Claude Sonnet 4.6: Benchmarks, Pricing & Complete Guide
Claude Sonnet 4.6 scores 72.5% on OSWorld-Verified and 79.6% on SWE-bench Verified at $3/$15 per million tokens. Benchmarks, features, and pricing guide.
OSWorld-Verified Score
SWE-bench Verified
Per Million Tokens
Context Window (Tokens)
Key Takeaways
Corrected October 21, 2026: we re-checked every figure against Anthropic’s launch table, system card and API docs, and the GPT-5.2 figures against OpenAI. Many Opus 4.6, Sonnet 4.5 and GPT-5.2 scores in the original tables were wrong (GPQA Diamond, Humanity’s Last Exam, GDPval-AA, Finance Agent, Terminal-Bench 2.0, ARC-AGI-2, τ²-bench) and now match the published figures; we also removed MMLU-Pro and MATH-500 rows Anthropic never published, credited the 94% insurance result to Pace, and corrected the price gap to Opus 4.6 (40% cheaper per token, not 5x).
Update, October 2026: Anthropic now lists Sonnet 4.6 as a legacy model, still available and not retiring before February 17, 2027. Its successor, Claude Sonnet 5, arrived on June 30, 2026, and Anthropic’s current Sonnet model is Claude Sonnet 5.5, released September 28, 2026. Since March 13, 2026 the full 1M-token window has been generally available on Sonnet 4.6 at the standard $3/$15, with no beta header, on Anthropic’s own platform, Microsoft Foundry and Google Cloud’s Vertex AI. Anthropic’s docs now list manual extended thinking on Sonnet 4.6 as deprecated in favor of adaptive thinking.
Anthropic released Claude Sonnet 4.6 on February 17, 2026 — its most capable Sonnet model at the time. It scores 72.5% on OSWorld-Verified for computer use (0.2 points behind Opus 4.6), 79.6% on SWE-bench Verified for coding, and posts the top office-productivity score in Anthropic’s launch table at 1633 Elo on GDPval-AA. The price stays at $3/$15 per million tokens, the same as Sonnet 4.5.
In Anthropic’s early testing in Claude Code, users preferred Sonnet 4.6 over Sonnet 4.5 roughly 70% of the time and over Opus 4.5 59% of the time. Combined with a 4.3x jump on ARC-AGI-2 (13.6% to 58.3%), 63.3% on the Finance Agent benchmark, and a 1M token context window in beta, Sonnet 4.6 brings work that Anthropic says previously needed an Opus-class model to the Sonnet tier. Anthropic still recommends Opus 4.6 for tasks that demand the deepest reasoning.
Sonnet 4.6 launched on all Claude plans, in Claude Cowork (Anthropic’s agent for handing off multi-step tasks across your files and tools), in Claude Code, and on the Claude API. It is also available on Amazon Bedrock, Google Cloud Vertex AI and Microsoft Foundry for enterprise workloads.
What Is Claude Sonnet 4.6?
Claude Sonnet 4.6 is the February 2026 model in Anthropic's Sonnet tier — positioned between the lightweight Haiku and the premium Opus. It ships as claude-sonnet-4-6 on the API and replaced Sonnet 4.5 as the default model on claude.ai for both Free and Pro plan users. In Anthropic’s launch table it comes close to Opus 4.6 on two capabilities central to production AI agents: computer use and coding.
The headline numbers: 72.5% on OSWorld-Verified (Opus 4.6 scores 72.7%), 79.6% on SWE-bench Verified (Opus 4.6 scores 80.8%), and the top office-task score in Anthropic’s table at 1633 Elo on GDPval-AA. In Anthropic’s early testing in Claude Code, users preferred Sonnet 4.6 over Sonnet 4.5 roughly 70% of the time, and over Opus 4.5, Anthropic’s frontier model from November 2025, 59% of the time.
Building with AI? Our AI & Digital Transformation team can help you evaluate and deploy Claude models for production agent architectures.
