Anthropic Just Put Claude on the Dollar Menu. Read the Fine Print.

Ten cents. That's what Anthropic now charges for a million input tokens on Claude Haiku 5.5, the small model it launched on Wednesday. Output runs 50 cents per million. Haiku 4.5 cost $1 and $5. That's 90% off, and Silicon Valley's AI price war just got a dollar menu.

Then you read the fine print.

The 90% discount only covers requests up to 100,000 tokens. Go past that and the price jumps five times, to 50 cents in and $2.50 out. Still half off Haiku 4.5. Just not the dollar menu anymore. Over on Hacker News, one developer put it bluntly: "100k tokens is an absurdly low cutoff."

There's a second catch in the docs. Haiku 5.5 ships with a new tokenizer, and Anthropic says the same text now counts as about 30% more tokens. Simon Willison measured roughly 1.25 times as many in his own tests and called it "a hidden price increase." Anthropic's own math puts Haiku 5.5 at about 75% cheaper than Haiku 4.5 on average. A great deal, just a smaller one than the headline.

The upgrade itself is real, at least on Anthropic's numbers. On OSWorld 2.1, a computer-use benchmark, Haiku went from 15.7% to 72.4%. On Terminal-Bench 4.0, Haiku 4.5 scored a flat 0.0%. Haiku 5.5 scores 39.2%. It's also the first Haiku with an effort setting, and it gets a 1-million-token context window with up to 128K tokens of output.

The target is obvious: OpenAI, Anthropic's neighbor across town in San Francisco, and its GPT-6 Luna, which sells for the same price. Anthropic says Haiku beats Luna on all six benchmarks it compared them on, including OSWorld, 72.4% to 48.9%. One post on X put it more simply: "Claude Haiku 5.5 cooked GPT-6 Luna."

Independent testing is messier. Artificial Analysis gives Haiku 5.5 a 41 on its Intelligence Index, ahead of Luna's 35, with both models on their Xhigh setting. But Haiku burned through about 180 million output tokens to get there, versus 69 million for Luna, so running the index cost $158 on Haiku and $67 on Luna. Cheap per token. Very fond of tokens. And on Artificial Analysis's business-workflow test, AutomationBench, Luna wins 48% to 36%.

Luna is also more forgiving on long requests. Its price doesn't step up until 272,000 tokens, and then only to 20 cents in and 75 cents out. Willison's verdict: "Above 100,000 tokens, Luna looks like a much better deal."

Anthropic threw in a few extras. Sonnet 5.5 cache reads are now half price, down from 20 cents to 10 cents per million tokens. Max and Team subscribers now get monthly API credits: $100 on Max 5x, $200 on Max 20x, and up to $500 pooled for Team. Willison notes the credits don't roll over.

Why so generous? Haiku 5.5 is Anthropic's third Claude 5.5 model in about a month, and reports say the company could go public as soon as mid-November. Nothing gets Wall Street's attention like a price war you look like you're winning.

So who should switch? If your requests stay under 100,000 tokens, Haiku 5.5 is a ridiculous deal. If they regularly run longer, redo the math with the new tokenizer before you move anything, because Luna may still come out cheaper.

Read the fine print. Then count your tokens. Is ten cents enough to pull you over to Claude, or is Luna staying on your bill?