← All models

Claude Sonnet 5.5

Language Available Full comparison ↗

anthropic/claude-sonnet-5-5 · by Anthropic · decoder-only-transformer

Pricing — 1 offering(s)

Anthropic

Input tokens

  • $2.00 / 1M tokens (input) Current 2026-09-28 → present
    Same as Sonnet 5's $2. The same figure appears in the models-overview table, the Sonnet 5.5 model page and…

    Same as Sonnet 5's $2. The same figure appears in the models-overview table, the Sonnet 5.5 model page and the launch post.

Output tokens

  • $10.00 / 1M tokens (output) Current 2026-09-28 → present
    Same as Sonnet 5's $10. The same figure appears in the models-overview table, the Sonnet 5.5 model page and…

    Same as Sonnet 5's $10. The same figure appears in the models-overview table, the Sonnet 5.5 model page and the launch post.

Cache hits and refreshes

  • $0.20 / 1M tokens (cached input) Historical 2026-09-28 → 2026-10-07
  • $0.10 / 1M tokens (cached input) Current 2026-10-08 → present

5-minute cache writes

  • $2.50 / 1M tokens (cache write) Current 2026-09-28 → present

Batch

Batch input tokens

  • $1.00 / 1M tokens (input) Current 2026-09-28 → present

Batch output tokens

  • $5.00 / 1M tokens (output) Current 2026-09-28 → present

Showing the active price and any recorded history. Full pricing history is available via the paid API — see API docs.

See how Claude Sonnet 5.5 fits into a cost-aware routing setup

See how →

About Claude Sonnet 5.5

Claude Sonnet 5.5 is the second model of Anthropic's Claude 5.5 family, released on 2026-09-28 with a system card the same day, six days after Claude Opus 5.5. It is a decoder-only transformer trained with large-scale pretraining, instruction tuning and preference optimisation. Adaptive thinking is on by default and steered by a five-level `effort` parameter that defaults to high.

On release Anthropic moved Claude Sonnet 5 to its "legacy models (still available)" list. Sonnet 5 isn't deprecated: the deprecations page lists it as Active with retirement not sooner than 2027-06-30. API pricing is unchanged from Sonnet 5 at $2 / $10 per million input / output tokens, with the same tokenizer.

Five API changes break Sonnet 5 code: thinking can't be set to `disabled` (`between_tools` replaces it); forced tool choice is rejected; thinking blocks are tied to the model and the conversation; the older `computer_20251124` computer-use tool is rejected on the Claude API and Google Cloud in favour of `computer_toolset_20260801`; and the advisor tool rejects Opus 4.8, Opus 4.7 and Sonnet 5 as advisors.

Capability profile

How Claude Sonnet 5.5 rates across core capability dimensions, with the task-level evidence behind each rating.

reasoning Strong
coding Strong

Terminal-Bench 2.1 (89 real terminal tasks; Terminus 2 harness): completes 83.1% — fails roughly 1 task in 6. Terminal-Bench 4.0 (66 long-horizon multi-domain tasks; mini-swe-agent): 64.1%, second on the board. Independent (Vals AI; Sonnet 5 served as a refusal fallback, and counting those tasks as failures gives 80.5% and 62.6%). Anthropic's own figures, vendor-reported: SWE-Bench Pro 81.3%, SWE-Bench Multilingual 90.3% and SWE-Bench Multimodal 54.3% (system card Table 8.1.A, max effort, no harness named), Terminal-Bench 4.0 70.6%, FrontierCode v1.1 (Main) 46.2% and CursorBench 4.0 55.5%.

tool use Strong

AutomationBench 44.7% (system card Table 8.1.A; ahead of Opus 5.5's 42.5% there) and OSWorld 2.1 80.1% partial score (vendor-reported). Forced tool choice (`any`/`tool`) is not supported: Anthropic recommends strict tool use or structured outputs instead.

instruction following Strong
context window Strong

1M-token context window and 128K max output (models-overview table and the Sonnet 5.5 model page); 300K output on the Batches API with the output-300k-2026-03-24 beta header.

multilingual Strong
speed Strong

Comparative latency is listed as 'Fast' in Anthropic's models-overview table (Opus 5.5 is 'Moderate', Haiku 4.5 'Fastest'). The launch post says it 'runs 30%+ faster' than Sonnet 5 (vendor claim). Fast mode is not offered on this model.

cost efficiency Strong

$2/$10 per MTok, unchanged from Sonnet 5, with standard 0.1x cache hits ($0.20/MTok) and a 50% Batch discount ($1/$5). Same tokenizer as Sonnet 5, so per-token prices compare directly. Anthropic says it 'costs up to 30% less for most work' than Sonnet 5 (vendor claim, through fewer tokens per task). On GDPval-AA v2.1 at max effort it scores 1,838 against Opus 5.5's 1,866 (Artificial Analysis, verified 2026-10-10), 28 Elo behind at half Opus 5.5's per-token price.

Coding benchmarks

View the full coding benchmark leaderboard →
Benchmark Score Harness Source
terminal-bench-2-1 83.2% 2026-09 terminus-2 📊 leaderboard ↗
terminal-bench-4-0 64.1% 2026-09 mini-swe-agent 📊 leaderboard ↗

Operator guidance

Route here in place of Sonnet 5 for general production, knowledge-work and agentic workloads: same per-token price, same tokenizer, and Anthropic reports it faster and cheaper per task. Set `effort` explicitly and re-tune it: levels are recalibrated from Sonnet 5 (Anthropic suggests `medium` for well-specified agentic coding, `medium` or `low` for chat). Its knowledge-work scores need max effort: on GDPval-AA v2.1 it drops from 1,838 at max to 1,550 at high (Artificial Analysis, verified 2026-10-10). Step up to Opus 5.5 for long-running agentic coding (SWE-Bench Pro 89.9% vs 81.3% in the same system-card table); drop to Haiku 4.5 for the lowest latency and cost. Don't route requests that need forced tool choice or `thinking: disabled`; both return an error.

Use cases

Limitations

Citations