Claude Haiku 5.5

Name: Claude Haiku 5.5
Author: Anthropic

Multimodal

Anthropic

Claude Haiku 5.5 is Anthropic's cheapest, fastest, and most capable Haiku-class model and the successor to Claude Haiku 4.5, aimed at high-volume, cost-sensitive work such as summarization, compaction, classification, and database queries. It is the first Haiku-class model with adjustable effort (low/medium/high/xhigh/max, Claude Platform default medium) and pairs with Opus 5.5 and Sonnet 5.5 as a coding subagent. It accepts text, image, and PDF input with text output, and supports multilingual and vision workloads, tool use, and adaptive thinking. Context window is 1M tokens with 128K max output and a June 2026 knowledge cutoff. Claude Platform list pricing is $0.10/$0.50 per million input/output for prompts up to 100K tokens (cache read $0.01, 5-minute cache write $0.125). Capability scores are self-reported from the Claude Haiku 5.5 System Card (Table 8.1.A and §8) and the launch announcement. Available on the Claude API as `claude-haiku-5-5`.

Key Specifications

Parameters

Context

1.0M

Release Date

October 7, 2026

Average Score

57.7%

API Documentation Research Paper Results Blog

Timeline

Key dates in the model's history

Announcement

October 7, 2026

Last Update

October 8, 2026

Today

October 11, 2026

Technical Specifications

Parameters

Training Tokens

Knowledge Cutoff

June 1, 2026

Family

Capabilities

MultimodalZeroEval

Pricing & Availability

Input (per 1M tokens)

$0.10

Output (per 1M tokens)

$0.50

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 / §8.10.3. AA-Briefcase v1.1 Elo (max_score 3000) at max effort. Artificial Analysis independent evaluation. At medium (API default) effort: 1372 Elo. • Self-reported

52.6%

BenchCAD

System Card §8.9.2. BenchCAD Vision2Code 1k subset; voxel IoU 0.670 without tools; adaptive thinking at max effort; averaged over five runs. • Self-reported

67.0%

BenchCAD (with Python tool)

System Card §8.9.2. BenchCAD Vision2Code 1k subset; voxel IoU 0.870 with tools (container with image files and standard libraries, plus an image cropping tool — not a Python CAD tool specifically); adaptive thinking at max effort; averaged over five runs. Kept on the existing with-tools BenchCAD catalog ID (named python-tool); no closer ID exists. • Self-reported

87.0%

Biomedical image analysis

System Card §8.13.5. Biomedical image analysis 43.5%; 1 attempt per problem; adaptive thinking at max effort. • Self-reported

43.5%

Chartography

System Card §8.9.1. Chartography with tools 86.2%; adaptive thinking at max effort; container with image file, standard libraries, and image cropping tool. • Self-reported

86.2%

Chartography (no tools)

Announcement + System Card §8.9.1. Chartography without tools 46.4%; adaptive thinking at max effort. • Self-reported

46.4%

FrontierCode 1.1

Announcement + System Card Table 8.1.A / §8.3. FrontierCode v1.1 Main at max effort 46.4% (best Haiku 5.5 effort). Extended 58.4% stored separately. • Self-reported

46.4%

FrontierCode 1.1 (Extended)

System Card §8.3. FrontierCode v1.1 Extended at max effort 58.4%. • Self-reported

58.4%

FrontierSWE V2

System Card §8.6. FrontierSWE v2; Proximal harness; max effort; 5 trials per task; mean across trials. • Self-reported

43.8%

GDPval-AA 2.1

Announcement + System Card Table 8.1.A / §8.10.2. GDPval-AA v2.1 Elo (max_score 3000) at max effort. Artificial Analysis independent evaluation. At medium (API default) effort: 1277 Elo. • Self-reported

54.0%

Global-MMLU

System Card §8.12.1. GMMLU average accuracy 87.8% across 42 languages; adaptive thinking at max effort; single trial; no tools. • Self-reported

87.8%

HealthBench

System Card §8.11.3. HealthBench length-adjusted 61.6% at max effort. Adaptive thinking at max effort. Raw HealthBench at max was not explicitly reported in the card text. • Self-reported

61.6%

HealthBench Professional

System Card Table 8.1.A / §8.11.3. HealthBench Professional length-adjusted score 64.8% at max effort. Raw 71.0% stored separately as healthbench-professional-raw. Adaptive thinking at max effort. • Self-reported

64.8%

HealthBench Professional (raw)

System Card §8.11.3. HealthBench Professional raw (before length adjustment) 71.0% at max effort. Length-adjusted 64.8% is stored as healthbench-professional. • Self-reported

71.0%

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

Announcement + System Card Table 8.1.A / §8.8.1. HLE without tools; the system card treats HLE as multimodal and does not specify a text-only subset. Adaptive thinking at max effort. Kept on the only existing no-tools HLE catalog ID (historical text-only name); no separate multimodal no-tools ID exists. • Self-reported

