The split is 100,000 prompt tokens
Claude Haiku 5.5 (claude-haiku-5-5) publishes two list prices, and the switch is the length of that one prompt. On the Claude pricing page, read November 14, 2026, input is $0.10 per million tokens and output is $0.50 per million tokens when the prompt is at most 100,000 tokens. At 100,001 tokens or more, that whole request is billed at $0.50 per million input tokens and $2.50 per million output tokens. Cache reads and cache writes count toward the length. A later short request does not inherit the higher rate, and a long request does not keep the lower one.
The model page dates the release to October 7, 2026. Context is 1 million tokens. Synchronous output tops out at 128,000 tokens. Anthropic's launch note says prompts at or under 100,000 tokens made up around 90% of requests to the previous Haiku. That sentence describes previous Haiku traffic. A repo-sized prompt can land far from that mix. The product page also lists Haiku 5.5 inside Claude Code. This comparison prices the Claude API, not a seat in an IDE.
Try it: platform.claude.com/docs/en/models/haiku-5-5/overview
Four rates on one page
Sonnet 5.5 and Fable 5.2 do not change price at 100,000 tokens. Sonnet 5.5 stays at $2 input and $10 output per million tokens. Fable 5.2 stays at $8 and $40. The 5-minute and 1-hour cache writes, and the cache-read rate, are different rows. The 1-hour cache write is $0.20 per million on short Haiku prompts, $1 over the line, $4 on Sonnet 5.5, and $16 on Fable 5.2.
| USD per 1M tokens | Haiku 5.5, prompt <=100k | Haiku 5.5, prompt >100k | Sonnet 5.5 | Fable 5.2 |
|---|---|---|---|---|
| Input | $0.10 | $0.50 | $2 | $8 |
| Output | $0.50 | $2.50 | $10 | $40 |
| 5-minute cache write | $0.125 | $0.625 | $2.50 | $10 |
| Cache read | $0.01 | $0.05 | $0.10 | $0.20 |
The October 7, 2026 list-price card adds a US-only inference row for Haiku 5.5: $0.11 input and $0.55 output at or under 100,000 tokens, then $0.55 input and $2.75 output. Global batch processing on that card is $0.05 input and $0.25 output per million tokens for prompts at or under 100,000, which is half the base rate. The model overview states the same 50% batch discount. Cursor and GitHub Copilot seats are not on this card.
Three mixes, counted from the table
Mix one: 1,000,000 input tokens and 100,000 output tokens, split so each prompt stays at or under 100,000 tokens. Haiku short row: $0.10 + $0.05 = $0.15. Sonnet 5.5: $2 + $1 = $3. Fable 5.2: $8 + $4 = $12. The Fable bill is 80 times the Haiku bill. The Sonnet bill is 20 times.
Mix two is a single 200,000-token prompt plus 2,000 output tokens. Haiku cannot use $0.10 here. The long row is 200,000 / 1,000,000 x $0.50 = $0.10, plus 2,000 / 1,000,000 x $2.50 = $0.005, so $0.105. Sonnet 5.5 is $0.40 + $0.02 = $0.42. Fable 5.2 is $1.60 + $0.08 = $1.68. Stuffing the repo into one prompt makes Haiku five times more expensive per token on input than the short row, and Sonnet is still four times the long-row Haiku total on this mix.
Mix three: an 80,000-token prefix read from cache 1,000 times, plus 500 new output tokens each time, with the full prompt still at or under 100,000. Cache reads are 80 million tokens. Haiku at $0.01 is $0.80, output is $0.25, total $1.05. Sonnet at $0.10 is $8 of reads plus $5 of output, total $13. Fable at $0.20 is $16 of reads plus $20 of output, total $36. Move that prefix to 120,000 tokens and Haiku's cache read becomes $0.05 per million ($6.00) and the output rate becomes $2.50 ($1.25), total $7.25, because the higher row covers the whole request.
Which call belongs on which row
Choose Haiku 5.5 if: the call is classification, extraction, or routing, and each prompt can stay at or under 100,000 tokens. The short row is the one Anthropic priced at $0.10 and $0.50. Mix one on that row is $0.15. Batch on the October 7 global card cuts the short row to $0.05 and $0.25 if the job can wait.
Choose Sonnet 5.5 if: the prompt has to carry a large checkout of the repo and you are not willing to send that edit to the small model. Mix two only compares list price: $0.42 on Sonnet 5.5 and $0.105 on long-row Haiku. Pay $0.42 when that edit should stay off the small model.
Choose Fable 5.2 if: the call is one of the few hard passes already budgeted at $8 per million input tokens and $40 per million output tokens. Mix one is $12, eighty times the $0.15 short-row Haiku bill. A hot cache still costs $0.20 per million tokens to read. Re-check the pricing page before a large batch: the figures here are the ones listed on November 14, 2026.