Anthropic Launches Haiku 5.5 with Average Price Drop of About 75%
Anthropic has released the small model Claude Haiku 5.5, claiming it to be the fastest, cheapest, and most capable small model to date, with an average operating cost approximately 75% lower than Haiku 4.5, targeting high-frequency, low-latency, cost-sensitive tasks such as summarization, classification, extraction, routing, and sub-agents.
For prompts under 100,000 tokens, the input price is $0.1 per million tokens, and output is $0.5; beyond 100,000 tokens, input rises to $0.5 and output to $2.5. Cache read prices are $0.01 and $0.05, while cache write prices are $0.125 and $0.625. The corresponding prices for Haiku 4.5 were $1 for input and $5 for output.
The company states that requests under 100,000 tokens account for about 90% of the previous generation Haiku's usage, with this pricing tier seeing a reduction of about 90%, and longer prompts dropping by about 50%. The new tokenizer increases the token count for equivalent text by about 30%, resulting in an adjusted average cost reduction of approximately 75%. The batch interface offers an additional 50% discount on input and output.
The model ID is claude-haiku-5-5, with a context window of 1 million tokens and a maximum output of 128,000 tokens, with knowledge cutoff in June 2026. It is now available on the Claude platform, Amazon Bedrock, Google Cloud, and Microsoft Foundry; Bedrock coverage includes the U.S., EU, Australia, Japan, and global inference configurations, with GovCloud also open.
On the same day, the cache read price for Sonnet 5.5 was cut from $0.2 to $0.1 per million tokens, a 50% reduction. Cache reads account for a significant portion of token consumption in agent tasks, and the company claims this will reduce the operating cost of Sonnet 5.5 by about 20% for most agent tasks. For short prompts, Haiku 5.5's input and output are about one-twentieth of Sonnet 5.5's.
Buyers are applications and cloud customers moving classification, summarization, and sub-tasks from large models, while sellers are Anthropic's token-priced inference capacity. The event is driven by the generational shift of small models and the price reduction of mid-tier model caches. Funding is shifting from high-priced long context calls to low-priced short calls, with Anthropic using the price difference to retain usage within its own tier, while cloud vendors receive new SKUs for resale. The pressure is on alternative small models still priced at Haiku 4.5 and customers continuing to run simple tasks on Sonnet.
Short prompt input drops from $1 to $0.1, and output from $5 to $0.5, with the price difference itself serving as a migration incentive. Prices jump fivefold after 100,000 tokens, pushing long context back towards Sonnet and Opus.
Source: Public Information
ABAB AI Insight
Anthropic's product tiering has always been priced according to task difficulty rather than parameter promotion. Opus handles long tasks requiring judgment, Sonnet handles routine coding and documentation, and Haiku handles parallel short calls. Sonnet 5.5 will launch on September 28, 2026, maintaining a price of $2 for input and $10 for output, reducing single-task costs by up to about 30% through lower token usage; Haiku is explicitly forecasted to follow in a few weeks. This replacement is not a new brand but a breaking of the price barrier at the bottom tier.
The capital path for price reduction is to lock in usage rather than clear out through discounts. Requests under 100,000 tokens account for about 90% of the previous generation Haiku, with this tier directly dropping to $0.1 for input, prompting applications to move classification, routing, and sub-agents down from Sonnet, increasing total token volume while decreasing average price per order. Simultaneously halving the cache read price for Sonnet prevents mid-tier models from being hollowed out by small models in agent loops: the main control remains in Sonnet, with repeated contexts going through the cache, and terminal calls using Haiku. Amazon Bedrock, Google Cloud, and Microsoft Foundry are available for sale on the same day, effectively passing the price difference to the three clouds for resale.
Comparable moves include OpenAI integrating lightweight models into real-time and batch interfaces, and Google using Flash to handle high concurrency. The industry position is in the price control phase following expansion: cutting-edge models have been set, and competition has shifted to price per million tokens and cache hit rates. CoreWeave and Lambda sell compute time, while Anthropic sells task layers that can be priced from compute time.
This represents a transfer of pricing power. The mechanism is that cache reads and short calls occupy an excessively high proportion of inference costs; whoever first prices these two items down will rewrite the model routing for applications. The fivefold price jump after 100,000 tokens serves as a reverse valve: cheap pricing only covers high-frequency short tasks, while long contexts remain in the high-price tier, preventing small model price reductions from flattening the entire pricing table.
ABAB News · Cognitive Laws
Price reduction is not about discounts; it is about locking in usage within one's own tier.
Cheap models handle frequency, while expensive models handle judgment.
Once cache prices drop, cyclical calls can begin to scale.