Anthropic Releases Claude Haiku 5.5: A Small Model With 1M Context Priced at $0.10 per Million Input Tokens

Anthropic released Claude Haiku 5.5, which it calls its cheapest and fastest small model yet. It keeps a 1 million token context window, up to 128,000 output tokens, and costs $0.10 per million input tokens and $0.50 per million output tokens for prompts up to 100,000 tokens. Above that, rates rise.

Anthropic estimates it runs about 75% cheaper on average than Haiku 4.5 after accounting for a new tokenizer that counts the same text as roughly 30% more tokens. It is generally available on Anthropic's API, Amazon Bedrock, Google Cloud and Microsoft Foundry. And it matches OpenAI's GPT-6 Luna on short-prompt list price, which tells you where the small-model price war has landed.

We think this is good news for useful AI, mostly. Small, cheap models are the workhorses. They summarize documents, sort tickets, pull a number out of a filing and hand it to a bigger model. When that layer gets dramatically cheaper and better, a lot of tedious work becomes economical to automate, and smaller companies get access to capabilities that used to need a serious budget.

The reported jump in capability is large. Anthropic's own numbers put Haiku 5.5 at 72.4% on an offline subset of OSWorld 2.1 for computer use, against 15.7% for Haiku 4.5. That is a different class of model, if it holds up. Anthropic is also honest about positioning: it still recommends Sonnet 5.5 and Opus 5.5 for complex agentic coding, and pitches Haiku as the subagent working under them.

Now the fine print, because it matters. The headline price applies under 100,000 tokens. Above that, the million-token window costs five times as much per token, and GPT-6 Luna's higher tier kicks in much later, so for long prompts Luna is cheaper on list price. The new tokenizer quietly raises token counts, and non-default sampling settings now return errors. Teams migrating should run their own numbers, not the headline.

The bigger downside is what near-free intelligence does at scale. When a computer-use-capable model costs this little, it becomes cheap to run thousands of agents clicking through websites, filling forms and generating content. That powers great customer support and automation. It also powers spam, scraping and fraud at volumes we have not seen. I'm watching whether usage controls keep pace with price cuts.

Our read: a strong release that makes the agent stack cheaper from the bottom up. All benchmark figures are Anthropic-reported, so independent testing is the next thing to wait for.

Haiku 5.5 is the first Haiku-class model with an adjustable effort setting, with adaptive thinking on by default. It accepts text and images and has a June 2026 knowledge cutoff. Anthropic cited early users including Rogo, which uses it as a subagent pulling data from filings, and AlphaSense, which tested it on a feature handling about 8 million calls a week.