What changed in the cheapest Claude model
Claude Haiku 5.5 is Anthropic's new smallest and fastest model, released on 7 October 2026. The most important part for businesses is the price. Requests with a prompt of up to 100,000 tokens cost USD 0.10 per million input tokens and USD 0.50 per million output tokens. The previous Haiku 4.5 cost USD 1 and USD 5, so in this range prices dropped by 90%. Above 100,000 tokens the rates are USD 0.50 and USD 2.50, half the Haiku 4.5 price. Cache reads (prompt caching) cost USD 0.01 per million tokens, and batch jobs through the Batch API get a 50% discount.
Anthropic says the average cost of typical workloads falls by about 75%. That figure accounts for the new tokenizer, which uses somewhat more tokens for the same text. When recalculating your budget, dividing the old invoice by ten is not enough.
The model has a 1M-token context window, output of up to 128K tokens and, for the first time in the Haiku family, an adjustable effort level (from low to max). The default is medium. Alongside the launch, Anthropic halved the cache-read price for Sonnet 5.5 (from USD 0.20 to USD 0.10 per million tokens), and Max and Team subscribers receive a monthly API credit usable on any model.
Benchmark results and their limits
According to Anthropic, Haiku 5.5 is well ahead of Haiku 4.5. On the GDPval-AA v2.1 office-work benchmark, run by Artificial Analysis, it scores 1620 points versus 735 for Haiku 4.5 and 1840 for Sonnet 5.5. On the OSWorld 2.1 computer-use benchmark the vendor reports 72.4% versus 15.7%. VentureBeat notes that these are vendor-reported results, and that the 39.2% Terminal-Bench 4.0 score was achieved at maximum effort; at the default it is around 20%. Anthropic itself still recommends Sonnet 5.5 and Opus 5.5 for complex coding.
Where Haiku 5.5 makes sense in a company
It is a model for large volumes of simple, repetitive tasks where cost and response time matter:
- classifying and routing tickets, e-mails and documents, e.g. sorting helpdesk messages by topic and urgency;
- extracting data from invoices, orders and forms into an ERP or CRM system;
- summarising long threads and meeting notes;
- first-line customer chat, with harder cases handed over to a person;
- helper tasks in agents, where Sonnet or Opus plans and Haiku carries out small steps.
For example: 10,000 tickets a day at 2,000 input tokens and 200 output tokens each comes to about USD 3 a day at Haiku 5.5 prices and about USD 30 at Haiku 4.5 prices. The real amount will be higher because of the new tokenizer and any retries, so run a sample of your own data and check usage in the console.
Availability in Poland, data and GDPR
The model is available through the Claude API (model ID claude-haiku-5-5) as well as Amazon Bedrock, Google Cloud and Microsoft Foundry. Poland is on Anthropic's list of supported countries.
With personal data, pay attention to where processing takes place. Anthropic's direct API processes requests globally, and the only option to restrict the region is US-only processing (the inference_geo parameter, with a 10% surcharge). If your data must stay in Europe, consider Bedrock or Google Cloud: both offer regional or multi-region endpoints with a 10% surcharge. Before deploying, check whether Haiku 5.5 is already available in the EU region you choose. Whatever the provider, you need a data processing agreement (DPA), a legal basis for transfers outside the EEA if data leaves Europe, and an updated record of processing activities.
The AI Act adds one more obligation: if Haiku 5.5 handles chat with customers, the person on the other side must know they are talking to an AI system.
What to watch when moving from Haiku 4.5
- The 100,000-token threshold. Each request is billed on its own, and cached tokens also count towards prompt length. Long documents in a single request cost five times more per token.
- Effort level. A higher level gives better results but uses more tokens. Set it deliberately for each task.
- Stricter cybersecurity safeguards. The model refuses, among other things, to help with penetration testing. Security teams can apply for access through the Cyber Verification Program.
- Regression tests. Before switching models in production, compare answers on a few hundred of your own examples. A new tokenizer and reasoning behaviour can change the output format.
The price drop means automations that did not pay off with Haiku 4.5 are worth recalculating. Base the choice of model and processing region on a test with your own data.