Anthropic releases Claude Haiku 5.5 at sharply lower prices, halves Sonnet 5.5 cache read cost, and adds monthly API credits for subscribers

Anthropic released Claude Haiku 5.5, its small model, on AWS, Google Cloud and Azure. It is the first Haiku with adjustable reasoning levels. For prompts up to 100k tokens it costs $0.10 per million input tokens and $0.50 per million output tokens (Haiku 4.5: $1 and $5), and prompts over 100k cost 5x as much. Anthropic also halved the Sonnet 5.5 cache read price from $0.20 to $0.10 per million tokens, added monthly API credits for subscribers ($100 for Max-5x, $200 for Max-20x, up to $500 for Team), and put beta computer use and browser use support into its Python and TypeScript SDKs.

Teams running high-volume or agentic workloads on Haiku 4.5 can re-test their costs now, because the list price per token has fallen by up to 90% for prompts under 100k. The cheaper Sonnet 5.5 cache reads (Anthropic claims about 20% off most agentic tasks) matter most to anyone running long, repeated contexts. The subscriber credits are a modest sweetener, not a pricing change. Impact is real but depends on your prompt lengths and on how much the new tokenizer inflates token counts.

The "pricing arms race is far from over" framing and the claim that the Sonnet cut is "almost certainly" a reaction to GPT-6.1 are speculation, not evidence. The benchmark gains (for example Terminal-Bench 4.0 at 39.2% vs 0%) are Anthropic's own figures, and a jump from near zero says as much about the weak baseline as about the new model. The "75% cheaper on average" headline is also softened by the tokenizer using slightly more tokens per task, and by the 5x surcharge above 100k tokens.

1. 01Independent benchmarks and cost-per-task tests of Haiku 5.5 against Haiku 4.5 and GPT-6 Luna, including the real token inflation from the new tokenizer.
2. 02Whether OpenAI or Google cut prices or change their small-model tiers in response over the coming weeks, which would support the arms race claim.
3. 03Developer reports on reliability of the beta computer use and browser use SDK support, and on how the 5x pricing above 100k tokens affects real workloads.