- 72.5% OSWorld-Verified computer use (+11.1pp)
- 79.6% SWE-bench coding (+2.4pp)
- 58.3% ARC-AGI-2 (+44.7pp)
- 1M token context window (beta)
- $3/$15 per million tokens
- Extended thinking mode
- Tool use and function calling
- Vision and image analysis
Full Benchmark Comparison
Anthropic’s launch table compares Sonnet 4.6 with Sonnet 4.5, Opus 4.6, Opus 4.5, Gemini 3 Pro and GPT-5.2 on 13 benchmarks covering coding, computer use, agentic tasks, reasoning, and domain-specific evaluations. Sonnet 4.6 posts the top score in that table on two of them — office productivity and financial analysis — and comes close to Opus 4.6 on most of the rest.
| Benchmark | Sonnet 4.6 | Sonnet 4.5 | Opus 4.6 | GPT-5.2 |
|---|---|---|---|---|
| SWE-bench Verified | 79.6% | 77.2% | 80.8% | 80.0% |
| Terminal-Bench 2.0 | 59.1% | 51.0% | 65.4% | 64.7% (Codex CLI) |
| OSWorld-Verified | 72.5% | 61.4% | 72.7% | 38.2% |
| Vending-Bench Arena (final balance) | $5,639 | $2,125 | $4,053 | — |
| ARC-AGI-2 | 58.3% | 13.6% | 68.8% | 54.2% (Pro) |
| GPQA Diamond | 89.9% | 83.4% | 91.3% | 93.2% (Pro) |
| MMMLU (multilingual Q&A) | 89.3% | 89.5% | 91.1% | 89.6% |
| GDPval-AA (Office, Elo) | 1633 | 1276 | 1606 | 1462 |
| Finance Agent v1.1 | 63.3% | 54.5% | 60.1% | 59.0% |
| MCP-Atlas Scaled Tool Use | 61.3% | 43.8% | 59.5% | 60.6% |
| τ²-bench Retail | 91.7% | 86.2% | 91.9% | 82.0% |
| τ²-bench Telecom | 97.9% | 98.0% | 99.3% | 98.7% |
| BrowseComp | 74.0% | 43.9% | 83.7% | 77.9% (Pro) |
| MMMU-Pro (no tools / with tools) | 74.5% / 75.6% | 63.4% / 68.9% | 73.9% / 77.3% | 79.5% / 80.4% |
| Humanity's Last Exam (no tools) | 33.2% | 17.7% | 40.0% | 36.6% (Pro) |
| Humanity's Last Exam (with tools) | 49.0% | 33.6% | 53.0% | 50.0% (Pro) |
Source: the benchmark table in Anthropic’s Claude Sonnet 4.6 announcement (February 17, 2026) and the Sonnet 4.6 system card. Vending-Bench Arena balances are Andon Labs’ published final balances for its February 17, 2026 round, in which Sonnet 4.6, Opus 4.6 and Sonnet 4.5 competed (Anthropic’s chart shows only the two Sonnet models; GPT-5.2 was not in that round). For GPT-5.2, Anthropic used the best reported version available via API: “Pro” marks GPT-5.2 Pro, and the unmarked SWE-bench, τ²-bench, MCP-Atlas, MMMLU and MMMU-Pro cells match OpenAI’s own GPT-5.2 Thinking results in its GPT-5.2 announcement; Anthropic does not name the variant behind the GDPval-AA, Finance Agent and OSWorld-Verified cells. Terminal-Bench 2.0 for GPT-5.2 is Anthropic’s run in Codex CLI (Anthropic lists 64.0% as the self-reported figure), and the 38.2% OSWorld-Verified figure appears only in Anthropic’s announcement: OpenAI did not publish one for GPT-5.2, and the system card leaves that cell blank. Sonnet 4.6’s 59.1% on Terminal-Bench 2.0 was run with thinking off according to the announcement and the system card’s Terminal-Bench section (at max effort, the system card adds), although the system card’s summary table labels it “default thinking.” GDPval-AA was run by Artificial Analysis; its Elo ratings are relative to the models rated at the time, so they do not match Artificial Analysis’s current leaderboard. Finance Agent v1.1 was run by Vals AI (the system card lists GPT-5.2 at 58.5% there). BrowseComp reflects Anthropic’s March 2026 system-card revisions (first published as 74.7% for Sonnet 4.6 and 84.0% for Opus 4.6). Sonnet 4.6’s ARC-AGI-2 score is at max effort; at high effort the ARC Prize Foundation recorded 60.4%.
The standout categories for Sonnet 4.6 are office productivity (1633 Elo on GDPval-AA, the top score in Anthropic’s table and ahead of Opus 4.6’s 1606), financial analysis (63.3% on Finance Agent v1.1, also the top score), and scaled tool use (61.3% on MCP-Atlas, ahead of Opus 4.6’s 59.5% and second only to Opus 4.5’s 62.3%). On reasoning benchmarks such as GPQA Diamond and Humanity’s Last Exam, it trails both Opus 4.6 and GPT-5.2 Pro.
These scores are Anthropic’s launch figures, from its own runs and third-party evaluators (Artificial Analysis for GDPval-AA, Vals AI for Finance Agent, Andon Labs for Vending-Bench Arena). Results may vary with different prompting strategies, effort settings and evaluation setups. For a broader comparison of recent AI models, see our ByteDance Seed 2.0 benchmarks guide.
