Claude Haiku 5.5 Released: Input Price Under Half of DeepSeek’s, Small Model Price War Overturns the Table

Overview

Anthropic officially released Claude Haiku 5.5 in the early hours of October 7th. This lightweight model entered the market with a highly disruptive pricing strategy, described by the industry as directly "flipping the table." Its input price is as low as $0.1 per million tokens, less than half that of DeepSeek's similar services, and also 20 times cheaper than Anthropic's own flagship model Sonnet 5.5, and 40 times cheaper than Opus 5.5.

Performance Breakthroughs and Capability Comparison

Although positioned as a lightweight model, Haiku 5.5 demonstrated significant progress in key benchmark tests. In the Terminal-Bench terminal programming test, the model achieved a score of 39.2%, significantly outperforming GPT-6 Luna's 16.4% by more than double. In complex tasks involving computer operations, Haiku 5.5 scored 72.4%. While slightly lower than its older brother Sonnet 5.5's 83.9%, considering its extremely low price threshold, this performance is considered sufficient to handle the vast majority of demands in the budget market.

Official complete benchmark scores have not yet been released, but analysis suggests that Haiku 5.5 can achieve roughly 80% of Sonnet 5.5's processing power at a fraction of the price, likely surpassing DeepSeek's similar products on cost-effectiveness.

Pricing Strategy and Market Impact

This price adjustment marks the entry into a white-hot stage of the lightweight model price war. Haiku 5.5 has reduced costs by 90% compared to the previous generation, becoming the fastest official model in the Claude series. For comparison, the input price for the similarly sized Sonnet 5.5 is $2 per million tokens, while Opus 5.5 is as high as $20 per million tokens.

"Cheap errand-boy models are starting to abandon martial virtue."

This tiered pricing precisely targets a wide range of text-processing scenarios, such as summarization generation, classification tasks, and intelligent customer service, or can be used as an auxiliary layer for Opus models. Additionally, caching read fees for Sonnet models have also been cut in half, reducing the overall cost of Agent-based workflows by approximately 20%. Some analysts predict that Claude Pro subscribers may soon enjoy near-infinite usage space, further reshaping the cost structure of AI services.