4.6 invested heavily in capacity for the first ten simulated months and then pivoted to profitability in the final stretch, finishing ahead of competitors (Source: anthropic.com).

In a third-party evaluation, Vals AI's SWE-bench Verified leaderboard (using the mini-SWE-agent bash-only harness, updated June 17, 2026) reports Sonnet 4.6 resolution rates by task difficulty of 87% on tasks under 15 minutes, 75% on 15-minute-to-1-hour tasks, 50% on 1-to-4-hour tasks, and 33% on tasks over 4 hours; it is placed below Anthropic's larger Opus models and competitors such as GPT 5.5 and Gemini 3.1 Pro on that leaderboard's overall ranking (Source: vals.ai).

Comparison with Opus 4.6 and competitors

In Claude Code, Anthropic's early testing found that users preferred Sonnet 4.6 over Sonnet 4.5 "roughly 70% of the time," reporting that it more effectively read context before modifying code and consolidated shared logic rather than duplicating it. Users preferred Sonnet 4.6 to Opus 4.5, Anthropic's frontier model from November 2025, 59% of the time, rating it less prone to overengineering and "laziness," better at instruction following, and producing fewer false claims of success and fewer hallucinations (Source: anthropic.com). Anthropic states that the model's prompt-injection resistance on its safety evaluations is "a major improvement compared to its predecessor, Sonnet 4.5," and "performs similarly to Opus 4.6" (Source: anthropic.com). Anthropic's announcement compares Sonnet 4.6 against GPT-5.2 and Gemini 3 Pro on its summary benchmark table (Source: anthropic.com).

Availability and pricing

Claude Sonnet 4.6 is available on all Claude plans, in Claude Cowork, in Claude Code, on Anthropic's API, and on major cloud platforms. It is the default model for Free and Pro plan users in claude.ai and in Claude Cowork, and the free tier was upgraded to Sonnet 4.6 by default, including file creation, connectors, skills, and compaction (Source: anthropic.com). For developers, it is available on the Claude Platform natively and on Amazon Bedrock, Google Cloud's Vertex AI, and Microsoft Foundry; the 1M-token context window is available in beta on the API only (Source: anthropic.com). The API model identifier is claude-sonnet-4-6.

Pricing starts at $3 per million input tokens and $15 per million output tokens, unchanged from Sonnet 4.5, with up to 90% cost savings via prompt caching and 50% cost savings via batch processing (Source: anthropic.com; Source: anthropic.com). For Claude in Excel users, the add-in supports MCP connectors to tools including S&P Global, LSEG, Daloopa, PitchBook, Moody's, and FactSet, available on Pro, Max, Team, and Enterprise plans (Source: anthropic.com). Alongside the release, Anthropic made several API tools generally available, including code execution, the memory tool, programmatic tool calling, tool search, and tool-use examples, and the web search and fetch tools now automatically write and execute code to filter and process search results (Source: anthropic.com).

Safety and evaluations

Sonnet 4.6 is deployed under the ASL-3 Deployment and Security Standard, the same standard applied to Sonnet 4.5. The ASL-3 Standard applied is defined in RSP v3.1 (Anthropic's Responsible Scaling Policy (Version 3.1)). Across the Responsible Scaling Policy evaluation domains, autonomy risks were evaluated and found appropriate for ASL-3 with no ASL-4 trigger reached, while CBRN risks and cyber risks were evaluated with no critical threshold crossed. A sabotage risk assessment was included in the release decision process (Claude Sonnet 4.6 System Card).

On alignment, the system card reports "low overall levels of misaligned behavior" and states that, on some alignment measures, Sonnet 4.6 showed "the best degree of alignment we have yet seen in any Claude model." The alignment assessment is described as covering a "very wide range of potentially misaligned behaviors" and testing model behavior "in unusual and extreme scenarios" (Claude Sonnet 4.6 System Card). Anthropic states that its safety evaluations of Sonnet 4.6 "overall showed it to be as safe as, or safer than, our other recent Claude models," and that its safety researchers concluded the model has "a broadly warm, honest, prosocial, and at times funny character, very strong safety behaviors, and no signs of major concerns around high-stakes forms of misalignment" (Source: anthropic.com). Anthropic notes that computer use poses prompt-injection risks, in which malicious actors attempt to hijack the model by hiding instructions on websites, and reports that Sonnet 4.6's resistance to such attacks improved over Sonnet 4.5 (Source: anthropic.com).

On March 6, 2026, Anthropic corrected the BrowseComp scores after an improved cheating-detection pipeline identified 9 additional instances of unintended solutions in the single-agent setting and 11 in the multi-agent setting. The single-agent score was revised from 74.72% to 74.01% and the multi-agent score from 82.62% to 82.07%. Anthropic published the revised methodology alongside the corrected scores (Claude Sonnet 4.6 System Card).

User wellbeing

Anthropic's user-wellbeing evaluations report a 98.7% appropriate response rate and a 0.075% benign refusal rate on single-turn suicide and self-harm prompts, and a 78% appropriate response rate on multi-turn suicide and self-harm prompts, up from Opus 4.1's 56%. On an automated behavioral audit, sycophancy was reduced 70–85% relative to Opus 4.1 (Source: anthropic.com).

Reception

Anthropic published statements from early enterprise customers alongside the release. Replit described the model's performance-to-cost ratio as "extraordinary"; Cursor called it "a notable improvement over Sonnet 4.5 across the board, including long horizon tasks and more difficult problems"; Windsurf said it brought "frontier-level reasoning in a smaller and more cost effective form factor"; and Cognition (Devin) said it "meaningfully closed the gap with Opus on bug detection." Databricks reported that Sonnet 4.6 matched Opus 4.6 on its OfficeQA enterprise-document benchmark, and Box reported that it outperformed Sonnet 4.5 on heavy-reasoning document Q&A "by 15 percentage points." Pace reported 94% on its insurance computer-use benchmark, "the highest of any Claude model" it tested, and Shortwave reported "zero hallucinated links" in its computer-use evaluations, down from roughly one in three previously. Letta reported the model was 70% more token-efficient than Sonnet 4.5 on its filesystem benchmark with a 38% accuracy improvement (Source: anthropic.com; Source: anthropic.com).

Related models

• Claude Opus 4.6 — higher-capability model in the same generation, positioned by Anthropic for the deepest-reasoning tasks.
• Claude Sonnet 4.5 — predecessor model (September 29, 2025).
• Claude Opus 4.7 and Claude Opus 4.8 — later Opus releases in the Claude 4.x line.

Relationships

• related: Claude Sonnet 4.6 System Card — primary source for capability and safety data
• related: Claude Opus 4.6 — higher-capability model in same generation
• related: Claude Sonnet 4.5 — predecessor model
• related: Anthropic's Responsible Scaling Policy (Version 3.1) — RSP v3.1 defines the ASL-3 Standard applied
• related: A Guide to Which AI to Use in the Agentic Era — describes Sonnet 4.6 alongside Opus 4.6