Claude Haiku 5.5 costs 90% less, until prompts get long

The headline price is real. It just has a fence around it.

Anthropic released Claude Haiku 5.5 on October 7 at $0.10 per million input tokens and $0.50 per million output tokens, for prompts up to 100,000 tokens.

Haiku 4.5 lists at $1 and $5, according to Anthropic's pricing docs.

A tenth of the price.

Cheap below 100,000 tokens, five times the price above

Past 100,000 tokens, Haiku 5.5 costs $0.50 per million input and $2.50 per million output. Anthropic's launch page says the model costs about 75% less to run than Haiku 4.5 on average: 90% cheaper up to 100,000 tokens and 50% cheaper above. The same footnote says Haiku 5.5 has an updated tokenizer and uses slightly more tokens per task.

Simon Willison measured one long prompt at about 1.25 times as many tokens on Haiku 5.5 as on Haiku 4.5.

The rest of the spec sheet, from Anthropic's models overview and launch page:

• The model ID is claude-haiku-5-5, with a 1M token context window and up to 128K output tokens.
• It is the first Haiku with an adjustable effort setting, from Low to Max.
• It is available on the Claude Platform, AWS, Google Cloud and Microsoft Azure, and GitHub Copilot made it generally available the same day.

Anthropic's own benchmark table puts Haiku 5.5 at 72.4% on OSWorld 2.1 (offline subset), against 15.7% for Haiku 4.5 and 48.9% for OpenAI's GPT-6 Luna. Those are Anthropic's numbers. Artificial Analysis, which tests independently, gives Haiku 5.5 at Max effort a score of 43 on its Intelligence Index, against 17 for Haiku 4.5's best result.

The small-model price fight has a floor now

Willison notes that GPT-6 Luna lists the same $0.10 and $0.50 base price. Luna's price rises only once a prompt passes 272,000 tokens, and then to $0.20 and $0.75. Under 100,000 tokens, the two match on price. Above it, Luna is cheaper.

So ignore the 90% and look at your prompt lengths. Short calls (classification, routing, extraction, the sub-agents that do the boring work) get dramatically cheaper. Long documents need the math first.

The quieter change is on Sonnet 5.5. Anthropic is halving its cache reads, from $0.20 to $0.10 per million tokens, and says that cuts the cost of most agentic tasks by about 20%. If your agents reread the same context all day, that is money back without touching your code.

Credits for Max and Team subscribers

Anthropic also began monthly API credits for some claude.ai plans, per its support page. Max 5x gets $100 a month and Max 20x gets $200. Team plans get $20 per Standard seat and $100 per Premium seat, pooled into one monthly balance capped at $500.

The credits expire at the end of each billing cycle. They cover the Messages and Batches APIs, but not interactive Claude Code or Claude on Bedrock, Vertex AI or Foundry. Anthropic says they are rolling out over a few days.

Effort is the new price dial

The effort setting moves cost more than the list price does. Willison's drawing test cost 0.0936 cents at Low effort and 3.3826 cents at Max.

That is about 36 times more for the same prompt.

Start at Low, raise effort only where quality fails, and log what each level costs you. Then pull last month's logs and count the prompts that cross 100,000 tokens. That count decides whether this is a price cut for you.