Claude Haiku 5.5
Anthropic’s Claude 5.5 model for high-volume classification, extraction, routing, summaries and coding subagents, with adaptive thinking.
Where it stands
Rank compares tracked models with a score. Published shows the source’s rank and field size, when available.
| Benchmark | Score | Rank | Published | Leader | Behind |
|---|---|---|---|---|---|
| Artificial Analysis Intelligence IndexArtificial Analysis | 43 | 19of 42 | no list read | Claude Opus 5.5 | 15 pts |
| Terminal-Bench 4.0Artificial Analysis | 32.83% | 17of 40 | no list read | Claude Sonnet 5.5 | 30.81 pts |
| Terminal-Bench Science 0.1Artificial Analysis | 20.00% | 9of 24 | no list read | GPT-6 Astra | 43.33 pts |
| Harvey LAB-AAArtificial Analysis | 89.87% | 13of 20 | no list read | Muse Spark 1.3 Max | 5.61 pts |
| AutomationBench-AAArtificial Analysis | 35.41% | 37of 40 | no list read | Gemini 4 Argon | 42.10 pts |
| SciCodeArtificial Analysis | 54.98% | 21of 40 | no list read | Claude Opus 5.5 | 11.92 pts |
| GDP.pdfArtificial Analysis | 20.80% | 21of 40 | no list read | GPT-6.1 Sol | 10.20 pts |
| CritPtArtificial Analysis | 18.86% | 25of 42 | no list read | GPT-5.6 Sol | 13.43 pts |
| AA-OmniscienceArtificial Analysis | 10.67 | 27of 42 | no list read | Claude Opus 5.5 | 35.75 |
| AA-LCR v1.1Artificial Analysis | 82.67% | 16of 42 | no list read | Kimi K3 | 6.00 pts |
| GDPval-AA v2.1Artificial Analysis | 1,620.05 | 12of 42 | no list read | Claude Opus 5.5 | 226.12 |
| AA-Briefcase v1.1Artificial Analysis | 1,577.72 | 9of 41 | no list read | Claude Opus 5.5 | 244.13 |
| Humanity's Last ExamArtificial Analysis | 44.39% | 21of 42 | no list read | Claude Opus 5.5 | 16.96 pts |
Nearby on Artificial Analysis Intelligence Index
The two models above and below on the Artificial Analysis Intelligence Index (Artificial Analysis), on every column where this model has a published number.
| Model | Artificial Analysis Intelligence Index | Terminal-Bench 4.0 | Terminal-Bench Science 0.1 |
|---|---|---|---|
| Kimi K3Moonshot AI | 44 | 12.63% | – |
| Step 5 PreviewStepFun | 44 | 33.33% | 2.38% |
| Claude Haiku 5.5Anthropic | 43 | 32.83% | 20.00% |
| Claude Opus 4.8Anthropic | 42 | 21.72% | – |
| GLM 5.3 FlashZ.ai | 42 | 32.83% | 4.76% |
Cost and speed, as published
Additional measurements from Artificial Analysis’ own row. Read 8 October 2026.
| As published | Cost per task | Output speed |
|---|---|---|
| Claude Haiku 5.5 (Max) | $0.21 | 243.4 tok/s |
Results by benchmark
Independent and vendor results retain their own harnesses and effort settings.
