Anthropic released Claude Haiku 5.5 on Oct. 7 and described it as the company’s cheapest, fastest and most capable small model to date. The company said average execution costs are roughly 75% lower than Haiku 4.5.
Built for high-volume work where cost matters
Anthropic said Haiku 5.5 is aimed at large-scale, cost-sensitive workloads including summarization, context compression, database queries and classification. It also said the model can serve as a sub-agent when Opus 5.5 or Sonnet 5.5 is handling programming tasks.
On speed, Anthropic said Haiku 5.5 is its fastest model at present and recommended it for latency-sensitive uses such as real-time customer support and browser operation.
This is also the first Haiku model with an adjustable effort setting, allowing users to trade off cost against capability.
Tiered pricing with lower rates under 100,000 tokens
Haiku 5.5 uses tiered pricing. For prompts under 100,000 tokens, input costs $0.1 per million tokens, output costs $0.5 per million tokens, and cache reads cost $0.01 per million tokens. Anthropic said all three are one-tenth the price of Haiku 4.5.
For prompts above 100,000 tokens, input pricing rises to $0.5 per million tokens and output pricing rises to $2.5 per million tokens.
Anthropic added that about 90% of requests sent to the previous Haiku model were within 100,000 tokens.
Benchmarks show gains in OSWorld 2.1 and Terminal-Bench 4.0
According to results released by Anthropic, Haiku 5.5 scored 72.4% on the computer-use benchmark OSWorld 2.1, up from 15.7% for Haiku 4.5 and above OpenAI’s GPT-6 Luna at 48.9%.
On the coding-agent benchmark Terminal-Bench 4.0, Haiku 5.5 rose from 0% for Haiku 4.5 to 39.2%. That was higher than GPT-6 Luna at 16.4%, but still behind Sonnet 5.5 at 70.6%.
Anthropic said Sonnet 5.5 and Opus 5.5 remain better suited for complex coding-agent work, while Haiku 5.5 fits narrowly defined tasks that had previously been too expensive to run with Claude.
Security rules are tighter than Haiku 4.5 but looser than some recent models
On security, Anthropic said Haiku 5.5 is stricter than Haiku 4.5 but more permissive than some of its more recent models. The company said it allows more defensive cybersecurity work than Sonnet 5.5, while still blocking techniques such as penetration testing that are more likely to be used by attackers.
Organizations that need broader access can apply to Anthropic’s expanded cybersecurity evaluation program, which the company widened earlier this week.
Sonnet 5.5 cache reads get cheaper, and API credits are coming to Max and Team plans
Anthropic also reduced the cache read price for Sonnet 5.5, which went live in late September. The rate was cut from $0.2 per million tokens to $0.1 per million tokens.
Because cache reads account for a large share of model token usage, Anthropic estimated that costs for most agent tasks will fall by about 20% as a result.
Max and Team subscribers will also begin receiving monthly API credits this week for use with any model on Claude Platform. Anthropic said Max 5x users will get $100 a month, Max 20x users will get $200 a month, and Team plans will receive up to $500 a month to be shared across team members.
Anthropic also updated its Python and TypeScript SDKs with beta support for computer use and browser use. Haiku 5.5 is now available on Claude Platform, AWS, Google Cloud and Microsoft Azure.