Anthropic logo

Claude Sonnet 4.6

Claudev4.6Current
byAnthropicAnthropic(ai lab)
Released February 17, 2026
Context200K
Price (In / Out)$3.00/M / $15.00/M
Training CutoffJanuary 2026
CategoryLarge Language Model
Max Output64K

Capabilities

What Claude Sonnet 4.6 can do and the modalities it supports.

Features

textvisioncodetool useextended thinkingcomputer use

Input Modalities

textimage

Output Modalities

text

Benchmarks

Performance scores for Claude Sonnet 4.6 across standard benchmarks.

BenchmarkScoreDate
SWE-bench Verified79.6February 17, 2026
OSWorld-Verified72.5February 17, 2026
Terminal-Bench 2.059.1February 17, 2026
ARC-AGI-258.3February 17, 2026
GDPval-AA Elo1633February 17, 2026
Finance Agent63.3February 17, 2026
Humanity's Last Exam19.1February 17, 2026

Pricing

Token pricing for Claude Sonnet 4.6 API usage.

Input Tokens

$3.00/M

per million tokens

Output Tokens

$15.00/M

per million tokens

Batch API: 50% discount. Prompt caching: up to 90% savings. 1M context window available in beta with premium long-context rates for requests exceeding 200K tokens.

Competing Models

Alternative models that compete with Claude Sonnet 4.6.

ModelFamilyContext WindowPrice (In / Out per M)Status
Gemini 2.5 ProGemini1.0M$1.25/M / $10.00/MCurrent
GPT-5.4GPT1.1M$2.50/M / $15.00/MCurrent

About Claude Sonnet 4.6

Anthropic's most capable Sonnet model, released February 2026. Claude Sonnet 4.6 delivers near-Opus-level performance at Sonnet pricing, featuring a hybrid reasoning architecture with adaptive extended thinking. It achieves 79.6% on SWE-bench Verified (approaching Opus 4.6's 80.8%), 72.5% on OSWorld-Verified for computer use, and a record 58.3% on ARC-AGI-2 for novel reasoning. With 30-50% lower latency and 25-45% reduced token burn on agentic tasks compared to Sonnet 4.5, it is the default model on claude.ai for Free and Pro users. Preferred over Sonnet 4.5 by 70% of developers and over Opus 4.5 by 59% in coding evaluations.

Technical Details

API Identifier
claude-sonnet-4-6
Category
Large Language Model
Context Window
200,000 tokens
Max Output Tokens
64,000 tokens

Tags

reasoningcodingmultimodaltool-usecomputer-usehybrid-reasoningagentic