Claude Haiku 5.5 released: surpasses GPT-6 Luna in every benchmark, priced to match and lower than DeepSeek-V4.1 Flash

Anthropic released Claude Haiku 5.5, whose official benchmarks beat GPT-6 Luna on every item, and surpassed DeepSeek-V4.1 Flash and GLM-5.3-Flash to become the new small-model gatekeeper. It is the first Haiku to support five effort tiers. On OSWorld 2.1, the Max tier reaches 72.4% accuracy at a per-run cost of $0.61, while Haiku 4.5 reached only 15.7% at $1.45. Prices for input and output drop 90% for requests under 100,000 tokens, but the new tokenizer uses more tokens for the same task. Terminal-Bench 4.0 scores only 39.2%, far below Sonnet 5.5's 70.6%.