Claude Haiku 5.5 matches GPT-6 Luna's price, with a catch

Anthropic shipped Claude Haiku 5.5 on October 7. For prompts under 100,000 tokens, it costs exactly what OpenAI’s GPT-6 Luna costs: $0.10 per million input tokens and $0.50 per million output. The previous Haiku charged $1 and $5. That’s a big cut, but read the fine print before you move anything.

Over 100K tokens, Haiku’s price jumps five times, to $0.50 and $2.50. Luna has a step too, at 272,000 tokens, but it’s smaller and ends at $0.20 and $0.75. So long prompts favor Luna. Simon Willison ran his token counter on the same long prompt and found Haiku 5.5 uses a new tokenizer that produces about 1.25x as many tokens as Haiku 4.5. That’s a hidden price increase, and it’s the number to check on your own traffic.

The benchmarks look good, with a caveat. Artificial Analysis, an independent evaluator, gives Haiku 5.5 a 43 on its Intelligence Index at max effort, up 26 points from Haiku 4.5 and above Luna’s 38. But it’s chatty. Haiku used about 162k output tokens per index task at max effort, against roughly 50k for Luna. The 5x price step isn’t in their cost numbers yet. A pre-release safety bug also made the model over-refuse, and Anthropic says a fix is coming.

This is the first Haiku with effort settings and adaptive thinking. You can’t turn reasoning off, and the default is medium. Willison’s pelican test (an SVG of a pelican on a bicycle, the usual sanity check) took 7 seconds and 0.0936 cents at low effort. At max, it took 5 minutes 9 seconds and 3.3826 cents, and the drawing was clearly better. Haiku 4.5’s pelican cost 0.7583 cents and, in his words, sucked.

The subscriber news matters too. Max 5x users get $100 a month in Anthropic API credits, Max 20x get $200, and Team gets up to $500 pooled. Credits don’t roll over. You can also turn off auto-reload, so requests stop when the balance runs out instead of billing you.

My take: if your prompts fit in 100K tokens, Haiku 5.5 is a real option at Luna’s price. Run your own prompts through the new tokenizer before you switch, and set a hard cap. Anthropic pitches it as the cheap worker under Opus or Sonnet 5.5, handling summaries and database queries. Test that split on your workload before you trust it.

Source: Simon Willison’s blog post and the Latent Space AINews roundup, both from October 7-8, 2026.