Computer Use: 72.5% on OSWorld-Verified
Anthropic says it was the first to introduce a general-purpose computer-using model, in October 2024. Its OSWorld chart shows the Sonnet line’s progress: 14.9% (Sonnet 3.5, new version, October 2024) → 28.0% (Sonnet 3.7) → 42.2% (Sonnet 4) → 61.4% (Sonnet 4.5) → 72.5% (Sonnet 4.6) over 16 months. Scores before Sonnet 4.5 were measured on the original OSWorld; from Sonnet 4.5 onward they use OSWorld-Verified. Sonnet 4.6 now matches Opus 4.6 (72.7%) for practical purposes, while Anthropic’s table lists GPT-5.2 at 38.2%.
Anthropic says early Sonnet 4.6 users saw human-level capability on tasks like navigating a complex spreadsheet or filling out a multi-step web form, then pulling it all together across several browser tabs. Among early-access customers, Pace reported that Sonnet 4.6 hit 94% on Pace’s own insurance benchmark, making it the highest-performing model Pace had tested for computer use; Pace called that accuracy mission-critical to insurance workflows such as submission intake and first notice of loss. The appeal is software built before modern APIs existed, which Anthropic notes previously needed bespoke connectors before AI could use it.
Computer Use Applications
Multi-step web forms, insurance intake, and data entry workflows; Pace reported 94% on its insurance benchmark.
Interacts with desktop applications and legacy systems that lack APIs, without building a bespoke connector for each.
End-to-end UI testing, spreadsheet validation, and cross-app workflows without custom test scripts.
The near-parity between Sonnet 4.6 and Opus 4.6 on computer use is significant because Sonnet 4.6’s per-token price is 60% of Opus 4.6’s ($3/$15 against $5/$25). For teams building computer-use agents, Sonnet 4.6 delivers essentially the same OSWorld-Verified score at a lower cost, which makes it a strong default for production deployments.
Coding and Agentic Performance
Sonnet 4.6 scores 79.6% on SWE-bench Verified — a 2.4 percentage point improvement over Sonnet 4.5's 77.2% and approaching Opus 4.6's 80.8%. On Terminal-Bench 2.0, which evaluates agentic coding in a terminal, it reaches 59.1% with thinking turned off (up from 51.0%), behind Opus 4.6’s 65.4% and the 64.7% Anthropic measured for GPT-5.2 in Codex CLI. Early users told Anthropic the model read context more effectively before modifying code, consolidated shared logic instead of duplicating it, and made fewer false claims of success.
Coding Benchmark Comparison
| Benchmark | Sonnet 4.6 | Sonnet 4.5 | Opus 4.6 |
|---|---|---|---|
| SWE-bench Verified | 79.6% | 77.2% | 80.8% |
| Terminal-Bench 2.0 | 59.1% | 51.0% | 65.4% |
| Vending-Bench Arena (final balance) | $5,639 | $2,125 | $4,053 |
Source: Anthropic’s Sonnet 4.6 announcement and system card; Vending-Bench Arena from Andon Labs’ published results for its February 17, 2026 round. The announcement and the system card’s SWE-bench section say the SWE-bench Verified score is averaged over 10 trials (a footnote to the system card’s summary table says 25); Anthropic saw 80.2% with a prompt modification.
The Vending-Bench Arena result is particularly noteworthy. In this Andon Labs evaluation, models run a simulated vending business over a year and compete against each other for profit. In the round Andon Labs ran for the Sonnet 4.6 launch, Sonnet 4.6 won with a final balance of $5,639, ahead of Opus 4.6 ($4,053) and Sonnet 4.5 ($2,125), about 2.7x Sonnet 4.5’s result. Anthropic says Sonnet 4.6 invested heavily in capacity for the first ten simulated months and then pivoted sharply to profitability, and that the timing of the pivot helped it finish well ahead. On the non-competitive Vending-Bench 2, the system card reports a final balance of $7,204.14 for Sonnet 4.6 at max effort against Opus 4.6’s $8,017.59, with a mean run cost of $265.03 versus $682.37.
Agentic Tool Use
Beyond pure coding, Sonnet 4.6 performs strongly on tool-use benchmarks that test a model's ability to interact with external systems. On τ²-bench, it scores 91.7% on retail scenarios and 97.9% on telecom scenarios, a little behind Opus 4.6 (91.9% and 99.3%). The MCP-Atlas scaled tool use benchmark, which tests multi-step workflows across many MCP tools, shows Sonnet 4.6 at 61.3% — ahead of Opus 4.6’s 59.5% at max effort (Opus 4.6 scores 62.7% at high effort) and just behind Opus 4.5’s 62.3%.
