Claude Haiku 5.5 costs the same as GPT-6 Luna, up to 100,000 tokens
Anthropic's cheapest model is now priced exactly like OpenAI's budget model, but only for short prompts. Past 100,000 tokens, the bill jumps fivefold.
Anthropic released Claude Haiku 5.5 on October 7. It closes the 5.5 family. Opus came first, then Sonnet. Anthropic calls it the cheapest, fastest and most capable small model it has shipped. Fine. Let’s check.
The headline is the price. Up to 100,000 tokens, Haiku 5.5 costs $0.10 per million input tokens and $0.50 per million output. Haiku 4.5 charged $1 and $5. OpenAI’s GPT-6 Luna charges $0.10 and $0.50 too. On short prompts, they’re the same price now.
The one number that changes the bill
The part that matters is the 100,000-token line. Under it, the cheap rate applies. Over it, the whole request costs five times as much: $0.50 input and $2.50 output.
Anthropic says about 90% of its previous Haiku requests fell under 100,000 tokens. So most people get the low rate. The ones who don’t are sending long documents, big codebases or long agent histories. For them, Luna looks cheaper, because Luna doesn’t change its rate until 272,000 tokens.
Anthropic also says Haiku 5.5 uses a new tokenizer that produces slightly more tokens for the same text. Simon Willison measured about 1.25 times as many tokens on one long prompt compared with Haiku 4.5. So the real saving is smaller than the list price suggests. Anthropic’s own figure is around 75% cheaper on average, and it doesn’t say how that number was calculated.
What it can do now
Anthropic’s own benchmarks show a large jump from Haiku 4.5. On OSWorld 2.1, a computer-use test on an offline subset, Haiku 5.5 scores 72.4%, up from 15.7%. On Terminal-Bench 4.0, an agentic coding test, it scores 39.2%, against 0.0% for Haiku 4.5.
These are the company’s numbers, and I haven’t checked them independently. Terminal-Bench is also measured at the highest reasoning setting. Simon noted that Haiku 5.5 can’t turn reasoning off, and it defaults to medium. A write-up I read says the default setting scores much lower on that test. I didn’t verify that figure, so treat the 39.2% as a best case.
Haiku 5.5 also brings a new effort setting, so you can choose between speed and quality per request. That’s useful if you run both quick classification jobs and harder ones on the same account.
Sonnet got cheaper too
Anthropic cut Sonnet 5.5’s cache-read price in half, from $0.20 to $0.10 per million tokens. Anthropic estimates that saves around 20% on most agentic work, where repeated context is a large share of the tokens. If you run Sonnet with a long system prompt that stays the same across calls, this change will reach your bill before anything else does. Our Sonnet 5.5 post covers that model in more detail.
My take
The price match with Luna is a sign that Anthropic is no longer willing to sit ten times above OpenAI on its budget tier, and that’s good for anyone who sends a lot of short requests, which is most people building small tools on top of these APIs.
Still, a price that changes by five times at one fixed line is the kind of thing I’d want to see on paper.
So count it.
The 100,000-token cliff is the detail I’d check before switching anything. Count the tokens in your own largest prompt, then compare the bill with the new tokenizer, not the old one. A price that changes by a factor of five at a fixed line is easy to miss in a monthly invoice.