Haiku 5.5 independent review scores 43 points: far ahead of GPT-6 Luna, but top-tier token consumption exceeds 3x

Beating AI News Flash: Third-party evaluation agency Artificial Analysis has published the test results for Claude Haiku 5.5. The new model scored 43 points on the comprehensive intelligence index, an increase of 26 points over the previous generation's 17 points. It surpasses GPT-6 Luna's 38 points and GLM-5.3 Flash's 42 points, and approaches Kimi K3's 44 points. The index combines 10 tests including coding, professional work, and scientific reasoning.

However, the high score comes with massive token consumption. At the highest reasoning intensity, Haiku 5.5 outputs an average of about 162,000 tokens per evaluation task, 3.2 times that of GPT-6 Luna. The two have the same base API unit price, but Haiku 5.5's estimated cost per task is about $0.21, while Luna's is only $0.07. When Haiku 5.5 is adjusted from the second-highest level to the highest level, the comprehensive score increases by only 2 points, while output increases by about 80%.

After lowering the reasoning intensity, the gap narrows significantly. Haiku 5.5 scores 38 points at the high level, with an average output of about 55,000 tokens per task, close to the highest-level GPT-6 Luna's 38 points and 50,000 tokens. Developers can reduce unnecessary token consumption by adjusting the reasoning intensity.

Haiku 5.5 is also not ahead across the board. In the AutomationBench-AA automation task evaluation, it scored only 35%, trailing Luna's 53%. However, during the test period, the model had an issue of excessively refusing to execute tasks. Anthropic is fixing it, and Artificial Analysis plans to retest. In addition, the currently published task cost for Haiku 5.5 does not yet include the long-context surcharge, so the actual cost may be higher.