Anthropic releases Claude Haiku 5.5: 43 points on the AA index and a price of $0.10 per million tokens

Anthropic has released Claude Haiku 5.5, and according to independent measurements by Artificial Analysis it is now the leader among small models: 43 points on the Intelligence Index at maximum effort. A year ago the previous Haiku stood at 17, so the gain is 26 points. The price dropped at the same time: $0.10 per million input tokens and $0.50 per million output tokens for prompts up to 100k. That is roughly 10% of the cost of the previous Haiku and the same rate as GPT-6 Luna.

Where Haiku 5.5 stands in the rankings

At maximum effort, the model edges out GLM-5.3 Flash (42), Gemini 3.8 Flash (41) and GPT-6 Luna (38). Its score is comparable to Kimi K3 (44), an open model with 2.8 trillion parameters, and it trails Claude Sonnet 5.5 (max, 56) by 13 points. Haiku 5.5 is the first model in the lineup with effort settings and adaptive thinking, which means reasoning depth can now be adjusted.

Key metrics according to Artificial Analysis:

• Intelligence Index: 43 (max), 38 at the high level.
• AA-Briefcase, a closed test of realistic knowledge work tasks: 1578 Elo, above Kimi K3 and GLM-5.3 and on par with Muse Spark 1.3 (max). Within confidence intervals, the model matches GPT-6 Astra (max) and Claude Fable 5.1 (high).
• Terminal-Bench 4.0: 33% versus 0% for Haiku 4.5. That is on par with GLM-5.3 Flash and above Gemini 3.8 Flash (20%) and GPT-6 Luna (13%).
• AA-Omniscience: 36% accuracy versus 55% for Gemini 3.8 Flash and 44% for GPT-6 Luna.
• Hallucination rate: 40% versus 55% and 77% for competitors.
• AutomationBench-AA: 35% versus 53–60% for Luna, Gemini 3.8 Flash and GLM-5.3 Flash.
• Context: 1M tokens instead of 200k; text and image input, text output.

The weak Omniscience accuracy is partly explained by the model's behavior. Haiku 5.5 is more willing to admit that it does not know an answer, so it makes things up less often. For agentic scenarios, where false confidence costs more than a pause, this is more of a plus. Artificial Analysis considers its AutomationBench result understated: a bug with safety refusals hit the pre-release test, and the model turned down tasks for no reason. Anthropic promises a fix, and Artificial Analysis will rerun the test.

Price per token and price per task

Haiku 5.5 uses tiered pricing. Above 100k tokens in the prompt, the price rises fivefold, to $0.50/$2.50. Cache reads cost $0.01 per million ($0.05 above the threshold), and a five-minute cache write costs $0.125 ($0.625). The Artificial Analysis site does not yet account for the higher tier, so preliminary cost estimates are understated. Cost per Task is promised later.

The model also has a more noticeable problem: it is verbose. At maximum effort, Haiku 5.5 spends about 162k output tokens per index task. That is three times as many as GPT-6 Luna (about 50k) and more than Opus 5.5 (max). Moving from xhigh to max adds only 2 points for ~1.8x the tokens. At an equal 38 points, Haiku 5.5 (high) consumes ~55k tokens while Luna (max) uses ~50k, and at lower settings the gap grows.

Output rates are the same for both models, so the real bill per task for Haiku 5.5 at max will be noticeably higher than for Luna. In this situation, a low price per million tokens does not guarantee a win on a per-task basis. Only Cost per Task calculations will settle the question.

Who should use it, and for what

Haiku 5.5 suits agentic pipelines, terminal work and office automation, where you need a strong result at a low price and a context window of a million tokens. For factual questions without external search, Gemini 3.

8 Flash is the better choice, with its 55% on Omniscience. Before choosing an effort level, run your own tasks and measure not only quality but also token consumption: at the high level the model fits into ~55k tokens, while at max it uses ~162k. Keep long prompts under 100k or make heavy use of the cache, otherwise the rate rises fivefold.

One more caveat: AutomationBench currently shows an understated result because of excessive refusals, so final conclusions about automation should wait until after the fix.