Anthropic rolled out Claude Haiku 5.5 on Wednesday across its direct developer platforms and major cloud infrastructure partners. The lightweight model brings a one-million-token context window alongside output limits reaching 128,000 tokens, but the primary shift centers on how the company restructured its API pricing tiers.
For prompts containing up to 100,000 tokens, Haiku 5.5 costs $0.10 per million input tokens, while output tokens run $0.50 per million. Prompt cache reads sit at $0.01 per million, while five-minute cache writes are priced at $0.125. That structure cuts base prompt rates 90% below the older Claude Haiku 4.5, which charged $1 for inputs and $5 for outputs.
Real-world cloud bills won't plummet quite that far, though. Anthropic deployed a new tokenizer that counts identical text as roughly 30% more tokens compared to Haiku 4.5. Taking that extra token density into account, the company estimates typical production workloads run about 75% cheaper on average. Batch processing shaves another 50% from those expenses, offering beta support for outputs up to 300,000 tokens.
Prompt size also triggers a steep price climb. Prompts stretching past 100,000 tokens automatically switch to a higher tier, where input costs jump to $0.50 per million and outputs hit $2.50. Anthropic noted that about 90% of requests handled by Haiku 4.5 remained under that 100,000-token ceiling, keeping the vast majority of calls in the lowest bracket.
Cloud availability and developer restrictions
Haiku 5.5 is available immediately as a hosted service on the Claude API, Amazon Bedrock, Google Cloud, Microsoft Foundry, and Claude Platform on AWS. The system accepts multimodal input including text and images, returning text responses with a knowledge cutoff date of June 2026.
Engineering teams migrating existing pipelines face several technical adjustments. Haiku 5.5 is the first Haiku-tier release to feature an adjustable effort setting. Adaptive thinking comes activated by default, and its effort parameter defaults to medium.
More critically, the model throws an HTTP 400 error whenever callers supply non-default values for temperature, top_p, or top_k. Anyone maintaining strict sampling settings in older scripts will need to strip those parameters out before running traffic against the new endpoints.
The pricing structure also sets up a direct contrast with rival offerings. OpenAI's GPT-6 Luna matches Haiku 5.5 on base short-context rates at $0.10 for inputs and $0.50 for outputs. Luna delays its rate increase until prompts cross 272,000 tokens, where it bills $0.20 for inputs and $0.75 for outputs. That means a 150,000-token prompt lands at a lower list price on Luna than on Haiku 5.5.
Anthropic designed Haiku 5.5 primarily for high-volume operational duties like text summaries, data compaction, classification tasks, and lightweight subagent workflows. For complex software engineering or agentic coding, the company continues to recommend Sonnet 5.5 and Opus 5.5, with Sonnet 5.5 still holding the benchmark lead at 70.6% on Terminal-Bench 4.0.
ADFiled by The AI Desk
Models, assistants and the companies and chips behind them, reported from what was released and what was claimed, with the difference kept clear.
More from this desk →
Be the first to comment
Join the argument. No password, just your email or a passkey.