Anthropic Releases Haiku 5.5 With 90% Lower Token Prices for Shorter Prompts

Anthropic released Claude Haiku 5.5 on October 7, 2026, lowering the price of repeated AI work such as summaries, classification and smaller assignments inside larger projects. Its API rates start at $0.10 per million input tokens and $0.50 per million output tokens—a 90% cut from Haiku 4.5 for prompts up to 100,000 tokens.

Cheaper tokens, a different workload bill

The discount shrinks for longer prompts. Above 100,000 tokens, Haiku 5.5 charges $0.50 per million input tokens and $2.50 per million output tokens, half Haiku 4.5’s rates. Anthropic says roughly 90% of requests to the previous model fell into the shorter category.

Those rate cuts are not the same as savings on completed work. Anthropic estimates that Haiku 5.5 costs around 75% less to run on average, accounting for request sizes and changed token consumption. Tokens are the billable units into which the model divides content.

The model’s updated tokenizer counts the same input text as approximately 30% more tokens than Haiku 4.5, depending on the content. Anthropic’s documentation therefore tells developers to recount prompts and revisit output limits and cost estimates rather than reuse their old measurements.

Rates for prompts up to 100,000 tokens

Input

$1.00→$0.10

USD per million tokens

Published input rates fall from $1.00 for Haiku 4.5 to $0.10 for Haiku 5.5.

Output

$5.00→$0.50

USD per million tokens

Published output rates fall from $5.00 for Haiku 4.5 to $0.50 for Haiku 5.5.

A supporting worker, not the lead coder

Anthropic positions Haiku for narrowly scoped, high-volume jobs rather than the hardest coding assignments. Financial AI company Rogo supplied one example: a larger model builds a presentation while a Haiku subagent—a model assigned one piece of the work—retrieves the revenue figure needed for a slide.

Haiku 5.5 can hold one million tokens of context, up from 200,000 in Haiku 4.5, and produce up to 128,000 output tokens, twice its predecessor’s limit. It also introduces adjustable effort to the Haiku family. Adaptive thinking is enabled by default, with medium effort as the starting setting.

The effort setting matters when reading performance claims. Anthropic’s table gives Haiku 5.5 a 39.2% score on Terminal-Bench 4.0, a test of multistep command-line work, against Sonnet 5.5’s 70.6%. VentureBeat notes that Haiku’s chart places the roughly 39% result at maximum effort, while medium scores roughly 20%. These are vendor-reported results.

Anthropic still recommends Sonnet and Opus for complex coding tasks. Its speed claim has a separate boundary: Haiku 5.5 is its fastest model at standard speeds, but Opus models running in Fast Mode are faster.

Available now, with migration work attached

Haiku 5.5 is available through the Claude API, Amazon Bedrock, Google Cloud, Microsoft Foundry and Claude Platform on AWS. Its Claude API identifier is claude-haiku-5-5. For existing integrations, the migration guide details changes beyond replacing the model name:

• Replace manually budgeted thinking with adaptive thinking. The old budget_tokens configuration returns an error.
• Remove unsupported sampling settings and stop ending requests with an assistant message for the model to continue.
• Allow room for thinking within output limits. A small limit can stop a response before any answer text appears.

There is also a capacity-planning distinction: Haiku 5.5 does not support Priority Tier. Organizations with a Haiku 4.5 commitment under that service must plan capacity separately.

The coordinated announcement also halves Sonnet 5.5 cache-read rates to $0.10 per million tokens. Cache reads reuse stored context; Anthropic estimates around 20% savings on most agent workloads. Monthly API credits roll out this week: $100 for Max 5x, $200 for Max 20x and up to $500 pooled across Team users, usable on any model.