Claude Haiku 4.5: Reported Sonnet 4 Score at One-Third Cost

A small model priced at a third of the tier it matched

Anthropic released Claude Haiku 4.5 on October 15, 2025, priced at $1 per million input tokens and $5 per million output tokens. That is a third of the $3 and $15 rates charged for both Claude Sonnet 4 and Claude Sonnet 4.5, the two models Anthropic compared it against. It carries a 200,000 token context window, up to 64,000 output tokens versus 8,192 for the prior Haiku 3.5, and a February 2025 knowledge cutoff. It is the first Haiku model with an extended-thinking mode. [1][2]

Anthropic's own framing: Haiku 4.5 gives "similar levels of coding performance" to Sonnet 4 "but at one-third the cost and more than twice the speed," and runs four to five times faster than Sonnet 4.5. On SWE-bench Verified, an automated benchmark of real-world GitHub coding tasks, Anthropic reported 73.3% for Haiku 4.5, slightly ahead of the 72.7% Anthropic had reported for Sonnet 4 five months earlier without extended thinking enabled. On agentic coding, an evaluation run by Augment Code and cited in Anthropic's launch materials put Haiku 4.5 at roughly 90% of Sonnet 4.5's score. Anthropic also said Haiku 4.5 surpassed Sonnet 4 on computer-use tasks. [1]

Atlas interpretation: Every comparative figure here is Anthropic's own, run on Anthropic's own chosen tests against Anthropic's own older models; the 90% agentic-coding figure is credited to a third party, Augment Code, but was still selected and published by Anthropic. That does not make the numbers false, but "matches Sonnet 4" and "90% of Sonnet 4.5" describe what Anthropic reported, not a score an outside lab reproduced independently. [1]

Five months from flagship to budget tier

Sonnet 4 launched May 22, 2025 as one of Anthropic's two flagship models that month, reported at 72.7% on SWE-bench Verified and priced at $3 and $15 per million tokens. Haiku 4.5, Anthropic's small and cheapest current-generation model, matched or slightly exceeded that same score five months later at a third of the price and, per Anthropic, more than double the speed. [1][3]

Atlas interpretation: The gap between a lab's frontier tier and its budget tier closing in months rather than years is a cost story more than a capability story: the coding score barely moved, but the price to get it fell by two thirds and the model got faster. Whether that reflects genuine efficiency gains in the smaller model or a company choosing to price down a capability it had already trained is not something the launch materials distinguish. [1]

Independent commentary: efficient, but not the cheapest on the market

Independent developer commentary on the same day corroborated the pricing and the coding-parity claim, and noted that despite the price cut, Haiku 4.5 still cost more than the cheapest small models from other vendors, such as GPT-4o-Nano and Gemini 2.0 Flash Lite. That commentary characterized Anthropic's small-model strategy as continuing to target coding capability specifically, rather than competing on the lowest possible per-token price. [2]

Atlas interpretation: Read against the launch claims, this is the honest caveat missing from Anthropic's own framing: "a third of the price of our other models" and "the cheapest model on the market" are different claims, and only the first one is true. Haiku 4.5 undercut Anthropic's own Sonnet tier substantially while remaining a premium product against the bottom of the wider small-model market. [2]