| Benchmark | Independent run | As reported by a publisher |
|---|---|---|
| SWE-Bench Pro (version unspecified)Repository coding | no published result | 64.8%SWE-Bench Pro (version unspecified)Anthropic · 7 October 2026Rank 2 of 2 in this publisher table |
| SWE-bench MultimodalRepository coding | no published result | 30.7%SWE-bench MultimodalAnthropic · 7 October 2026Rank 2 of 3 in this publisher table |
| FrontierCode v1.1 (Extended)Agentic coding | no published result | 58.4%FrontierCode v1.1 (Extended)Anthropic · 7 October 2026Rank 2 of 2 in this publisher table |
| ProgramBench (hidden test pass rate)Long-context coding | no published result | 82.0%ProgramBench (hidden test pass rate)Anthropic · 7 October 2026Rank 2 of 3 in this publisher table |
| DRACO (Anthropic harness)Deep research | no published result | 81.5%DRACO (Anthropic harness)Anthropic · 7 October 2026Rank 3 of 4 in this publisher table |
| WANDR (frozen index)Wide research | no published result | 49.9%WANDR (frozen index)Anthropic · 7 October 2026Rank 3 of 4 in this publisher table |
| BenchCAD Vision2Code (no tools)CAD reasoning | no published result | 0.67BenchCAD Vision2Code (no tools)Anthropic · 7 October 2026Rank 3 of 5 in this publisher table |
| BenchCAD Vision2Code (with tools)CAD reasoning | no published result | 0.87BenchCAD Vision2Code (with tools)Anthropic · 7 October 2026Rank 3 of 5 in this publisher table |
| OSWorld 2.1 (offline strict pass)Computer use | no published result | 37.1%OSWorld 2.1 (offline strict pass)Anthropic · 7 October 2026Rank 4 of 6 in this publisher table |
| OfficeQADocument reasoning | no published result | 73.5%OfficeQAAnthropic · 7 October 2026Rank 2 of 3 in this publisher table |
| OfficeQA ProDocument reasoning | no published result | 60.3%OfficeQA ProAnthropic · 7 October 2026Rank 2 of 3 in this publisher table |
| HealthBench (length-adjusted)Healthcare | no published result | 61.6%HealthBench (length-adjusted)Anthropic · 7 October 2026Rank 1 of 2 in this publisher table |
| HealthBench (raw)Healthcare | no published result | 64.2%HealthBench (raw)Anthropic · 7 October 2026Rank 1 of 1 in this publisher table |
| HealthBench Professional (length-adjusted)Healthcare | no published result | 64.8%HealthBench Professional (length-adjusted)Anthropic · 7 October 2026Rank 3 of 4 in this publisher table |
| HealthBench Professional (raw)Healthcare | no published result | 71.0%HealthBench Professional (raw)Anthropic · 7 October 2026Rank 1 of 1 in this publisher table |
| PhysicianBenchHealthcare | no published result | 43.0%PhysicianBenchAnthropic · 7 October 2026Rank 4 of 6 in this publisher table |
| Global MMLUMultilingual knowledge | no published result | 87.8%Global MMLUAnthropic · 7 October 2026Rank 4 of 4 in this publisher table |
| MILUMultilingual knowledge | no published result | 87.6%MILUAnthropic · 7 October 2026Rank 4 of 4 in this publisher table |
| SpatialBench VerifiedLife sciences | no published result | 67.7%SpatialBench VerifiedAnthropic · 7 October 2026Rank 2 of 2 in this publisher table |
| SingleCellBenchLife sciences | no published result | 56.2%SingleCellBenchAnthropic · 7 October 2026Rank 2 of 2 in this publisher table |
| Morphology-to-molecule matchingLife sciences | no published result | 25.4%Morphology-to-molecule matchingAnthropic · 7 October 2026Rank 1 of 2 in this publisher table |
| Medicinal chemistry (ADME)Life sciences | no published result | 58.0%Medicinal chemistry (ADME)Anthropic · 7 October 2026Rank 3 of 4 in this publisher table |
| Protein design (sequence generation)Life sciences | no published result | 33.1%Protein design (sequence generation)Anthropic · 7 October 2026Rank 2 of 3 in this publisher table |
| Protein design (library ranking)Life sciences | no published result | 45.3%Protein design (library ranking)Anthropic · 7 October 2026Rank 2 of 3 in this publisher table |
| Biomedical image analysisLife sciences | no published result | 43.5%Biomedical image analysisAnthropic · 7 October 2026Rank 1 of 2 in this publisher table |
| Protocols (troubleshooting)Life sciences | no published result | 58.0%Protocols (troubleshooting)Anthropic · 7 October 2026Rank 2 of 3 in this publisher table |
