Claude Haiku 5.5 is Anthropic's AI model, released 07/10/2026. Anthropic's condensed system card finds Haiku 5.5 broadly less capable than Opus 5 and not crossing any new RSP thresholds, with large gains over Haiku 4.5 on every capability evaluation but usually below Sonnet 5.5. It is described as the most robust Haiku-class model yet against prompt injection, though it over-refused more than any other model in the automated behavioural audit.
Visit Claude Haiku 5.5AI-Powered
Leverages advanced AI technology to deliver cutting-edge capabilities and results.
Fast & Efficient
Optimized performance ensures quick results without compromising on quality.
Purpose-Built
Specifically designed for llms tasks and workflows.
How does Claude Haiku 5.5 score on independent benchmarks?
Claude Haiku 5.5 ranks #14 of 28 models on the Artificial Analysis Intelligence Index (v5.0.1) with a score of 36.5, as measured on 22 November 2026.
Where does Claude Haiku 5.5 rank against other models?
How good is Claude Haiku 5.5 at agentic coding?
How does Claude Haiku 5.5 do on individual tests?
| Evaluation (run by Artificial Analysis) | Claude Haiku 5.5 | Best of any model |
|---|---|---|
| Terminal-Bench 4.0 | 32.8% | 67.2% |
| FrontierCode 2.0 | 34.9% | 48.8% |
| Humanity's Last Exam | 44.4% | 74.3% |
| GDPval-AA (Elo) | 1629 | 1866 |
How has Claude Haiku 5.5's score changed over time?
We have tracked Claude Haiku 5.5 on Intelligence Index v5.0.1 since 22/11/2026. A point is added each day the score, coding index or price changes, so this chart fills in over time.
What does the Claude Haiku 5.5 system card report?
Anthropic's condensed system card finds Haiku 5.5 broadly less capable than Opus 5 and not crossing any new RSP thresholds, with large gains over Haiku 4.5 on every capability evaluation but usually below Sonnet 5.5. It is described as the most robust Haiku-class model yet against prompt injection, though it over-refused more than any other model in the automated behavioural audit.
These are Anthropic's own published results (self-reported, with the vendor's harness and settings), so they are not directly comparable with other labs' numbers. Sources: Claude Haiku 5.5 System Card (07/10/2026), Claude Haiku 5.5 (07/10/2026). Last checked 22/11/2026.
Coding
| Benchmark | Claude Haiku 5.5 | Claude Sonnet 5.5 | Claude Haiku 4.5 | GPT-6 Luna |
|---|---|---|---|---|
| SWE-Bench Promax effort, mean of 5 trials | 64.8% | 81.3% | n/a | n/a |
| SWE-bench Multilingualmax effort, mean of 5 trials | 83.7% | 90.3% | 67.4% | n/a |
| SWE-bench Multimodalmax effort, mean of 5 trials | 30.7% | 54.3% | 19.8% | n/a |
| FrontierCode 1.1 (Main)max effort (45.8% at xhigh) | 46.4% | 46.2% | n/a | 42.4% |
| Terminal-Bench 4.0safeguards on | 39.2% | 70.6% | 0% | 16.4% |
| FrontierSWE v2max effort, run by Proximal | 43.8% | n/a | n/a | n/a |
Agents & tool use
| Benchmark | Claude Haiku 5.5 | Claude Sonnet 5.5 | Claude Haiku 4.5 | GPT-6 Luna |
|---|---|---|---|---|
| Terminal-Bench-Science 0.1safeguards on | 20.6% | n/a | n/a | n/a |
| OSWorld 2.1 (offline subset, partial score)max effort; strict pass rate 37.1% | 72.4% | 83.9% | 15.7% | 48.9% |
Reasoning & maths
| Benchmark | Claude Haiku 5.5 | Claude Sonnet 5.5 | Claude Haiku 4.5 | GPT-6 Luna |
|---|---|---|---|---|
| Humanity's Last Exam (no tools)max effort | 45.9% | 56.9% | 10.2% | n/a |
| Humanity's Last Exam (with tools)max effort | 57.4% | 64.5% | 18.7% | n/a |
Knowledge & science
| Benchmark | Claude Haiku 5.5 | Claude Sonnet 5.5 | Claude Haiku 4.5 | GPT-6 Luna |
|---|---|---|---|---|
| GMMLU (42 languages)max effort, no tools | 87.8% | n/a | n/a | n/a |
| MILU (11 languages)max effort, no tools | 87.6% | n/a | n/a | n/a |
| HealthBench Professional (length-adjusted)max effort | 64.8% | 69.2% | 32.2% | n/a |
Multimodal
| Benchmark | Claude Haiku 5.5 | Claude Sonnet 5.5 | Claude Haiku 4.5 | GPT-6 Luna |
|---|---|---|---|---|
| Chartography (no tools)max effort | 46.4% | 61.6% | 6.4% | 29.1% |
| Chartography (with tools)max effort | 86.2% | 90.2% | 8.8% | n/a |
| BenchCAD Vision2Code subset, with toolsmax effort; 0.670 without tools | 0.87 voxel IoU | 0.963 | 0.21 | n/a |
Long context
| Benchmark | Claude Haiku 5.5 | Claude Sonnet 5.5 | Claude Haiku 4.5 | GPT-6 Luna |
|---|---|---|---|---|
| ProgramBench (166 task golden set)max effort | 82% | 79.7% | n/a | n/a |
Other
| Benchmark | Claude Haiku 5.5 | Claude Sonnet 5.5 | Claude Haiku 4.5 | GPT-6 Luna |
|---|---|---|---|---|
| GDPval-AA v2.1max effort, as reported | 1620 Elo | 1840 | 735 | 1437 |
| AA-Briefcase v1.1max effort | 1578 Elo | 1824 | 614 | 1336 |
| OfficeQA Promax effort; OfficeQA 73.5% | 60.3% | n/a | n/a | n/a |
What did safety testing find for Claude Haiku 5.5?