For developers using coding-focused AI models, Sonnet 4.6 represents a strong option that balances coding performance with broader agentic capabilities and competitive pricing.
New Features and Capabilities
Beyond raw benchmark improvements, the Sonnet 4.6 launch brought several features that expand what Claude can do in practice. These range from infrastructure changes (1M token context) to product features (MCP connectors in Claude in Excel) that make the model more useful for enterprise workflows.
1M Token Context Window (Beta)
Sonnet 4.6 launched with a 1M token context window in beta — about 750,000 words on its tokenizer, which Anthropic says is enough to hold entire codebases, lengthy contracts, or dozens of research papers in a single request. It is paired with context compaction (beta), which automatically summarizes older context as conversations approach limits, increasing effective context length.
Adaptive and Extended Thinking
Sonnet 4.6 retains extended thinking from Sonnet 4.5 but adds an adaptive mode that automatically adjusts thinking depth based on task complexity. Simple questions get fast responses, while complex reasoning problems trigger deeper thinking chains — optimizing both latency and cost.
Improved Prompt Injection Resistance
Anthropic’s safety evaluations show Sonnet 4.6 is a major improvement over Sonnet 4.5 in resisting prompt injection attacks and performs similarly to Opus 4.6. This is critical for production deployments where models process untrusted input, and particularly relevant for computer-use agents that interact with arbitrary web pages and desktop applications.
Web Search, Code Execution & Memory
Alongside the launch, Claude’s API web search and web fetch tools gained dynamic filtering: they write and run code to filter results before they reach the context window. Code execution, web fetch, memory, programmatic tool calling, tool search and tool use examples became generally available on the API with no beta header, and code execution became free when used with web search or web fetch. The memory tool lets Claude store and retrieve information across conversations in a directory of memory files.
Claude in Excel & MCP Connectors
Anthropic’s Claude in Excel add-in now supports MCP (Model Context Protocol) connectors, so Claude can pull context from tools such as S&P Global, LSEG, Daloopa, PitchBook, Moody’s and FactSet without leaving the spreadsheet. Connectors already set up in Claude.ai work in Excel automatically, and the feature is available on Pro, Max, Team, and Enterprise plans. It gives enterprise users a way to use Claude for financial modeling, data transformation, and report generation inside their existing tools.
Pricing and Availability
Sonnet 4.6 keeps Sonnet 4.5’s pricing — $3 per million input tokens and $15 per million output tokens — which Anthropic described as “starting at” $3/$15. At launch the 1M window was a beta for organizations in usage tier 4 or with custom rate limits, and requests over 200K input tokens were billed at long-context rates of $6 input and $22.50 output per million tokens, until the full window moved to standard pricing on March 13, 2026. Batch API requests are 50% off ($1.50/$7.50), prompt-cache writes cost $3.75 (5-minute) or $6 (1-hour) and cache reads $0.30 per million tokens. The API model ID is claude-sonnet-4-6.
$3 / $15
Input / Output per 1M tokens
claude-sonnet-4-6
API Model ID
4 API Platforms
Claude API, Bedrock, Vertex AI, Foundry
Where to Access Sonnet 4.6
- claude.ai — Default model for Free and Pro plans at launch, and available on all Claude plans
- Claude Cowork — Anthropic’s agent for handing off multi-step tasks; Sonnet 4.6 became its default model for Free and Pro users
- Claude Code — Anthropic’s coding agent for developers, with Sonnet 4.6 available at launch
- Anthropic API — Direct access via
claude-sonnet-4-6 - Amazon Bedrock — Enterprise deployment with AWS infrastructure
- Google Cloud Vertex AI — Enterprise deployment with GCP infrastructure
- Microsoft Foundry — Enterprise deployment on Azure, available from launch day
Sources: Anthropic’s Sonnet 4.6 model page and pricing page (the launch-era long-context rates and beta access terms are from the pricing page as archived on February 18, 2026), Anthropic’s API release notes and its March 13, 2026 announcement of 1M context general availability, and Microsoft’s Foundry announcement of February 17, 2026.