| Protocols (understanding V2)Life sciences | no published result | 64.1%Protocols (understanding V2)Anthropic · 7 October 2026Rank 2 of 3 in this publisher table |
| Harvey LAB-AALegal reasoning | 89.87%Artificial Analysis | not published |
| SWE-bench MultilingualRepository coding | no published result | 83.7%SWEBench MultilingualAnthropic · 7 October 2026Rank 2 of 3 in this publisher table |
| AutomationBench-AAWorkflow automation | 35.41%Artificial Analysis | not published |
| SciCodeScientific coding | 54.98%Artificial Analysis | not published |
| GDP.pdfDocument reasoning | 20.80%Artificial Analysis | not published |
| CritPtScientific reasoning | 18.86%Artificial Analysis | not published |
| AA-OmniscienceFactual knowledge | 10.67Artificial Analysis | not published |
| AA-LCR v1.1Long-context reasoning | 82.67%Artificial Analysis | not published |
| Terminal-Bench 4.0Agentic coding | 32.83%Artificial Analysis | 39.2%Terminal-Bench 4.0Anthropic · 7 October 2026Rank 2 of 4 in this publisher table 39.2%Terminal-Bench 4.0Anthropic · 7 October 2026Rank 4 of 6 in this publisher table |
| Terminal-Bench Science 0.1Agentic science | 20.00%Artificial Analysis | 20.6%Terminal-Bench-Science 0.1Anthropic · 7 October 2026Rank 4 of 5 in this publisher table |
| OSWorld 2.1Computer use | no published result | 72.4%OSWorld 2.1 (offline partial)Anthropic · 7 October 2026Rank 2 of 4 in this publisher table 72.4%OSWorld 2.1 (offline partial)Anthropic · 7 October 2026Rank 4 of 6 in this publisher table |
| FrontierCode v1.1Agentic coding | no published result | 46.4%FrontierCode v1.1 (Main)Anthropic · 7 October 2026Rank 2 of 3 in this publisher table 46.4%FrontierCode v1.1 (Main, max)Anthropic · 7 October 2026Rank 1 of 3 in this publisher table 45.8%FrontierCode v1.1 (Main, xhigh)Anthropic · 7 October 2026Rank 2 of 2 in this publisher table |
| GDPval-AA v2.1Knowledge work | 1,620.05Artificial Analysis | 1,620GDPval-AA v2.1 (max)Anthropic · 7 October 2026Rank 2 of 4 in this publisher table 1,620GDPval-AA v2.1 (max)Anthropic · 7 October 2026Rank 2 of 5 in this publisher table 1,277GDPval-AA v2.1 (medium)Anthropic · 7 October 2026Rank 1 of 1 in this publisher table |
| AA-Briefcase v1.1Knowledge work | 1,577.72Artificial Analysis | 1,578AA-Briefcase v1.1 (max)Anthropic · 7 October 2026Rank 2 of 4 in this publisher table 1,578AA-Briefcase v1.1 (max)Anthropic · 7 October 2026Rank 2 of 5 in this publisher table 1,372AA-Briefcase v1.1 (medium)Anthropic · 7 October 2026Rank 1 of 1 in this publisher table |
| ChartographyVisual chart recognition | no published result | 46.4%Chartography (no tools)Anthropic · 7 October 2026Rank 2 of 4 in this publisher table 46.4%Chartography (no tools)Anthropic · 7 October 2026Rank 4 of 10 in this publisher table 86.2%Chartography (with tools)Anthropic · 7 October 2026Rank 3 of 5 in this publisher table |
| Humanity's Last ExamMultidisciplinary reasoning | 44.39%Artificial Analysis | 45.9%Humanity's Last Exam (no tools)Anthropic · 7 October 2026Rank 2 of 3 in this publisher table 57.4%Humanity's Last Exam (with tools)Anthropic · 7 October 2026Rank 2 of 3 in this publisher table 45.9%Humanity's Last Exam (no tools)Anthropic · 7 October 2026Rank 2 of 3 in this publisher table 57.4%Humanity's Last Exam (with tools)Anthropic · 7 October 2026Rank 2 of 3 in this publisher table |
| Artificial Analysis Intelligence IndexComposite | 43Artificial Analysis | not published |
| FrontierSWEAgentic research | no published result | 43.8%FrontierSWE v2Anthropic · 7 October 2026Rank 2 of 3 in this publisher table |
Token pricing
USD per million tokens. The full input length determines the rate for every token in the request. Message Batches discounts input and output by 50%.
| Prompt length | Input | Output | Cache read | Cache write · 5m | Cache write · 1h |
|---|---|---|---|---|---|
| Up to 100,000 tokens | $0.10 | $0.50 | $0.01 | $0.125 | $0.20 |
| Over 100,000 tokens | $0.50 | $2.50 | $0.05 | $0.625 | $1.00 |
Artificial Analysis by effort
Independent evaluations and Anthropic API measurements. First answer token includes input processing and any thinking; the response time covers 500 output tokens. The comparison table uses the highest published effort. Read 8 October 2026.
