Anthropic launches Claude Haiku 5.5: cheaper, but with a token catch

Anthropic has shipped the fast, low-cost model it promised earlier — Claude Haiku 5.5. The previous version, Haiku 4.5, was clearly showing its age: it launched almost a year ago at $1 per million input tokens and $5 per million output. That was pricey even then — ten times the cost of OpenAI's GPT-6 Luna, which arrived a month earlier.

Price: caught up, but not everywhere

The new Haiku exactly matches Luna's price — $0.10 and $0.50 per million tokens — but only up to 100,000 tokens. Beyond that the rate jumps fivefold, to $0.50 and $2.50. Luna's increase kicks in later, at 272,000 tokens, and is far gentler — up to $0.20 and $0.75.

There's also a hidden increase. Haiku 5.5 uses a new, less generous tokenizer. By Simon Willison's measurements with his Claude Token Counter tool, the same long prompt consumes roughly 1.25 times as many tokens as on Haiku 4.5.

The upshot is simple: if your workload fits within 100,000 tokens, Haiku costs the same as Luna and posts higher benchmark scores. Past that threshold, Luna looks like the better deal.

A pelican on a bicycle

Willison, as is his custom, tested the model with his signature benchmark — asking it to generate an SVG of a pelican riding a bicycle. The new Haiku doesn't let you turn off reasoning entirely and defaults to medium effort. A decent bicycle frame came out at every effort level above the lowest. The low-effort pelican cost 0.0936 cents and took 7 seconds, while the max-effort one took 5 minutes 9 seconds to generate yet still cost only 3.3826 cents. For comparison, last year's Haiku 4.5 was frankly terrible at drawing pelicans.

Credits for subscribers

At the same time, Anthropic halved the cost of cache reads for Sonnet 5.5 and introduced monthly API credits for subscribers. Max 5x users get $100 a month, Max 20x users $200, and teams up to $500 pooled across their members. The credits exactly match the cost of the subscription itself, which makes using the API considerably easier. Claiming them is painless: in billing settings you pick the organization that should receive the funds. You can also disable auto-reload — requests simply stop when the balance hits zero, with no nasty bills.

Willison notes that OpenAI still lets you use a Codex subscription for personal API work, which works out better for heavy users. This new credit scheme at least partly closes that gap.

Source: Simon Willison