LLM Wiki

Claude Sonnet 4

Anthropic

Claude Sonnet 4 · Anthropic
Context200K
Max output8K
Licenseproprietary
Released2026-03-18
Modalitiestext, vision
Pricinginput $3.00/1M · output $15.00/1M · cache read $0.30/1M · effective 2026-03-18
Benchmarks
MMLU88.7%
GPQA Diamond72.5%
SWE-bench61.9%
LiveCodeBench56.4%
AIME 202545.1%
Changelog
Input pricing trend 2026-03-18: 4.00 → 2026-04-02: 3.00
  1. 2026-04-08
    SWE-bench score updated after v2 evaluation
    benchmarks.swe_bench 0.582→0.619
  2. 2026-04-05
    Context window expanded
    model.context_window 100.0K→200.0K
  3. 2026-04-02
    Input pricing cut
    model.pricing.input_per_1m 4→3
  4. 2026-03-18
    Launched — initial pricing
    model.pricing.input_per_1m null→4

Claude Sonnet 4

Anthropic's 2026 mid-tier frontier model. Successor to Claude Sonnet 3.5, sits between Haiku (fast/cheap) and Opus (max-capability). Targets the sweet spot of cost vs. reasoning quality for agentic workloads.

Notable features

  • 200K context window — matches the earlier 3.x generation; Opus 4.x pushes further for research workloads but Sonnet 4 is the default for most product use cases.
  • Vision — image input supported in the same endpoint as text.
  • Cache read at $0.30/1M — 10× cheaper than input, so long-running sessions with heavy context reuse can drop effective per-request cost by an order of magnitude.

Benchmarks

The numbers in the frontmatter come from Anthropic's launch announcement on 2026-03-18. Update this page (and append a time-series row to the changelog field from #56 once it lands) whenever Anthropic publishes a new set.

Connections

  • [[Anthropic]] — the provider
  • [[AgenticWorkloads]] — primary use case
  • [[CachePricing]] — the 10× cache-read discount changes cost models