Quick Facts

  • Claude Haiku 5.5 drops input token pricing 90% to $0.10 per million for requests under 100,000 tokens, matching GPT-6 Luna's cheapest tier.
  • The model jumps from 15.7% to 72.4% on the OSWorld computer use benchmark and introduces a one-million-token context window, up from 200,000.
  • Haiku 5.5 is available now through the Claude Platform, AWS, Google Cloud, Microsoft Azure, and GitHub Copilot.

Anthropic shipped Claude Haiku 5.5 on Oct. 7, 2026, cutting API costs by as much as 90% for the high-volume, cost-sensitive workloads that make enterprise AI expensive at scale. Input tokens for requests under 100,000 now cost $0.10 per million. Output runs $0.50 per million at the same tier.

The pricing matches OpenAI's GPT-6 Luna at identical list rates. Anthropic says around 90% of requests to the previous Haiku model fell within that lower pricing band, meaning most customers will see the full discount.

Pricing Details

The cuts are tiered. Requests above 100,000 tokens revert to $0.50 per million for input and $2.50 for output. Cache reads fall to $0.01 per million for shorter prompts. Anthropic says the average cost reduction across all request lengths is around 75%, lower than the headline 90% because its updated tokenizer uses more tokens per request. Cache read pricing for Claude Sonnet 5.5 also dropped, from $0.20 to $0.10 per million tokens.

Benchmark Gains

The performance gap between Haiku 5.5 and its predecessor is wide. On OSWorld, a computer use test, the model jumped from 15.7% to 72.4%. On Humanity's Last Exam, it reached 45.9% without tools and 57.4% with tools, up from 10.2% and 18.7%. The GDPval-AA v2.1 knowledge benchmark score more than doubled, from 735 to 1,620.

Agentic coding shows the sharpest change. Haiku 4.5 scored 0.0% on Terminal-Bench 4.0, which tests complex, multi-step command-line tasks. Haiku 5.5 scores 39.2% at maximum effort and approximately 20% at the default medium effort setting. Anthropic's launch table shows Haiku 5.5 beating GPT-6 Luna on every listed benchmark at the same price point.

One caveat: Haiku 5.5 introduces adjustable effort levels. Anthropic uses medium as the default, but the Terminal-Bench figure of roughly 39% reflects maximum effort. Buyers should account for that distinction when evaluating real-world cost and performance tradeoffs.

Expanded Capabilities

Haiku 5.5 arrives with a one-million-token context window, a fivefold increase from the 200,000 tokens Haiku 4.5 shipped with. Anthropic positions the model for narrowly scoped, repetitive tasks — summarization, document classification, database queries, and subagent work — rather than complex agentic coding, where Sonnet 5.5 and Opus 5.5 remain the better options.

GitHub has made Haiku 5.5 generally available in Copilot across Pro, Pro+, Max, Business, and Enterprise plans. GitHub's early tests found the model performed on par with Claude Sonnet 5 across many coding tasks while using fewer tokens and steps.

Early Customer Results

Box VP of AI Products Yashodha Bhavnani reported an 11-point improvement over Haiku 4.5 with approximately half the latency. Asana staff software engineer Aaron Vinh said task-completion latency fell more than 30% and inference per agent turn accelerated by as much as 2.5 times. HubSpot distinguished software engineer Ze'ev Klapow reported a 92.8% average across three runs of its CRM evaluation, the strongest result among the models tested.

Subscriber API Credits

Anthropic also announced monthly API credits for eligible subscribers, available without a separate Console payment card. Max 5x users receive $100 per month. Max 20x users receive $200. Team plans get $20 per Standard seat and $100 per Premium seat, pooled across the organization and capped at $500 per month.

Haiku 5.5 is the third model in the Claude 5.5 family, following Opus 5.5 and Sonnet 5.5, which launched earlier this fall. The API identifier is claude-haiku-5-5.

Read more: Anthropic launches Claude Haiku 5.5 with 90% API price reduction, matching GPT-6 Luna

anthropicapienterprisellmopenai

This article was written by an AI agent. Spotted an error? Send a correction and we will fix it.