Anthropic launches Claude Haiku 5.5: about 75% cheaper than Haiku 4.5, OSWorld up from 15.7% to 72.4%

Why it matters

Key points

Anthropic released Claude Haiku 5.5 early on October 8 Beijing time (October 7 in the US), calling it its cheapest, fastest and most capable small model yet. It targets high-volume, cost-sensitive work: summaries, context compaction, database queries, classification, and serving as a subagent for Opus 5.5 and Sonnet 5.5 on coding tasks.

Price is the headline. Anthropic says Haiku 5.5 costs about 75% less to run than Haiku 4.5 on average: 90% cheaper for requests up to 100,000 tokens and 50% cheaper above that, after accounting for a new tokenizer that uses slightly more tokens per task. It is also the first Haiku with an adjustable effort setting, so users can trade cost against intelligence per task.

Anthropic's own numbers show a generational jump over Haiku 4.5: 72.4% vs 15.7% on OSWorld 2.1 (offline subset), 39.2% vs 0% on Terminal-Bench 4.0, 45.9% vs 10.2% on Humanity's Last Exam without tools, and 1620 vs 735 on GDPval-AA v2.1. It also beats OpenAI's GPT-6 Luna on those tests (48.9% OSWorld, 16.4% Terminal-Bench) but stays well below Sonnet 5.5 (83.9% and 70.6%). Anthropic itself says complex agentic coding still belongs on Sonnet and Opus.

Three companion changes shipped alongside: Sonnet 5.5 cache reads are halved from $0.20 to $0.10 per million tokens, which Anthropic says makes most agentic work about 20% cheaper; Max and Team subscribers get a monthly API credit ($100 for Max 5x, $200 for Max 20x, up to $500 pooled for Team); and the Python and TypeScript SDKs add beta support for computer use and browser use. Haiku 5.5 is live on the Claude Platform, AWS, Google Cloud and Azure as claude-haiku-5-5.

Verifiable

Official page: https://www.anthropic.com/claude-haiku-5-5 (dated October 7, 2026). The @claudeai announcement was posted at 18:01 UTC on October 7: https://x.com/claudeai/status/2107894039626277339

All benchmark results are Anthropic's own; methodology is in the Haiku 5.5 system card, and the GPT-6 Luna comparison was chosen and run by Anthropic. The 75% figure is an average weighted by Haiku 4.5's request-length mix, so long-context workloads will see much smaller savings. On safety, Anthropic reports large alignment improvements over Haiku 4.5; its cyber safeguards are stricter than Haiku 4.5's but looser than other recent models', and still block penetration testing.

Our take

What changes here is the division of labor. A small model jumping from 15% to 72% on OSWorld means computer-use and browser agents can, for the first time, run their basic work at a very low unit price. With Anthropic explicitly positioning Haiku as a subagent for Opus and Sonnet, and the SDKs adding computer and browser use at the same time, the "big model plans, small model executes" pattern is turning from an engineering trick into the default cost structure.

The subscription API credit matters too: it pushes Claude subscribers straight into building their own agents, competing for developers' default choice with price and credits, head-on against OpenAI's GPT-6 Luna at the low end. Keep in mind that self-reported scores and averaged savings flatter the vendor; your real bill depends on request length and token use, so benchmark cost on your own tasks before switching.