Anthropic launches Claude Haiku 5.5, its fastest and cheapest small model

Haiku 5.5 cuts token prices for high-volume work. The 100,000-token threshold and Sonnet's cheaper cache reads change the model-choice calculation.

Oct 8, 2026

Anthropic launched Claude Haiku 5.5 on October 7 for high-volume tasks such as summaries, classification and coding subagents. It calls Haiku its fastest model at standard speed, though Opus in Fast Mode runs faster. Developers can use claude-haiku-5-5 on the Claude Platform, with availability also on AWS, Google Cloud and Microsoft Azure.

The prompt length changes the price

Anthropic publishes these rates in dollars per million tokens:

Token type	Haiku 5.5, prompts up to 100,000 tokens	Haiku 5.5, prompts over 100,000 tokens	Haiku 4.5	Sonnet 5.5
Input	$0.10	$0.50	$1.00	$2.00
Output	$0.50	$2.50	$5.00	$10.00
Cache reads	$0.01	$0.05	$0.10	$0.10
Cache writes	$0.125	$0.625	$1.25	$2.50

Sonnet’s cache-read rate falls from $0.20 to $0.10. Anthropic estimates average running costs fall about 75% for Haiku versus 4.5 and about 20% for Sonnet’s agentic work. Haiku’s estimate accounts for its changed tokenizer and token usage per task.

Coding results favor a split workload

Haiku gains an adjustable effort setting. In Anthropic’s Terminal-Bench 4.0 comparison, Haiku scores 39.2%, GPT-6 Luna 16.4% and Sonnet 70.6%. On OSWorld 2.1’s offline subset, they score 72.4%, 48.9% and 83.9%, respectively. These are Anthropic’s reported evaluations. Its announcement supplies no rival token-price table.

Our analysis

Price the work you can delegate before switching the whole agent

Using the published rates, an illustrative batch with 10 million uncached input tokens and one million output tokens costs $1.50 on Haiku’s lower tier, $7.50 on its higher tier, or $30 on Sonnet. This calculation assumes equal token counts and excludes cache writes, tool charges and retries. The batch’s total size does not determine Haiku’s tier. Each request’s prompt length does, so builders should check the share of requests above 100,000 tokens before budgeting a migration.

The table also narrows the savings for cached work. Haiku’s lower-tier input and output rates are one twentieth of Sonnet’s, while cache reads cost one tenth as much. Above the threshold, Haiku’s cache reads cost half as much as Sonnet’s. A workflow that repeatedly reads cached context therefore has a different savings ratio from one that generates fresh answers.

Start a trial with a bounded task such as summarization or classification, and compare accepted results at the chosen effort setting. Record token spend and repeat attempts together. The coding-score gap supports keeping Sonnet on complex coding work while testing Haiku on smaller delegated tasks. For a rival comparison, the announcement supports an initial benchmark shortlist, but a purchase decision still needs rival prices and results on the same workload.

Sources

Anthropic: Introducing Claude Haiku 5.5. Source of pricing, reported evaluations and the lead launch graphic, credited to Anthropic.