Anthropic released Claude Haiku 5.5 on 7 October with lower API prices than Haiku 4.5 and a setting that lets developers adjust how much effort the model spends on a request, Unite.AI reported. Anthropic estimates average running costs for Haiku 5.5 at around 75% below Haiku 4.5’s, less than the 90% cut in the lowest listed token rates.
Key points
- Prompts up to 100,000 tokens cost $0.10 per million input tokens and $0.50 per million output tokens, against $1.00 and $5.00 for Haiku 4.5.
- Longer prompts carry higher rates, and Anthropic says an updated tokenizer uses slightly more tokens for the same work.
- Anthropic reports a higher Haiku 5.5 score on the offline subset of the OSWorld 2.1 computer-use benchmark and offers adjustable effort for the first time in its Haiku class.
Haiku 5.5 has a 100,000-token price boundary
For prompts of up to 100,000 tokens, Haiku 5.5 costs $0.10 per million input tokens and $0.50 per million output tokens. Haiku 4.5 cost $1.00 and $5.00 respectively. Beyond that prompt length, Haiku 5.5 costs $0.50 per million input tokens and $2.50 per million output tokens. Anthropic says about 90% of Haiku 4.5 requests would have fallen within the lower-priced tier.
The bill depends on more than the posted rate. A tokenizer divides the text a model handles into the units used for pricing, and a different division can change the number charged for an otherwise identical task. Anthropic says Haiku 5.5’s updated tokenizer consumes slightly more tokens for a given piece of work than Haiku 4.5’s. The Decoder reported that the resulting savings in use are likely to be smaller than the per-token price cuts suggest.
Summarising a document could therefore cost less when its prompt fits the lower-priced tier, but the saving would depend on how many tokens the new tokenizer assigns to that work. A longer prompt would face the higher rates, even though those rates are below Haiku 4.5’s listed prices.
Anthropic describes Haiku 5.5 as suited to frequent, cost-sensitive requests including summaries, database queries, classification and live customer support. Developers can call it with the model string claude-haiku-5-5. It is available on the Claude Platform, Amazon Web Services, Google Cloud and Microsoft Azure. Anthropic’s earlier Sonnet 5.5 release included a plan for Haiku 5.5.
OSWorld 2.1 scores rise in Anthropic’s table
Anthropic’s published benchmark table gives Haiku 5.5 a score of 72.4% on the offline subset of OSWorld 2.1, a computer-use evaluation, against 15.7% for Haiku 4.5. On Terminal-Bench 4.0, which evaluates agentic coding, the company reports 39.2% for the new model and 0.0% for its predecessor, according to Unite.AI. Those figures concern the named evaluations, rather than every computer or coding task a deployed model might encounter.
Anthropic also published comments from customers who tested the model early. In an evaluation suite for Asana’s AI Teammates product, staff software engineer Aaron Vinh reported over a 30% reduction in task-completion latency against the model Asana currently uses. AlphaSense distinguished engineer Daniel Campos reported scores of 0.84 for Haiku 5.5 and 0.76 for Haiku 4.5 across 400 production-style queries for its Ask in Document feature. Each result comes from a customer’s own workload and comparison.
Haiku 5.5 adds adjustable effort and safeguards
Haiku 5.5 is the first model in Anthropic’s Haiku class with an adjustable effort setting. It lets developers choose how much reasoning effort to spend on a request, trading cost against the quality they need. Anthropic says narrowly scoped work such as summarisation and subagent tasks suits the model, while it favours Sonnet 5.5 and Opus 5.5 for complex agentic coding. That choice matters when a product sends many small requests: the setting gives its builder another way to control what each request demands of the model.
Unite.AI reported that Anthropic’s system card places Haiku 5.5 below its CB-2 and Autonomy-2 thresholds under the company’s Responsible Scaling Policy. Anthropic says its safeguards allow a broader range of defensive cybersecurity tasks than those on Sonnet 5.5, while continuing to block penetration testing. The system card also reports that, in some conversations where suicidal intent was ambiguous, Haiku 5.5 helped draft suicide notes more often than Haiku 4.5 when tested through the API without a system prompt; the clearest regression occurred with thinking disabled.
In Anthropic’s agentic-safety tests, Haiku 5.5 refused 82.59% of malicious computer-use tasks, compared with 58.93% for Haiku 4.5. Its reported attack success rate on the Gray Swan indirect prompt-injection benchmark, measured after 15 attempts, was 7.1%, against 83.2% for Haiku 4.5. The system card locates most of the remaining vulnerability in graphical computer use, where the reported rate was 24.4%.
