Claude Sonnet 4.6 — API Pricing & Benchmarks

Claude Sonnet 4.6 is Anthropic's previous-generation reasoning model, released in 2026-02. The API costs $3.00 per 1M input tokens and $15.0 per 1M output tokens ($6.00 blended at a 3:1 ratio), with cached input at $0.30. It supports a 1M-token context window. On the independently measured Artificial Analysis Intelligence Index it scores 48.4, ranking #30 of 68 models tracked on this site, with a median output speed of 49 tokens/sec.

Explore provider routes and additional benchmark sources for Claude Sonnet 4.6

ProviderAnthropic
Input price / 1M tokens$3.00
Output price / 1M tokens$15.0
Cached input / 1M tokens$0.30
Blended price (3:1)$6.00
Context window1M
Max output tokens128K
Knowledge cutoff2025-08
Licenseproprietary
Input modalitiesT+I
Output speed (tok/s)49
Released2026-02
Artificial Analysis Intelligence Index48.4
LMArena Elo1458
GPQA Diamond87.5%
SWE-bench Verified79.6%
MMLU-Pro79.2%
Humanity's Last Exam33.6%

Where Claude Sonnet 4.6 fits

On a blended 3:1 basis Claude Sonnet 4.6 costs $6.00 per 1M tokens, cheaper than 15% of the 80 priced models in this index. It is a previous-generation model: usually only worth choosing if you have an existing integration, negotiated pricing, or need its specific behavior, since newer siblings offer better price-performance. Measured output speed of 49 tokens/sec is below the index median of 104 — typical of reasoning-heavy models, and worth checking against interactive latency budgets. Cached input is priced at $0.30 (90% below standard input), which rewards prompt structures with long stable prefixes — see our caching guide.

Price history

PeriodInput $/1MOutput $/1MCached $/1M
2026-07-18 → current$3.00$15.0$0.30

No list-price changes recorded since tracking began. Full dataset: price-history.json (CC BY 4.0).

Alternatives at a similar price

About Anthropic

Anthropic builds the Claude family, organized into Opus (flagship), Sonnet (mid-tier) and Haiku (small) tiers alongside the Fable line. Claude models currently lead the Artificial Analysis Intelligence Index and are particularly strong on agentic-coding benchmarks such as SWE-bench Verified.

More Anthropic models

← Full comparison table (60+ models, live filters)