claude-haiku-4-5 — pricing, context window and limits

The previous Haiku, still served

Haiku 5.5 has replaced this model as the current Haiku, at a lower per-token price. Haiku 4.5 is still served, and it is still the only current model with the smaller context window and output cap the table above shows. What it keeps over its successor is predictability: one price at any prompt length, and a request surface that still accepts prefill, sampling parameters and fixed thinking budgets — all of which Haiku 5.5 rejects. See migrating: Haiku 4.5 to Haiku 5.5 before you move.

No effort parameter — the one to check for in shared code

Unlike every other current model in the main lineup, the effort parameter errors outright here rather than being accepted or ignored. Request-building code shared across a model fleet needs to branch on this specifically. See the effort parameter errors on Haiku 4.5 for exactly this failure. Thinking, where you want it, still uses the older fixed-budget form.

A cache minimum that catches short prompts

This model's minimum cacheable prefix is among the highest in the lineup, so a short prompt that caches cleanly on the newest models can silently fail to cache here, with no error to flag it. The combination of a low per-token rate and the highest bar to clear before caching applies is the detail most likely to make this model underperform its price on paper.

Watch its lifecycle row

Among current models, this is the one whose earliest possible retirement date comes soonest, as the lifecycle table above shows. That floor is not a retirement — Anthropic announces one with notice — but a high-volume pipeline pinned to this id should know where its fallback would go.

Related

See Claude Haiku 5.5 for its successor, Sonnet 5.5 vs Haiku 4.5 for what the step up buys, and Claude model retirement dates for the full calendar.