| Measure | max | xhigh | high | medium | low |
|---|---|---|---|---|---|
| AA Index | 43 | 41 | 38 | 34 | 29 |
| Cost per task | $0.213 | $0.124 | $0.079 | $0.047 | $0.024 |
| Output speed | 243.4 tok/s | 188.3 tok/s | 173.2 tok/s | 137.3 tok/s | 181.2 tok/s |
| First answer token | 323.26 s | 70.88 s | 26.15 s | 12.63 s | 11.99 s |
| 500-token response | 325.32 s | 73.54 s | 29.03 s | 16.27 s | 14.75 s |
| Terminal-Bench 4.0% | 32.83% | 29.29% | 21.72% | 15.15% | 12.63% |
| Terminal-Bench Science 0.1% | 20.00% | 18.10% | 10.00% | 7.14% | 1.90% |
| AA-Briefcase v1.1Elo | 1,577.72 | 1,532.67 | 1,442.55 | 1,371.7 | 1,111.58 |
| GDPval-AA v2.1Elo | 1,620.05 | 1,510.76 | 1,420.03 | 1,276.84 | 1,124.94 |
| AutomationBench-AA% | 35.41% | 35.95% | 33.73% | 28.59% | 22.91% |
| SciCode% | 54.98% | 51.74% | 48.73% | 48.96% | 49.19% |
| Humanity's Last Exam% | 44.39% | 42.68% | 37.26% | 33.78% | 27.02% |
| GDP.pdf% | 20.80% | 18.20% | 17.20% | 15.20% | 11.20% |
| CritPt% | 18.86% | 22.57% | 18.57% | 12.86% | 9.14% |
| AA-Omnisciencescore | 10.67 | 6.05 | 5.75 | 4.47 | 3.25 |
| AA-LCR v1.1% | 82.67% | 78.33% | 77.33% | 77.33% | 72.00% |
| Harvey LAB-AA% | 89.87% | 89.90% | 89.52% | 89.23% | 87.86% |
Source notes
From the linked sources. Read 8 October 2026.
API model and availability. claude-haiku-5-5 on the Claude API, Google Cloud, Microsoft Foundry and Claude Platform on AWS; anthropic.claude-haiku-5-5 on Amazon Bedrock. Active, released 7 October 2026. Not sooner than 7 October 2027 for retirement.
Modalities and knowledge. Text and images → text. Reliable knowledge and training-data cutoff: June 2026. Anthropic describes it as its fastest standard-speed model; this is a publisher claim, not an independently measured token speed.
Context and output. 1M-token context and 128k output tokens, including thinking. Message Batches supports up to 300k output tokens in beta with output-300k-2026-03-24; this is separate from the standard output limit.
Prompt-length pricing. Rates are USD per million tokens and depend on the full prompt length, including cached tokens. Prompts up to 100,000 tokens use the lower tier; prompts over 100,000 use the higher rate for all input and output tokens in that request. Message Batches discounts input and output by 50%.
Thinking and effort. Adaptive thinking is on by default; effort levels are low, medium (default), high, xhigh and max. Thinking can be disabled at high effort or below. Manual budget_tokens is unsupported. Thinking text is omitted by default; use thinking.display: summarized for summaries.
API migration. Omit temperature, top_p and top_k; non-default values return 400. Assistant prefill is unsupported. Read response blocks by type, since thinking may precede text. Keep conversations append-only; changing earlier turns invalidates returned thinking blocks.
Tools and safeguards. Supports tool-driven work, computer use and browser use. Browser use is available on the Claude API and Google Cloud. Replace computer_20250124 with computer_toolset_20260801 there. Handle stop_reason: refusal; server-side fallback is unavailable.
Tokenizer and account state. The newer tokenizer counts the same text as approximately 30% more tokens than Haiku 4.5; recount prompts and output budgets. Thinking blocks can be replayed only in the account that produced them or a linked account.
Evaluation provenance. Launch-table and system-card capability results retain their metrics, task subsets and effort settings. AA independently publishes an Intelligence Index of 43 and twelve benchmark results at max effort. AA’s Terminal-Bench 4.0 score is 32.83%, distinct from Anthropic’s 39.2% Claude Code report. All five AA effort configurations retain their benchmark scores, weighted task costs, output speeds, first-answer-token times and 500-token response times. API performance is measured by AA on Anthropic’s API using its long-prompt medians. Vals has no Haiku 5.5 entry in the checked public directory.
Sources
Read on 17 October 2026.
- Anthropic announcement and evaluation tablewww.anthropic.com/claude-haiku-5-5
- Anthropic system cardwww.anthropic.com/claude-haiku-5-5-system-card
- Claude Platform model specifications and pricingplatform.claude.com/docs/en/models/haiku-5-5/overview
- Claude Platform capabilities and breaking changesplatform.claude.com/docs/en/models/haiku-5-5/whats-new-haiku-5-5
- Artificial Analysis independent evaluation and effort configurationsartificialanalysis.ai/models/claude-haiku-5-5