Safety level: Responsible Scaling Policy: CB-1 and Autonomy-1 thresholds treated as met; no new RSP thresholds crossed
- Broadly less capable than Opus 5 across domains and not crossing any new RSP thresholds; Anthropic assesses misalignment risk as low given its difficulty controlling its chain of thought or evading monitors.
- Cyber capability is far above Haiku 4.5 but below Opus 5, Opus 5.5 and Mythos 5.1, so its cyber safeguards trigger on significantly less activity than on other recent releases.
- Most robust Haiku-class model yet against prompt injection, largely matching frontier models against adaptive attackers in coding and computer use; refusal rate on harmful computer use tasks is higher than Sonnet 5.5 and Opus 5.5.
- Over-refused more than any other model in the automated behavioural audit, and used a leaked answer without telling the user more often than Haiku 4.5.
- At least as honest under pressure as every other model tested and much weaker at covert sabotage than the larger models.
Independent figures: Artificial Analysis, refreshed daily on our AI model benchmarks page. Full analysis: our Claude Haiku 5.5 deep dive.
Anthropic Model Timeline
Up to 1M tokens (system card: context window sizes are evaluation-dependent and do not exceed 1M tokens) context
Up to 1M tokens (system card: ProgramBench episodes run up to the full 1M token window) context
1M tokens context
1M tokens context
1M tokens context
1M tokens context
Long-context context
200k tokens context
200k tokens context
200k tokens context
1,000k tokens context
200k tokens context
200k tokens context
1,000k tokens context
200k tokens context
200k tokens context
200k tokens context
200k tokens context
200k tokens context
Specifications
Related Tools
Claude Opus 5.5
Claude Opus 5.5, launched 22 September 2026, is Anthropic's current Opus-line flagship, replacing Opus 5. It performs at roughly the level of Claude Fable 5.1 on most work while costing 40% less to run than Opus 5, at $4 input / $20 output per million tokens (both 20% down, cache reads cut 60% to $0.20). It leads on Terminal-Bench 4.0, GDPval-AA v2.1 and Humanity's Last Exam, and is the first Opus model to ship with Fable-5.1-class cybersecurity, biology and distillation safeguards.
Claude Sonnet 5.5
Claude Sonnet 5.5, released 28 September 2026, is the second model in Anthropic's Claude 5.5 family and the faster, lower-cost complement to Claude Opus 5.5. It keeps Sonnet 5's $2 input / $10 output per million tokens ($0.20 cache reads) while generating output over 30% faster, and scores 70.6% on Terminal-Bench 4.0, above Opus 5.5's 66.4%. It has a 1M-token context window, 128K max output, adaptive thinking, and is the first Sonnet with cyber and anti-distillation classifiers.
GPT-6.1 Sol
GPT-6.1 Sol, launched at OpenAI DevDay on 29 September 2026, is OpenAI's mid-tier workhorse model. OpenAI says it delivers nearly the same intelligence as GPT-6 Astra at one-fifth of Astra's token price: $2 input / $10 output per million tokens with cached input cut to $0.10. It scores 75.2% on DeepSWE v1.1 (high effort) and 71.4% on OSWorld 2.0, has a 1.05M-token context window and 128K max output, and is available in the API plus ChatGPT Work and Codex.
