Model Drop

Claude Sonnet 4.6

February 17, 2026 2 min read

Claude Sonnet 4.6, which Anthropic shipped on February 17, 2026, is an upgrade to Anthropic's mid-priced model that brings near-Opus coding, computer use and office work at the same price as Sonnet 4.5, plus a 1M-token context window in beta.

Model: Claude Sonnet 4.6 (claude-sonnet-4-6)
Maker: Anthropic
Ship date: February 17, 2026
Pricing: $3 / $15 per M tokens
Context: 200K tokens (1M in beta on the Claude API)
Weights: Closed
Where you can use it:
- Claude.ai (default for Free and Pro) and Cowork
- Claude Code and the Claude API
- All major cloud platforms
When to reach for it:
- Everyday coding and agent work
- Computer use and office documents
- Skip it for the deepest reasoning jobs
How it compares:
- Close to Claude Opus 4.6 on most evals
- Ahead of GPT-5.2 on office tasks
What makes it different:
- Coders often preferred it to Opus 4.5
- Opus-class work at Sonnet pricing
Early reactions: Early reviewers called it the best value model around and a better daily driver than Opus, while noting Opus 4.6 still led on deep reasoning.
Read more: Introducing Claude Sonnet 4.6, Claude Sonnet 4.6 system card

The benchmarks

Anthropic ran Sonnet 4.6 with adaptive thinking at max effort, averaged over 10 trials (Terminal-Bench 2.0 at max effort with thinking off, 5 runs per task), against Claude Sonnet 4.5, Claude Opus 4.5, Claude Opus 4.6, Gemini 3 Pro and the best reported GPT-5.2 variant (sometimes GPT-5.2 Pro); Terminal-Bench 2.0 rival scores are Anthropic's reruns (GPT-5.2 in OpenAI's Codex CLI), and GDPval-AA was run by Artificial Analysis.

Anthropic's numbers

SWE-bench Verified (real GitHub fixes): 79.6%. +2.4 over Claude Sonnet 4.5 (77.2%), −0.4 under GPT-5.2 (80.0%), −1.2 under Claude Opus 4.6 (80.8%).
Terminal-Bench 2.0 (command-line agent tasks): 59.1%. +8.1 over Claude Sonnet 4.5 (51.0%), +2.9 over Gemini 3 Pro (56.2%), −5.6 under GPT-5.2 (64.7%), −6.3 under Claude Opus 4.6 (65.4%).
OSWorld-Verified (operating a computer): 72.5%. +11.1 over Claude Sonnet 4.5 (61.4%), +6.2 over Claude Opus 4.5 (66.3%), −0.2 under Claude Opus 4.6 (72.7%).
GDPval-AA (office work, Elo): 1633. +27 over Claude Opus 4.6 (1606), +171 over GPT-5.2 (1462), +432 over Gemini 3 Pro (1201).
ARC-AGI-2 (novel visual puzzles): 58.3%. +44.7 over Claude Sonnet 4.5 (13.6%), +4.1 over GPT-5.2 Pro (54.2%), −10.5 under Claude Opus 4.6 (68.8%).
Humanity's Last Exam (no tools) (expert questions): 33.2%. +15.5 over Claude Sonnet 4.5 (17.7%), −3.4 under GPT-5.2 Pro (36.6%), −4.3 under Gemini 3 Pro (37.5%), −6.8 under Claude Opus 4.6 (40.0%).
GPQA Diamond (grad-level science): 89.9%. +6.5 over Claude Sonnet 4.5 (83.4%), −2.0 under Gemini 3 Pro (91.9%), −3.3 under GPT-5.2 Pro (93.2%).

Independent numbers

Artificial Analysis Intelligence Index (current v4.3.2): 30.1. +9.4 over Claude Sonnet 4.5 (20.7), +2.1 over Gemini 3 Pro (28.0), −0.3 under GPT-5.2 (30.4), −1.8 under Claude Opus 4.6 (31.9).
ARC-AGI-1 (ARC Prize, max effort): 86.00%. +11.00 over Gemini 3 Pro (75.00%), −0.17 under GPT-5.2 (xhigh) (86.17%), −7.00 under Claude Opus 4.6 (93.00%).