Key Highlights

  • Claude Haiku 5.5 costs about 75% less than Haiku 4.5 on average, with prices down up to 90% for shorter prompts.
  • It scores 72.4% on the OSWorld 2.1 computer use benchmark, up from 15.7% for Haiku 4.5, Anthropic reported.
  • Haiku 5.5 leads OpenAI's GPT-6 Luna in every category tested, including agentic coding at 39.2% against 16.4%.
  • Anthropic is also cutting Sonnet 5.5 cache read costs by 50%, to USD 0.10 per million tokens.

Pricing and Positioning

Anthropic has released Claude Haiku 5.5, which it calls its fastest and most affordable small model. The model is built for high-volume, cost-sensitive work such as summarization, database queries, classification and live customer support. On average it costs about 75% less than Haiku 4.5. For requests with prompts up to 100,000 tokens, which Anthropic says make up roughly 90% of all previous Haiku requests, prices fall by up to 90%. Prompts longer than 100,000 tokens cost five times as much.

Per million tokens, Haiku 5.5 charges USD 0.10 for input and USD 0.50 for output on prompts up to 100,000 tokens, against USD 1.00 and USD 5.00 for Haiku 4.5. Above 100,000 tokens the rates are USD 0.50 and USD 2.50. Cache reads cost USD 0.01 and cache writes USD 0.125 on shorter prompts. Sonnet 5.5 is priced at USD 2.00 for input and USD 10.00 for output.

Anthropic notes that Haiku 5.5 uses an updated tokenizer that consumes slightly more tokens per task than its predecessor, so real-world savings can run below what the per-token prices suggest. A similar tokenizer change on the Opus 4.x models raised token usage by about 30% on its own.

Benchmark Gains

On the knowledge benchmark GDPval-AA v2.1, Haiku 5.5 scores 1,620, more than double the 735 posted by Haiku 4.5. On Humanity's Last Exam it reaches 45.9% without tools and 57.4% with tools, up from 10.2% and 18.7%.

The largest jump is in computer use, where a model operates a computer on its own. Haiku 5.5 scores 72.4% on an offline subset of OSWorld 2.1, up from 15.7%. Computer use consumes large numbers of tokens, which makes a cheap, fast model a natural fit. On the agentic coding benchmark Terminal-Bench 4.0 it reaches 39.2%, while Haiku 4.5 scored zero.

Anthropic also lists OpenAI's budget model GPT-6 Luna as a comparison, and Haiku 5.5 leads in every tested category. It scores 1,620 against 1,437 on GDPval-AA v2.1, 72.4% against 48.9% on OSWorld, and 46.4% against 42.4% on FrontierCode 1.1. Sonnet 5.5 still holds a wide lead, with 83.9% on OSWorld and 70.6% on Terminal-Bench 4.0.

Reasoning Levels, Safeguards and Sonnet Changes

Haiku 5.5 is the first Haiku-class model with adjustable reasoning levels, so users can balance cost against quality. Anthropic says it works best for narrowly scoped tasks such as compaction, summarization or sub-agent work, while Sonnet 5.5 and Opus 5.5 remain the better picks for complex agentic coding. Cybersecurity safeguards are tighter than on its predecessor yet allow a broader range of defensive tasks than Sonnet 5.5, partly because the model is less capable overall. Penetration testing stays blocked, and organizations with broader needs can apply for Anthropic's verification programs for life sciences and cybersecurity. The model is available on Amazon Web Services, Google Cloud and Microsoft Azure.

Alongside the launch, Anthropic is cutting Sonnet 5.5 cache read costs by 50%, from USD 0.20 to USD 0.10 per million tokens, which it says lowers costs for most agentic tasks by about 20%. The cuts arrive as OpenAI rolls out its GPT-6.1 series. Anthropic is also introducing monthly API credits: USD 100 for Max-5x subscribers, USD 200 for Max-20x subscribers and up to USD 500 for Team subscribers.

Pricing Signals to Follow

The details to track are how the tokenizer change affects effective costs, how Haiku 5.5 performs in production against GPT-6 Luna, and what OpenAI's GPT-6.1 series carries for pricing.