Anthropic releases Claude Haiku 5.5
Anthropic's Claude Haiku 5.5 scores 43 on Artificial Analysis's Intelligence Index, putting it at the top of the small-model class at a tenth of its predecessor's price, though it uses many tokens.
Anthropic has released Claude Haiku 5.5. Independent testing by Artificial Analysis gives it a score of 43 on the firm's Intelligence Index at max effort. That is 26 points higher than the previous Haiku, which came out a year ago.
It is the first Haiku model with Anthropic's effort settings and adaptive thinking. The score puts it slightly ahead of GLM-5.3 Flash (42), Gemini 3.8 Flash (41) and GPT-6 Luna (38). It is close to Kimi K3 (44), a 2.8T-parameter open-weights model, and 13 points behind Claude Sonnet 5.5 (max, 56).
Pricing. Pricing is now tiered. Prompts up to 100k tokens cost $0.10/$0.50 per 1M input/output tokens, which matches GPT-6 Luna and is 10% of the previous Haiku's price. Above 100k tokens, prices rise fivefold to $0.50/$2.50. The context window is 1M tokens, up from 200k. The model takes text and image input and produces text output.
Token use. At max effort, Haiku 5.5 uses about 162k output tokens per Index task. That is more than Opus 5.5 (max) and roughly 3x GPT-6 Luna (max). Moving from xhigh to max effort adds 2 points for about 1.8x the tokens. Artificial Analysis says its cost figures are provisional because they do not yet account for the higher price above 100k tokens.
Other results:
- AA-Briefcase (knowledge work): 1578 Elo, ahead of Kimi K3 and GLM-5.3.
- Terminal-Bench 4.0: 33%, up from 0% for Haiku 4.5.
- AA-Omniscience: 36% accuracy, below its peers, but with a lower hallucination rate (40%).
- AutomationBench-AA: 35%, which Artificial Analysis expects to rise. A safety-refusal issue in pre-release testing made the model refuse too often; Anthropic is working on a fix, and the test will be re-run.