Claude Haiku 5.5
- ID
- 32928
- Status
- summarized
- Published
- 08 Oct 2026, 4:56 AM
- Fetched
- 08 Oct 2026, 7:22 AM
- Provider
- Simon Willison
- Category
- developer-ai
- Original URL
- https://simonwillison.net/2026/Oct/7/claude-haiku-5-5/
- Source URL
- https://simonwillison.net/atom/everything/
Summary
- Score
- 7.8
- Created
- 08 Oct 2026, 7:23 AM
- Tags
- Audience
- developersai_ml_learnersai_agent_userssaas_founders
What happened
Anthropic released Claude Haiku 5.5 on 7 October 2026 at $0.10/million input and $0.50/million output up to 100,000 tokens, matching OpenAI's GPT-6 Luna, but jumping 5x to $0.50/$2.50 past that threshold (Luna only rises to $0.20/$0.75 at 272,000 tokens). Simon Willison measured a further hidden increase: Haiku 5.5's new tokenizer consumes about 1.25x more tokens on the same prompt than Haiku 4.5, which was priced at $1/$5. Haiku 5.5 cannot disable reasoning and defaults to medium; Willison's pelican test cost 0.0936 cents in 7 seconds at low effort and 3.3826 cents over 5 minutes 9 seconds at max effort. Anthropic also halved Sonnet 5.5 cache-read pricing and added monthly API credits to subscriptions: $100 for Max 5x, $200 for Max 20x, up to $500 pooled for Team.
Why it matters
If your prompts run under 100,000 tokens, Haiku 5.5 is roughly a 10x price cut versus Haiku 4.5 and lands at parity with GPT-6 Luna; above 100k tokens Luna is the cheaper choice, so long-context workloads should route elsewhere. The 1.25x tokenizer inflation means your real cost is roughly 25% higher than a naive per-token comparison suggests — re-run your own token counts before switching a production pipeline. And if you already pay for Max 5x, Max 20x or Team, the new monthly API credits ($100/$200/up to $500 pooled) exactly offset the subscription cost, which changes whether a separate API budget line is still needed.
Discussion angle
Take one real prompt from your app, run it through Willison's Claude Token Counter against both Haiku 4.5 and 5.5, and compare the actual billed cost rather than the headline price — then decide whether the 100k-token cliff or the 1.25x tokenizer is the bigger factor for your workload.