Claude Haiku 5.5: A 90% Price Cut With a Tokenizer Catch

Anthropic just cut the price of its smallest model by an order of magnitude, and the headline number still overstates the savings. Claude Haiku 5.5 is the cheapest capable Claude ever shipped. What a team actually saves depends on two details most launch coverage skipped: a new tokenizer and a reasoning mode that cannot be switched off.

The Number

Haiku 5.5 costs $0.10 per million input tokens and $0.50 per million output tokens for requests up to 100,000 tokens, according to The Decoder’s breakdown of Anthropic’s launch. Above that threshold the rate rises fivefold to $0.50 and $2.50. Haiku 4.5 charged a flat $1 and $5 at any length, so short requests get a 90% cut and long ones get 50%.

Anthropic told The New Stack that about 90% of Haiku 4.5 requests fell under 100,000 tokens. That makes the cheap tier the one that matters for most existing workloads. The context window also grows from 200,000 tokens to one million, and the model ships on the Claude Platform, AWS, Google Cloud, and Azure.

How the Math Works

Anthropic’s own blended estimate is roughly 75% savings, not 90%. The gap comes from the tokenizer. Simon Willison measured the same long prompt at about 1.25 times as many tokens on Haiku 5.5 as on Haiku 4.5, so part of the per-token cut is spent before the model writes a word.

The second variable is effort. Haiku 5.5 is the first Haiku with adjustable reasoning levels, defaulting to medium, and Willison found reasoning cannot be disabled. His test drawing cost 0.0936 cents at low effort and took 7 seconds. At max effort it cost 3.3826 cents and took over five minutes. Artificial Analysis, cited by The Decoder, found Haiku 5.5 at max effort used about 162,000 output tokens per task against roughly 50,000 for GPT-6 Luna. At about $0.21 per task, it cost roughly three times as much as Luna for a similar intelligence score.

So the price card says Haiku 5.5 equals Luna at $0.10 and $0.50. The bill says it depends on which effort level the workload actually needs.

What It Signals

The capability jump is real. On Anthropic’s reported figures, Haiku 5.5 scores 72.4% on the OSWorld 2.1 offline subset against 15.7% for Haiku 4.5 and 48.9% for GPT-6 Luna. On Terminal-Bench 4.0 it scores 39.2%, where Haiku 4.5 scored zero. Artificial Analysis also measured a 40% hallucination rate against 77% for Luna. Sonnet 5.5 still leads at 70.6% on Terminal-Bench, which is why Anthropic positions Haiku for compaction, summarization, and sub-agent work rather than complex agentic coding.

Cheaper options still exist below this tier. The New Stack notes Alibaba’s Qwen3.7 Flash at $0.03 input and $0.13 output for short prompts. Haiku 5.5 does not win on raw price. It competes on capability per dollar inside an ecosystem many teams already run.

The Pattern

This is the fourth major model price reset Stack Archive has covered in three weeks. It follows the simultaneous Opus 5.5 and GPT-6 Sol/Luna launch, Grok 4.7’s aggressive pricing, and GPT-6.1 Sol’s near-Astra performance at a fraction of the cost. The recurring angle is that list prices fall while effective cost per task moves far less, because tokenizers and reasoning budgets absorb the difference.

Source: The Decoder

Content created with AI assistance and reviewed for accuracy.