When to Choose Sonnet 4.6 vs Opus 4.6
- Computer use agents (72.5% vs 72.7%)
- Office productivity and data tasks
- Standard coding tasks (79.6% SWE-bench)
- Cost-sensitive production deployments
- Novel reasoning (68.8% vs 58.3% ARC-AGI-2)
- Codebase refactoring and coordinating multiple agents
- Terminal-based agentic coding (65.4% vs 59.1%)
- Hardest reasoning problems (HLE: 40.0% vs 33.2% without tools)
How Sonnet 4.6 Compares
vs Opus 4.6
Sonnet 4.6 achieves near-parity with Opus 4.6 on computer use (72.5% vs 72.7%) and trails slightly on coding (79.6% vs 80.8% SWE-bench). On reasoning, Opus maintains a clear lead — 68.8% vs 58.3% on ARC-AGI-2 and 40.0% vs 33.2% on Humanity’s Last Exam without tools — and it leads on Terminal-Bench 2.0 (65.4% vs 59.1%) and τ²-bench Telecom (99.3% vs 97.9%). Sonnet 4.6 leads on office tasks (1633 vs 1606 Elo) and financial analysis (63.3% vs 60.1%). At 40% lower per-token cost, Sonnet 4.6 is a sensible default for most production workloads; Anthropic itself recommends Opus 4.6 for codebase refactoring, coordinating multiple agents, and problems where getting it just right is paramount.
vs GPT-5.2
In Anthropic’s launch table, the gap with GPT-5.2 is widest on computer use: Sonnet 4.6’s 72.5% on OSWorld-Verified against 38.2%, a figure OpenAI did not publish itself. On coding, GPT-5.2 is ahead: slightly on SWE-bench Verified (80.0% vs 79.6%) and by 5.6 points on Terminal-Bench 2.0 (64.7% vs 59.1% in Anthropic’s Codex CLI run). Sonnet 4.6 leads on office tasks (1633 vs 1462 Elo), financial analysis (63.3% vs 59.0%), and ARC-AGI-2 (58.3% vs GPT-5.2 Pro’s 54.2%), while GPT-5.2 Pro leads on GPQA Diamond (93.2% vs 89.9%) and BrowseComp (77.9% vs 74.0%).
vs Gemini 3 Pro
Against Gemini 3 Pro, Anthropic’s figures favor Sonnet 4.6 on most agentic work: Terminal-Bench 2.0 (59.1% vs 56.2% in Anthropic’s run, or Google’s self-reported 54.2%), SWE-bench Verified (79.6% vs Gemini 3 Pro’s 76.2% in the system card; the announcement’s table shows Gemini 3 Flash’s 78.0% instead), MCP-Atlas (61.3% vs 54.1%), Finance Agent (63.3% vs 55.2%) and GDPval-AA (1633 vs 1201 Elo). Gemini 3 Pro leads on τ²-bench Telecom (98.0% vs 97.9%), GPQA Diamond (91.9% vs 89.9%), MMMU-Pro without tools (81.0% vs 74.5%), multilingual MMMLU (91.8% vs 89.3%) and Humanity’s Last Exam without tools (37.5% vs 33.2%), and Anthropic lists no Gemini OSWorld-Verified score. The Qwen 3.5 guide covers additional frontier model comparisons.
Sonnet 4.6 lands within 0.2 points of Opus 4.6 on OSWorld-Verified and 1.2 points on SWE-bench Verified at 60% of Opus 4.6’s per-token price. For teams that previously needed Opus for computer-use and standard coding agents, Sonnet 4.6 removes most of the cost premium. Opus 4.6 remains clearly ahead on novel reasoning (ARC-AGI-2, Humanity’s Last Exam), terminal-based coding and agentic search (BrowseComp 83.7% vs 74.0%).
Conclusion
Claude Sonnet 4.6 represents a significant milestone in Anthropic's model lineup. By achieving near-Opus performance on computer use (72.5% vs 72.7%) and coding (79.6% vs 80.8%) at Sonnet-tier pricing ($3/$15 per million tokens, 40% below Opus 4.6), it makes frontier-level AI agents cheaper to run. Its top scores in Anthropic’s table for office productivity (1633 Elo) and financial analysis (63.3%) further distinguish it from both Opus 4.6 and GPT-5.2.
For most production workloads — computer use agents, coding assistants, tool-use pipelines, and office automation — Sonnet 4.6 was the practical default at launch. Opus 4.6 retains its edge for the hardest reasoning problems and complex multi-step analysis, but Sonnet 4.6's combination of performance, cost, and new features (1M context, compaction, MCP connectors in Claude in Excel) made it a practical choice for most of those workloads in early 2026.
Ready to Build with Claude Sonnet 4.6?
Whether you're deploying computer-use agents, building coding assistants, or automating enterprise workflows, our team can help you leverage Claude's capabilities for measurable business results.
Frequently Asked Questions
Deep dives on AI, marketing and development.
Practical guides and fresh insights by email. No recycled takes.
Related Guides
Continue exploring AI model developments and frontier benchmarks