Claude Haiku 5.5: prices up to 90% lower and adjustable effort levels

On October 7, 2026, Anthropic released Claude Haiku 5.5 (API identifier claude-haiku-5-5), the third model in the 5.5 family after Opus 5.5 and Sonnet 5.5. Built for high-volume, latency-sensitive tasks, it brings a structural change: pricing that depends on prompt length. Up to 100,000 tokens, it costs $0.10 per million input tokens and $0.50 per million output tokens; above that threshold, the rates rise to $0.50 and $2.50 respectively. For comparison, the previous Haiku 4.5 cost a flat $1 for input and $5 for output. Caching follows the same tiered logic: cache reads cost $0.01 per million tokens up to 100,000 tokens ($0.05 above), versus $0.10 for Haiku 4.5; cache writes drop to $0.125 ($0.625 above), compared with $1.25 for the previous generation. At the same time, Anthropic halved the cache read price for Sonnet 5.5, bringing it to $0.10 per million tokens.

On performance, the company calls the newcomer its cheapest, fastest and most capable small model. In benchmarks whose results are reported by Anthropic, Haiku 5.5 shows clear gains over version 4.5: on OSWorld 2.1, for example, it goes from 15.7% to 72.4%, and on Humanity's Last Exam it reaches 45.9% without tools. Sonnet 5.5 still scores higher on every reported metric. And while Haiku 5.5 is described as the fastest model at standard speed, it is slower than the Opus models when those run in Fast mode. It is worth stressing that these evaluations come from the manufacturer and have not yet been independently replicated. The official documentation page also does not state the context window (some aggregators cite one million tokens, an unverified figure) or the ASL level.

Another technical novelty is adjustable effort controls with five levels (Low, Med, High, Xhigh, Max), designed to let developers balance compute and cost on their own. On safety, the model applies cyber-risk safeguards that are stricter than Haiku 4.5's but more permissive than Sonnet 5.5's: defensive work is allowed, while attempts at penetration testing or other offensive techniques are blocked. In the biological domain, containment measures remain aligned with those of Sonnet 5, Sonnet 5.5 and Opus 5. For organizations with specific needs in these areas, Anthropic offers controlled access through dedicated programs such as the Life Sciences Verification Program and the Cyber Verification Program.

According to Anthropic, the average saving is about 75% (90% less up to 100,000 tokens, 50% above). As reported by Let's Data Science, the company says the under-100,000-token tier covers about 90% of the requests made to the previous Haiku. Real savings will therefore depend on how traffic splits between the two tiers. The price cut and the effort levels shift competition among budget models, where Haiku 5.5 goes up against OpenAI's GPT-6 Luna. — Olya