Claude Sonnet 5.5

Name: Claude Sonnet 5.5
Author: Anthropic

Multimodal

Anthropic

Claude Sonnet 5.5 is the second model in Anthropic's Claude 5.5 family and a faster, lower-cost complement to Claude Opus 5.5. It targets well-scoped everyday tasks, coding, and polished documents, slides, and spreadsheets, and runs 30%+ faster than Claude Sonnet 5 while typically using far fewer tokens at the same $2/$10 per million input/output prices. It accepts text and image input with text output, supports multilingual and vision workloads, tool use, and adaptive thinking (Claude Platform default effort high; Claude Code and apps default medium). Context window is 1M tokens with 128K max output on the sync Messages API; reliable knowledge and training-data cutoffs are June 2026. Cache reads are $0.20 and 5-minute cache writes are $2.50 per million tokens. Capability scores are self-reported from the Claude Sonnet 5.5 System Card (Table 8.1.A and §8) and the launch announcement. Available on the Claude API as `claude-sonnet-5-5`.

Key Specifications

Parameters

Context

1.0M

Release Date

September 28, 2026

Average Score

67.4%

API Documentation Research Paper Results Blog

Timeline

Key dates in the model's history

Announcement

September 28, 2026

Last Update

September 29, 2026

Today

October 5, 2026

Technical Specifications

Parameters

Training Tokens

Knowledge Cutoff

June 1, 2026

Family

Capabilities

MultimodalZeroEval

Pricing & Availability

Input (per 1M tokens)

$2.00

Output (per 1M tokens)

$10.00

Max Input Tokens

1.0M

Max Output Tokens

128.0K

Supported Features

Function CallingStructured OutputCode ExecutionWeb SearchBatch InferenceFine-tuning

Benchmark Results

Model performance metrics across various tests and benchmarks

Other Tests

Specialized benchmarks

AA-Briefcase v1.1

Announcement + System Card Table 8.1.A. AA-Briefcase v1.1 Elo (max_score 3000). Same Artificial Analysis pre-release structured-output bug caveat as GDPval-AA. • Self-reported

60.4%

ArXivMath

System Card §8.9. ArXivMath August 2026 (57 problems); without tools; max effort; averaged over four attempts per problem. Evaluated internally without safeguards classifiers. • Self-reported

86.8%

ArXivMath (with tools)

System Card §8.9. ArXivMath August 2026 (57 problems); with tools (code-execution sandbox, no internet); max effort; averaged over four attempts per problem. • Self-reported

95.2%

AutomationBench

System Card Table 8.1.A / §8.14.6. Zapier AutomationBench 44.7% at max effort with API default fallbacks enabled. Adaptive thinking at max effort. • Self-reported

44.7%

BenchCAD

System Card §8.13.2. BenchCAD Vision2Code 1k subset; voxel IoU 0.747 without tools; adaptive thinking at max effort; averaged over five runs. • Self-reported

74.7%

BenchCAD (with Python tool)

System Card §8.13.2. BenchCAD Vision2Code 1k subset; voxel IoU 0.963 with tools; adaptive thinking at max effort; averaged over five runs. • Self-reported

96.3%

Biomedical image analysis

System Card §8.17.7. Biomedical image analysis 72.2%; 1 attempt per problem; adaptive thinking at max effort. • Self-reported

72.2%

BioMysteryBench

System Card §8.17.1. BioMysteryBench Human Solvable subset 89.2%; adaptive thinking at max effort; bash/file tools with package allow-list. (RSP summary rounded to 0.89; §8.17.1 is authoritative.) • Self-reported

89.2%

BioMysteryBench (Human Difficult)

System Card §8.17.1. BioMysteryBench Human Difficult subset 44.7%; adaptive thinking at max effort; bash/file tools with package allow-list. (RSP summary rounded to 0.45; §8.17.1 is authoritative.) • Self-reported

44.7%

Chartography

System Card §8.13.1. Chartography with tools 90.2%; adaptive thinking at max effort; container with image file, standard libraries, and image cropping tool. • Self-reported

90.2%

Chartography (no tools)

Announcement + System Card §8.13.1. Chartography without tools 61.6%; adaptive thinking at max effort. • Self-reported

61.6%

CursorBench 4.0

Announcement table + System Card §8.8. CursorBench 4.0 at max effort 55.5% (independently measured by Cursor in its production harness). Also reported: xhigh 53.1%, high 47.8%, medium 39.2%. • Self-reported

55.5%

DeepSWE 1.1

System Card §8.3. DeepSWE v1.1; 5-trial average 71.0%. Adaptive thinking at max effort. • Self-reported