45.9%

Humanity's Last Exam (with tools)

Announcement + System Card Table 8.1.A / §8.8.1. HLE with tools (web search, web fetch, programmatic tool calling, code execution); adaptive thinking at max effort. • Self-reported

57.4%

LatchBio SingleCellBench

System Card §8.13.1. LatchBio SingleCellBench 56.2%; adaptive thinking at max effort; bash/file tools. Biology safeguards off for this capability eval. • Self-reported

56.2%

LatchBio SpatialBench Verified

System Card §8.13.1. LatchBio SpatialBench Verified 67.7%; adaptive thinking at max effort; bash/file tools. Biology safeguards off for this capability eval. • Self-reported

67.7%

Medicinal Chemistry (ADME)

System Card §8.13.3. Medicinal Chemistry ADME 58.0% (504 questions); adaptive thinking at max effort. • Self-reported

58.0%

MILU

System Card §8.12.2. MILU average accuracy 87.6% across 11 languages; adaptive thinking at max effort; averaged over 5 trials; no tools. • Self-reported

87.6%

Morphology-to-molecule matching

System Card §8.13.2. Axiom Bio morphology-to-molecule matching 25.4%; adaptive thinking at max effort. • Self-reported

25.4%

OfficeQA

System Card §8.10.1. OfficeQA full set; max effort; agentic harness with extracted-text documents (mean of five runs). • Self-reported

73.5%

OfficeQA Pro

System Card §8.10.1. OfficeQA Pro (133-question subset); max effort (mean of five runs). • Self-reported

60.3%

OSWorld 2.1 (partial)

Announcement + System Card Table 8.1.A / §8.9.3. OSWorld 2.1 offline subset (82 of 108 tasks, no internet) partial-credit score 72.4% (Pass@1, 5-run average) at max effort. • Self-reported

72.4%

OSWorld 2.1 (strict)

System Card §8.9.3. OSWorld 2.1 offline subset strict pass rate 37.1% (every checkpoint satisfied; Pass@1, 5-run average) at max effort. • Self-reported

37.1%

PhysicianBench

System Card §8.11.2. PhysicianBench 100 physician EHR tasks; 43.0% pass rate at max effort (five attempts per task). Safety classifiers enabled. • Self-reported

43.0%

Program Bench

System Card §8.7.1. ProgramBench 166 tasks; mini-swe-agent harness; no 6-hour time limit; hidden test pass rate against tests the reference binary passes. • Self-reported

82.0%

Protein Design — Library Ranking

System Card §8.13.4. Protein Design Library Ranking 45.3%; 3 attempts per problem; adaptive thinking at max effort. • Self-reported

45.3%

Protein Design — Sequence Generation

System Card §8.13.4. Protein Design Sequence Generation 33.1%; 1 attempt per problem; adaptive thinking at max effort. • Self-reported

33.1%

Protocols Troubleshooting

System Card §8.13.6. Protocols Troubleshooting 58.0%; adaptive thinking at max effort. • Self-reported

58.0%

Protocols Understanding (V2)

System Card §8.13.6. Protocols Understanding (V2) (Benchling) 64.1%; adaptive thinking at max effort. • Self-reported

64.1%

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

83.7%

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

30.7%

SWE-Bench Pro

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

64.8%

Terminal-Bench 4.0

Announcement + System Card §8.4. Terminal-Bench 4.0 accuracy 39.2% with safeguards enabled (Claude Code --bare, max thinking effort; 10-trial average). No fallback model; safeguard-flagged requests failed (1.8% of trials). • Self-reported

39.2%

Terminal-Bench-Science 0.1

System Card §8.5. Terminal-Bench-Science 0.1; 20.6% with safeguards enabled (no fallback; 2.3% of trials flagged). Claude Code --bare, max thinking effort. • Self-reported

20.6%

License & Metadata

License

proprietary

Announcement Date

October 7, 2026

Last Updated

October 8, 2026

Compare Claude Haiku 5.5

All comparisons

vs Claude Fable 5.1 vs Claude Opus 5 vs Claude Sonnet 5.5 vs Claude Opus 5.5 vs DeepSeek-V4-Flash-0731 vs Laguna S 2.1 vs Hy4 preview vs IBM Granite 4.2 8B

Similar Models

All Models

Claude Sonnet 5.5

Anthropic

Released:Sep 2026

Price:$2.00/1M tokens

Claude Sonnet 4.6

Anthropic

Best score:0.9 (TAU)

Released:Feb 2026

Price:$3.00/1M tokens

Claude Opus 4

Anthropic

Best score:0.8 (GPQA)

Released:May 2025

Price:$15.00/1M tokens

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 Mythos Preview

Anthropic

Best score:0.9 (GPQA)

Released:Apr 2026

Price:$25.00/1M 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.