Haiku 5.5
Haiku 5.5 is behind the leaders in reasoning. Too few results yet to rate coding, agentic tasks, safety, long context, math, multimodal tasks, multilingual tasks, instruction following, or factuality.
Capability profile
Bars show the model's median result as a share of the leading model's, per capability. Select a row to see its results.
Too few results yet to rate Coding, Agentic, Safety, Long Context, Math, Multimodal, Multilingual, Instruction Following or Factuality.
Price
No current price is tracked for this model. See the rate card
Evidence
32results on31benchmarks
- 32 vendor-reported
From 4 sources · latest Oct 7, 2026 · How verification works
Haiku 5.5 benchmark results
32 results on 31 benchmarks, grouped by capability. Each bar is the result as a share of the capability leader's; every result links to its source.
31.9% behind the leader1 of 6 ranked benchmarks measured
- Humanity's Last Exam57.40Oct 7, 2026Humanity's Last Exam with tools
0 of 10 ranked benchmarks measured
- 83.70Oct 7, 2026
- 64.80Oct 7, 2026
More results
Benchmarks outside the capability baskets. They are not ranked against a leader.
- Terminal-Bench 4.039.20Oct 7, 2026
- OSWorld 2.172.40Oct 7, 2026
- GDPval-AA 2.11620.00Oct 7, 2026GDPval-AA v2.1
- Pairwise political bias: refusals (Claude.ai)5.00Oct 7, 2026
- Pairwise political bias: refusals (Public API)3.30Oct 7, 2026
- OSWorld 2.1 (offline subset)72.00Oct 7, 2026
Show 23 more resultsHide 23 results
- Child safety multi-turn appropriate response rate (API, without a system prompt)96.00Oct 7, 2026
- Child safety multi-turn appropriate response rate (Claude.ai)99.00Oct 7, 2026
- Child safety single-turn harmful prompts decline rate (API)99.29Oct 7, 2026
- Election integrity multi-turn (appropriate response rate) - API, without a system prompt99.00Oct 7, 2026
- Election integrity multi-turn (appropriate response rate) - Claude.ai98.00Oct 7, 2026
- Election integrity single-turn benign requests (refusal rate) - API, without a system prompt0.33Oct 7, 2026
- Election integrity single-turn benign requests (refusal rate) - Claude.ai0.33Oct 7, 2026
- Election integrity single-turn harmful requests (harmless rate) - API, without a system prompt100.00Oct 7, 2026
- Election integrity single-turn harmful requests (harmless rate) - Claude.ai100.00Oct 7, 2026
- FrontierCode 1.1 (Main) max46.40Oct 7, 2026
- FrontierCode 1.1 (Main) xhigh45.80Oct 7, 2026
- FrontierCode Extended58.40Oct 7, 2026
- HealthBench Professional length-adjusted64.80Oct 7, 2026HealthBench Professional (Length-adjusted)
- OSWorld 2.1 (offline subset)73.00Oct 7, 2026
- 82.00Oct 7, 2026
- Suicide and self-harm multi-turn appropriate response rate (API, without a system prompt)70.00Oct 7, 2026
- Suicide and self-harm multi-turn appropriate response rate (Claude.ai)90.00Oct 7, 2026
- Suicide and self-harm single-turn benign requests refusal rate (API, without a system prompt)0.00Oct 7, 2026
- Suicide and self-harm single-turn benign requests refusal rate (Claude.ai)0.41Oct 7, 2026
- Suicide and self-harm single-turn requests posing potential risk harmless rate (API, without a system prompt)99.61Oct 7, 2026
- Suicide and self-harm single-turn requests posing potential risk harmless rate (Claude.ai)100.00Oct 7, 2026
- 30.70Oct 7, 2026
- Terminal-Bench-Science 0.120.60Oct 7, 2026
Haiku 5.5: common questions
Who makes Haiku 5.5?
Haiku 5.5 is made by Anthropic.
When was Haiku 5.5 released?
Haiku 5.5 was released on Oct 7, 2026, according to Anthropic's own announcement.
What is Haiku 5.5 good at?
Haiku 5.5 is behind the leaders in reasoning. Too few results yet to rate coding, agentic tasks, safety, long context, math, multimodal tasks, multilingual tasks, instruction following, or factuality.
How many benchmarks has Haiku 5.5 been tested on?
We track 32 results for Haiku 5.5 on 31 benchmarks from 4 sources. The latest was recorded on Oct 7, 2026.
About this record
Where Haiku 5.5's numbers come from, and every name it appears under.
- Tracked since
- Sep 28, 2026
- Newest source mention
- Oct 7, 2026
Where the results come from
Verification: 32 scores · 0 independently verified · 32 vendor-reported. How these tiers are assigned
From 4 sources on 2 sites. www-cdn.anthropic.com supplies 28 of them. Bars are coloured by trust tier.
- www-cdn.anthropic.com28
- anthropic.com4