Haiku 5.5's 100K-token cliff cuts its price advantage over Sonnet 5.5 from 20x to 4x

Each request is priced on its own, and that is how the cliff gets reached. Cache reads and cache writes are re-priced along with everything else once the prompt passes the line ^[2]. An agent loop that keeps pushing conversation history back into the prompt crosses it quietly, according to the post ^[23]. The step is abrupt: a prompt of exactly 100,000 tokens costs $0.010 in input at $0.10 per million, and one token more costs about $0.05 ^[15]. The post's worked case is a 120,000-token prompt, which costs $0.06 where the cheap rate would have given $0.012 ^[3].

The post says Haiku is 20x under Sonnet and 40x under Opus on every token class ^[4]. Its opening paragraph says a tenth of Sonnet, but it also calls Sonnet a $2 model ^[5], and 20x is the multiple that fits that figure. That puts Sonnet at $2.00 input and $10.00 output ^[13]. Haiku is the only model in the lineup with two price tiers ^[6], so Sonnet's price does not move. Past the line, Haiku's $0.50/$2.50 is still 4x under Sonnet's $2/$10 ^[14]. Crossing it costs money but does not make Haiku dearer than Sonnet. The Sonnet prices here are our derivation from the post's multiples.

Whether Haiku can do the job is a separate question. Anthropic, as the post reports it, points multi-step shell work, large refactors and debugging across a codebase at Sonnet 5.5 and Opus 5.5 ^[21]. We'd expect the cleanest Haiku wins in a subagent that starts from a short prompt and returns a short result, because it stays far from the line. The benchmark figures from Anthropic are vendor-reported and were not independently reproduced, the post says ^[22].

Bito's sessions show how the saving erodes with length. Against Opus 5.5, Haiku was about 23x cheaper per one-shot session and about 4x cheaper per long build ^[16][17]. The post says the gap widens on long builds. In dollars it does: the saving per session goes from about $0.31 to $4.07 ^[18]. In ratio it narrows. Bito ran fresh sessions on its own task battery ^[7], so the numbers transfer only if your work looks like that battery. Both comparisons are against Opus, not the Sonnet the post recommends as the default coder. The post does not say whether any Haiku session crossed 100,000 tokens.

The tokenizer moves the line. At the 1.25x Willison measured, a prompt of 80,000 tokens under Haiku 4.5's tokenizer comes to 100,000 under Haiku 5.5's; at a flat 30%, the break-even is about 76,900 ^[19].