71.0%

De novo protein binder design

System Card §8.17.6. De novo protein binder design 82.3%; 15 targets, 30 miniprotein binders each, 24h wall-clock budget; adaptive thinking at max effort. • Self-reported

82.3%

FrontierCode 1.1

Announcement + System Card §8.4 / Table 8.1.A. FrontierCode v1.1 Main. Catalog stores Xhigh 52.1% as the fair headline (Table 8.1.A max-effort Main is 46.2%). Anthropic notes Max is lower due to Claude Code code-review subagent timeouts / out-of-scope edits penalized by mergeability grading. Extended Xhigh 64.4% stored separately. • Self-reported

52.1%

FrontierCode 1.1 (Extended)

System Card §8.4. FrontierCode v1.1 Extended at Xhigh effort 64.4% (max effort Extended 59.1%). Same mergeability grading caveats as Main. • Self-reported

64.4%

FrontierSWE V2

System Card §8.7. FrontierSWE v2; Proximal harness; max effort; 5 trials per task; mean across trials. • Self-reported

61.9%

GDPval-AA 2.1

Announcement + System Card Table 8.1.A. GDPval-AA v2.1 Elo (max_score 3000). Artificial Analysis evaluated a pre-release Claude Platform deployment with a structured-output bug later fixed; Anthropic expects any impact to be small and to understate performance. • Self-reported

61.5%

Global-MMLU

System Card §8.16.1. GMMLU average accuracy 92.1% across 42 languages; adaptive thinking at max effort; single trial; no tools. • Self-reported

92.1%

HealthBench

System Card §8.15.4. HealthBench length-adjusted scores across all five effort levels sit in 64.7–65.4%; store the published band max 65.4% to match the catalog convention (length-adjusted on healthbench). Raw 69.4% at max effort (§8.15.1) is stored separately as healthbench-raw. • Self-reported

65.4%

HealthBench Professional

System Card Table 8.1.A length-adjusted score 69.2% at max effort. Raw (pre-length-adjustment) 77.1% is stored separately as healthbench-professional-raw (§8.15.2). Adaptive thinking at max effort. Length-adjusted per HealthBench Professional paper method. • Self-reported

69.2%

HealthBench Professional (raw)

System Card §8.15.2. HealthBench Professional raw (before length adjustment) 77.1% at max effort; level with Opus 5.5 raw. Length-adjusted 69.2% is stored as healthbench-professional (Table 8.1.A). • Self-reported

77.1%

HealthBench (raw)

System Card §8.15.1. HealthBench raw (before length adjustment) 69.4% at max effort (slightly ahead of Opus 5.5 raw 68.1%). Length-adjusted band max 65.4% is stored as healthbench (§8.15.4). • Self-reported

69.4%

Humanity's Last Exam (no tools, text-only)

System Card Table 8.1.A / §8.11.1. HLE no tools (reasoning-only); adaptive thinking at max effort. • Self-reported

56.9%

Humanity's Last Exam (with tools)

Announcement + System Card Table 8.1.A / §8.11.1. HLE with tools (web search, web fetch, programmatic tool calling, code execution); adaptive thinking at max effort. • Self-reported

64.5%

LatchBio SingleCellBench

System Card §8.17.2. LatchBio SingleCellBench 59.1%; adaptive thinking at max effort; bash/file tools. • Self-reported

59.1%

LatchBio SpatialBench Verified

System Card §8.17.2. LatchBio SpatialBench Verified 72.5%; adaptive thinking at max effort; bash/file tools. • Self-reported

72.5%

Legal Agent Benchmark

System Card §8.14.2. Legal Agent Benchmark held-out set; all-pass rate 11.7% at high effort (AA harness). Mean criterion-pass rate 92.1% noted in card but not stored as a separate catalog metric (all-pass is the standard LAB reporting id). At max effort: 10.0% all-pass / 93.1% mean criterion-pass. • Self-reported

11.7%

Medicinal Chemistry (ADME)

System Card §8.17.4. Medicinal Chemistry ADME 65.3% (504 questions); adaptive thinking at max effort. • Self-reported

65.3%

MILU

System Card §8.16.2. MILU average accuracy 91.6% across 11 languages; adaptive thinking at max effort; averaged over 5 trials; no tools. • Self-reported

91.6%

Morphology-to-molecule matching

System Card §8.17.3. Axiom Bio morphology-to-molecule matching 25.0%; adaptive thinking at max effort. • Self-reported

25.0%

OfficeQA

System Card §8.14.1. OfficeQA full set; max effort; agentic harness with extracted-text documents and code-execution tools. • Self-reported

