gpt.college

Anthropic · Claude 5.5 · Reviewed 2026-10-02

Claude Sonnet 5.5

Anthropic's faster Sonnet release for scoped coding and everyday professional tasks.

All models
Overall · 0–10093.0
Status
active
Released
2026-09-28
Context
1,000,000 tokens
Parameters
Undisclosed
Inputs
Text · Image
Type
multimodal
Max output
128,000 tokens
API model ID
claude-sonnet-5-5
Reasoning
—
Knowledge cutoff
2026-06
Training cutoff
2026-06
Weights / license
Unavailable
Price / 1M tokens
$2 in · $10 out
Cached input
$0.2
Terminal codingComputer use1M-token context
CodingReasoningResearchMultimodalAgentic

Independent benchmark scores

Raw benchmark results

Source-backed rows with the original evaluation context preserved.

15 results
BenchmarkScoreSourceEvaluation notes
AA-Briefcase v1.1Checked 2026-10-02 1811 Anthropic AA v1.1 Elo reported by Anthropic; pre-release structured-output bug since fixed; may understate performance.
AutomationBench-AAChecked 2026-10-02 71% Artificial Analysis Artificial Analysis Intelligence Index v4.3.2; configuration=max · adaptive reasoning · default fallback. Rounded public score; checked date is not the evaluation execution date.
AA GDP.pdfChecked 2026-10-02 26% Artificial Analysis Artificial Analysis Intelligence Index v4.3.2; configuration=max · adaptive reasoning · default fallback. Rounded public score; checked date is not the evaluation execution date.
GDPval-AA v2.1Checked 2026-10-02 1844 Artificial Analysis Artificial Analysis Intelligence Index v4.3.2; configuration=max · adaptive reasoning · default fallback. Rounded public score; checked date is not the evaluation execution date.
AA Humanity's Last ExamChecked 2026-10-02 55% Artificial Analysis Artificial Analysis Intelligence Index v4.3.2; configuration=max · adaptive reasoning · default fallback. Rounded public score; checked date is not the evaluation execution date.
Artificial Analysis Intelligence Index v4.3.2Checked 2026-10-02 56 Artificial Analysis Artificial Analysis Intelligence Index v4.3.2; configuration=max · adaptive reasoning · default fallback. Rounded public score; checked date is not the evaluation execution date.
AA Terminal-Bench 4.0Checked 2026-10-02 64% Artificial Analysis Artificial Analysis Intelligence Index v4.3.2; configuration=max · adaptive reasoning · default fallback. Rounded public score; checked date is not the evaluation execution date.
Chartography without ToolsChecked 2026-10-02 61.6% Anthropic Anthropic table; no tools, separate from tool-enabled results.
CursorBench 4.0Checked 2026-10-02 55.5% Anthropic Anthropic-reported CursorBench 4.0.
FrontierCode 1.1 MainChecked 2026-10-02 46.2% Anthropic Anthropic table; effort=max, Main set; out-of-scope changes can reduce score.
FrontierCode 1.1 MainChecked 2026-10-02 52.1% Anthropic Anthropic table; effort=xhigh, Main set.
GDPval-AA v2.1Checked 2026-10-02 1844 Anthropic AA v2.1 Elo reported by Anthropic; pre-release structured-output bug since fixed; may understate performance.
Humanity's Last Exam with ToolsChecked 2026-10-02 64.5% Anthropic Anthropic table; tools enabled.
OSWorld 2.1 PartialChecked 2026-10-02 80.1% Anthropic Anthropic table; OSWorld 2.1 partial credit, separate from 2.0.
Terminal-Bench 4.0Checked 2026-10-02 70.6% Anthropic Anthropic-reported Terminal-Bench 4.0; preserve source-specific effort and harness settings.

Strengths

  • Terminal coding
  • Computer use
  • 1M-token context

Limitations

  • Maximum effort does not always give the best benchmark score
  • GDPval-AA and AA-Briefcase were evaluated on a pre-release deployment with a since-fixed structured-output bug
  • Haiku 5.5 is announced for the coming weeks, not released in this review

Model evidence · official · Checked 2026-10-02

Introducing Claude Sonnet 5.5

Vendor table. FrontierCode distinguishes max and xhigh effort. GDPval-AA/AA-Briefcase used a pre-release structured-output deployment with a since-fixed bug.

Model evidence · official · Checked 2026-10-02

Claude Sonnet 5.5 specifications

Official standard API specifications; default effort high on the platform differs from medium in Claude apps.

Task lenses preserve raw scores. Radar axes use score / current leader × 100; overall is the equal-weight average of all five indices. Missing evidence is not a zero or a low rank.