What changed
On October 7, 2026 Anthropic released Claude Haiku 5.5 (claude-haiku-5-5) on its API and major cloud platforms. The model has a 1M-token context window, 128K maximum output and adaptive thinking with an effort parameter. For prompts up to 100,000 tokens, API pricing is $0.10 per million input and $0.50 per million output tokens, versus Haiku 4.5's $1/$5; above 100,000 prompt tokens, rates rise to $0.50/$2.50. Anthropic estimates about 75% lower average cost per task than Haiku 4.5 after accounting for tokenization differences. It also cut Sonnet 5.5 cache-read prices in half and announced monthly API credits for Max and Team subscribers.
Why it matters
High-volume AI products spend substantial money on short repetitive calls and delegated agent tasks. A 90% token-rate reduction at short prompts can change the economics of classification, extraction, summarization and support workloads, but a 1M context window is not a promise of flat-rate processing. The 100K pricing boundary and newer tokenizer matter to actual cost per completed task. Developers should benchmark quality, latency and token usage under their own data rather than treat vendor comparisons as universal.
Short prompts get a steep price cut; long prompts have a different tier
Up to 100K prompt tokens, input/output rates are $0.10/$0.50 per million; above 100K, they are $0.50/$2.50. Cache writes/reads and batch discounts have their own rates. Anthropic says the typical task costs around 75% less than on Haiku 4.5, not 90%, because output/token consumption can differ.
Context and effort controls move downmarket
Haiku 5.5 offers a 1M-token context window, up to 128K output, and adaptive thinking with adjustable effort. The API model ID is claude-haiku-5-5.
The rest of the Claude stack also changes price
Anthropic cut Sonnet 5.5 cache-read cost from $0.20 to $0.10 per million tokens and says it lowers typical agentic cost about 20%. It also announced monthly API credits for Max/Team users, with rollout and plan-specific limits.
Vendor benchmark comparisons need independent evaluation
Anthropic reports substantial gains over Haiku 4.5 and publishes early customer examples, but model choice for production agents depends on task success rate, retries, latency and risk rather than one benchmark score.