76.9%

OfficeQA Pro

System Card §8.14.1. OfficeQA Pro (133-question subset); max effort. • Self-reported

65.6%

OSWorld 2.1 (partial)

Announcement + System Card Table 8.1.A / §8.13.3. OSWorld 2.1 partial-credit score 80.1% (Pass@1, 5-run average) at max effort. Not OSWorld 2.0. • Self-reported

80.1%

OSWorld 2.1 (strict)

System Card §8.13.3. OSWorld 2.1 strict pass rate 43.5% (every checkpoint satisfied; Pass@1, 5-run average) at max effort. Not OSWorld 2.0. • Self-reported

43.5%

PhysicianBench

System Card §8.15.3. PhysicianBench 100 physician EHR tasks; 63.2% pass rate at max effort. Safety classifiers enabled; graded with Claude Opus 5. • Self-reported

63.2%

Program Bench

System Card §8.10.1. ProgramBench 166 golden tasks; mini-swe-agent harness; no 6-hour time limit; hidden test pass rate against tests the reference binary passes. • Self-reported

79.7%

Protein Design — Library Ranking

System Card §8.17.5. Protein Design Library Ranking 54.8%; 3 attempts per problem; adaptive thinking at max effort. • Self-reported

54.8%

Protein Design — Sequence Generation

System Card §8.17.5. Protein Design Sequence Generation 51.0%; 1 attempt per problem; adaptive thinking at max effort. • Self-reported

51.0%

Protocols Troubleshooting

System Card §8.17.8. Protocols Troubleshooting 67.3%; adaptive thinking at max effort. • Self-reported

67.3%

Protocols Understanding (V2)

System Card §8.17.8. Protocols Understanding (V2) (Benchling) 66.6%; adaptive thinking at max effort. • Self-reported

66.6%

SWE-bench Multilingual

System Card Table 8.1.A / §8.2. Adaptive thinking at max effort; 5-trial average. 300 problems across nine programming languages. • Self-reported

90.3%

SWE-Bench Multimodal

System Card Table 8.1.A / §8.2. Adaptive thinking at max effort; 5-trial average. Visual context (screenshots and design mockups). • Self-reported

54.3%

SWE-Bench Pro

System Card Table 8.1.A / §8.2. Adaptive thinking at max effort; 5-trial average. • Self-reported

81.3%

Terminal-Bench 4.0

Announcement table + System Card §8.5. Terminal-Bench 4.0 accuracy 70.6% with production safeguards enabled (Claude Code --bare, max thinking effort; 5-trial average). Safeguard-flagged requests used the default server-side fallback (1.2% of requests). • Self-reported

70.6%

Terminal-Bench-Science 0.1

System Card §8.6. Terminal-Bench-Science 0.1; 59.9% with safeguards enabled (no fallbacks fired). Claude Code --bare, max thinking effort. • Self-reported

59.9%

Toolathlon-Verified

System Card §8.14.5. Toolathlon Verified Pass@1 77.8% (3-trial average across 108 tasks). Adaptive thinking at max effort; safety/sandbox stops counted as failures (0.6% of trials). • Self-reported

77.8%

License & Metadata

License

proprietary

Announcement Date

September 28, 2026

Last Updated

September 29, 2026

Compare Claude Sonnet 5.5

All comparisons

vs Claude Opus 5.5 vs Claude Fable 5.1 vs Claude Opus 5 vs Qwen3.8 Flash Next vs MAI-Code-1.1-Flash vs Kimi K3 vs Gemini 3.6 Flash vs Ling 3.0 Flash

Similar Models

All Models

Claude Fable 5

Anthropic

Released:Jun 2026

Price:$10.00/1M tokens

Claude Fable 5.1

Anthropic

Released:Sep 2026

Price:$10.00/1M tokens

Claude Sonnet 5

Anthropic

Released:Jun 2026

Price:$2.00/1M tokens

Claude Opus 5

Anthropic

Released:Jul 2026

Price:$5.00/1M tokens

Claude Opus 5.5

Anthropic

Released:Sep 2026

Price:$4.00/1M tokens

Claude Opus 4.8

Anthropic

Best score:0.9 (GPQA)

Released:May 2026

Price:$5.00/1M tokens

Claude Sonnet 4

Anthropic

Best score:0.8 (GPQA)

Released:May 2025

Price:$3.00/1M tokens

Claude Opus 4.7

Anthropic

Best score:0.9 (GPQA)

Released:Apr 2026

Price:$5.00/1M tokens

Recommendations are based on similarity of characteristics: developer organization, multimodality, parameter size, and benchmark performance. Choose a model to compare or go to the full catalog to browse all available AI models.