Quick answer
Anthropic released Claude Haiku 5.5 on October 7, 2026, the last model of the Claude 5.5 family. Prompts up to 100K tokens cost $0.10 per million input and $0.50 per million output, against $1 and $5 for Haiku 4.5; Anthropic says that tier covers about 90% of requests to the previous Haiku. Prompts over 100K tokens bill at $0.50 and $2.50, and the higher rate applies to the whole request. Cache reads are listed at $0.01. It adds a configurable effort setting, beta computer-use and browser-use in the SDKs, and an Anthropic-reported 72.4% on the OSWorld 2.1 offline subset (Haiku 4.5: 15.7%). Context is 1M tokens with 128K output. Model ID: claude-haiku-5-5, on the Claude Platform and AWS Bedrock.
Haiku is the model most products actually run at scale: the classifier, the summariser, the subagent that does the boring half of the work. A tenth of the price for that tier is a bigger change to most AI bills than any flagship launch this autumn.
What changed
- Price: $0.10 / $0.50 per million tokens for prompts up to 100K, a roughly 90% cut for short prompts
- Effort setting: choose lower or higher reasoning per request, trading tokens for capability
- Computer use and browser use in beta in the Python and TypeScript SDKs
- Context 1M tokens, output 128K, matching Sonnet 5.5 and Opus 5.5
- Anthropic's headline "about 75% cheaper to run" is a different measure from the list-price cut; both are vendor figures
The 100K cliff
Above 100K input tokens the rate jumps five times, to $0.50 and $2.50, and the whole request is billed at the higher rate, output and cache included. That is the same shape as the 272K surcharge on GPT-6.1 Sol and the 200K one on Grok 4.7, with a lower threshold. For most Haiku workloads it will never matter; for anyone feeding whole documents into a cheap model, it decides whether Haiku 5.5 is cheap at all. Measure your prompt sizes before you migrate.
API changes to fix first
- Requests that still send temperature, top_p, or top_k from Haiku 4.5 code return 400 errors; remove them
- Thinking cannot be disabled; use the effort setting instead of a thinking budget
- Cache-write and batch pricing were not in the launch coverage; check the rate card before assuming parity with Sonnet 5.5
Where it fits
The 5.5 family is now complete: Opus 5.5 at $4/$20 for the hardest work, Sonnet 5.5 at $2/$10 as the default, Haiku 5.5 at $0.10/$0.50 for volume. The closest rival is DeepSeek V4.1 Flash at $0.15/$0.60 off-peak with open weights and no long-prompt cliff; the choice between them is mostly about licence, computer use, and data policy rather than price.
Bottom line
If you run Haiku 4.5, the migration is a model-ID change, a parameter cleanup, and a check on prompt sizes, and it will probably cut that line of your bill by most of its value. Just know where the 100K cliff is.
