Haiku 4.5 in Sume's registry now points to Haiku 5.5: $1 to $0.10

Short answer

In Sume's agent model registry, the Haiku 4.5 row is disabled and names Haiku 5.5 as its successor with when: "always". New requests that name a Haiku 4.5 spelling resolve to the Haiku 5.5 row; rows already stored keep the exact 4.5 id and the 4.5 rate card. On list prices, the move takes input from $1 to $0.10 and output from $5 to $0.50 per million tokens, for prompts up to 100,000 tokens.

This is separate from Anthropic's own lifecycle. Anthropic's pricing page, read 2026-10-08, still lists Claude Haiku 4.5 at $1 and $5, and the models overview lists it under legacy models that are still available.

Side by side

Prices below are per million tokens from the Anthropic pricing page. Batch and over-100k tiers are shown because the 5.5 step changes the picture for long prompts.

Arithmetic for a long prompt

Over 100,000 tokens, Haiku 5.5 input is $0.50, still half of Haiku 4.5's $1. A 150,000-token prompt with 3,000 output tokens costs 150,000 x 0.50 / 1M = $0.075 plus 3,000 x 2.50 / 1M = $0.0075, so $0.0825 on 5.5; on 4.5 it is $0.15 plus $0.015 = $0.165. Under 100,000 tokens the gap is a factor of ten.

Output follows the same pattern: 5.5's over-100k output is $2.50, half of 4.5's $5.

What to check

If you pinned claude-haiku-4.5 in a Format run and compared costs last week, rerun the comparison. The receipt names the id that ran, and the Anthropic page says Haiku 5.5 is not retired sooner than 7 October 2027. Note that the tokenizer is newer on 5.5: Anthropic says models from 4.7 on produce about 30 percent more tokens for the same text, so a per-token price cut of 10x does not mean an equal cut per document.

Migration notes

Anthropic's models overview gives Haiku 5.5 a retirement date of not sooner than October 7, 2027. That is the vendor's commitment on Anthropic-operated platforms. Sume's picker has no reason to keep a 4.5 row listed, so it points users onward and keeps the old row only for stored threads.

Two differences are worth testing in your own prompts: the Haiku 5.5 tool-use system prompt is shorter (286 tokens against 496 for auto tool choice), and the tokenizer is the newer one, which Anthropic says produces about 30 percent more tokens for the same text on 4.7-and-later models. The first lowers cost and the second raises token counts, so measure a real prompt.

For most users nothing needs doing. Threads that were started on the old row keep working because the registry maps the old id onto the new one, and a Format that names the old id in model resolves the same way. If you pinned Haiku 4.5 for a reason, such as an output style you tuned prompts against, run a small side-by-side before you accept the swap, and keep the output of the old runs so you can compare. The price change is the part you can compute in advance; the quality change is the part you have to test.

• Tool-use tokens, auto: 286 on 5.5, 496 on 4.5.
• Tool-use tokens, any/tool: 406 on 5.5, 588 on 4.5.