The benchmark gains are large, though the figures come from Anthropic's own reporting. Haiku 5.5 scores 1,620 on the knowledge benchmark GDPval-AA v2.1, more than double Haiku 4.5's 735, and jumps from 15.7 to 72.4 percent on the computer-use benchmark OSWorld 2.1, where both scores cover the offline subset. On agentic coding test Terminal-Bench 4.0 it reaches 39.2 percent where its predecessor scored zero. Anthropic's comparison table also puts it ahead of OpenAI's budget model GPT-6 Luna in every tested category, while Sonnet 5.5 reference scores show the small model still trails Anthropic's larger models by a wide margin.

There is a catch on the savings. Haiku 5.5 uses an updated tokenizer that consumes slightly more tokens per task, and the same change pushed token usage up about 30 percent on the Opus 4.x models, so real-world savings are likely smaller than the per-token prices suggest.

It is the first Haiku-class model with adjustable reasoning levels, letting users trade cost against quality, and Anthropic says it suits narrowly scoped tasks such as summarisation or sub-agent work rather than complex agentic coding. Cybersecurity safeguards are tighter than on the predecessor, though penetration testing stays blocked.

Alongside the launch, Anthropic is halving Sonnet 5.5 cache read costs from $0.20 to $0.10 per million tokens, which it says should cut costs for most agentic tasks by about 20 percent, and is rolling out monthly API credits for subscribers. The Decoder's view is that the move is almost certainly a reaction to OpenAI's new GPT-6.1 series, a sign the AI price war is intensifying.