Haiku 5.5 Launches with Entry Price of $0.10 per Million

Anthropic launched the Claude Haiku 5.5 on Wednesday (October 7), its small model, with an entry price of $0.10 per million tokens and an output price of $0.50 for prompts up to 100,000 tokens. The predecessor, Haiku 4.5, charged $1 and $5 for the same volume. For prompts exceeding 100,000 tokens, the new model costs $0.50 on entry and $2.50 on exit, half the charge of the previous generation.

The figure the company highlights is different: an average cost that is 75% lower. The discrepancy between the 90% in the table and the 75% average is noted in a footnote of the announcement. Haiku 5.5 uses the tokenizer adopted starting with Claude 4.7, and Anthropic's technical documentation indicates that the same text now counts about 30% more tokens than in Haiku 4.5. The company also assumes that around 90% of requests for Haiku 4.5 stayed below the 100,000-token threshold.

Price Comparison with GPT-6 Luna

The short tier of Haiku 5.5 matches the GPT-6 Luna, which OpenAI launched on September 22 at $0.10 and $0.50 per million tokens. Above 100,000 tokens, Luna remains cheaper. In tests published by Anthropic itself, Haiku 5.5 scores 1620 Elo points on GDPval-AA v2.1, compared to 1437 for Luna and 735 for Haiku 4.5. In OSWorld 2.1, which measures the autonomous operation of a computer, the scores are 72.4% for Haiku 5.5, 48.9% for Luna, and 15.7% for its predecessor.

The same graphs reveal the limits of the model. In Terminal-Bench 4.0, Haiku 5.5 scores 39.2%, less than half of Sonnet 5.5’s 70.6%, which costs $2 and $10 per million tokens. These numbers are from the supplier, still awaiting complete independent evaluation. Artificial Analysis, which publishes its own index, gives Haiku 5.5 a score of 43 in the maximum effort mode.

The context window has increased from 200,000 to 1 million tokens, and the maximum output has risen from 64,000 to 128,000. The model is available via the Anthropic API, on Amazon Bedrock, Google Cloud, and Microsoft Foundry, with guaranteed withdrawal not before October 7, 2027.

User Experiences

Anthropic has published testimonials from clients with metrics. According to Aaron Vinh, an engineer at Asana, latency in task completion has decreased by over 30%. Ze'ev Klapow, an engineer at HubSpot, states that the model achieved 92.8% on the company’s internal CRM battery, the best score ever recorded for a small model there. At AlphaSense, Daniel Campos reports a score of 0.84 across 400 queries, compared to 0.76 for Haiku 4.5. Rogo uses the model as a sub-agent to extract segment numbers from 10-K reports.

Cognition, in turn, employs Haiku 5.5 as an auxiliary model in Devin, a design that Anthropic suggests for classification, routing, extraction, and sub-agents coordinated by a larger model.

Migration Costs

Those who have Haiku 4.5 in production do not change the model identifier and continue as usual. The migration guide lists changes that may break code: the reasoning mode with manual token budget now returns errors, as do outlier values for temperature, top_p, and top_k, and the pre-filling of the assistant's response. Adaptive reasoning is enabled by default, and reasoning tokens count against the output limit, which may truncate responses in integrations utilizing low limits.

For an integrator, this means repricing automation contracts that were budgeted in tokens. A document that cost 1,000 tokens in Haiku 4.5 now costs nearly 1,300, at a unit price ten times lower in the short tier. The figures favor the client, but budgets and consumption caps need to be revised.

Impact on Hourly Services

In the fiscal fourth-quarter balance, released just days ago, Accenture reported that price pressure appeared on multiple fronts and that the competition should remain intense. A model capable of operating systems at one-tenth the previous cost shifts the negotiation for managed service contracts, from India to shared service center operations in Brazil and Poland: the cost per transaction now competes with an API priced per million tokens.

Anthropic took the opportunity in the announcement to halve the cache read price of Sonnet 5.5, from $0.20 to $0.10 per million tokens, making the model about 20% cheaper for most agent tasks, according to the company. Combined, these two moves indicate where the competition lies: in the cost of running agents by the hour, not in the price of the first query.