Claude Haiku 5.5, the new small model in Anthropic's Claude family, was released on 7 October 2026 at a fraction of the price of the model it replaces. Anthropic says it costs around 75 per cent less to run on average than Claude Haiku 4.5, while scoring far higher on tests of office work, computer use and reasoning. Small models like this do the unglamorous, high-volume jobs inside business software: sorting, summarising, extracting and answering quick questions. Their price and ability shape the AI features in tools a small business already uses.
What happened
Anthropic announced the release in its post introducing Claude Haiku 5.5, calling it the cheapest, fastest and most capable small model the company has released. The Claude range runs from large, expensive models such as Opus 5.5 and Sonnet 5.5 down to Haiku, the small, quick one meant for tasks that are repeated thousands of times.
The company says Haiku 5.5 is designed for high-volume, cost-sensitive work such as summaries, database queries and classification, for speed-sensitive work such as live customer support, and for use as a helper to the larger models on coding tasks. It is the first Haiku model with an adjustable effort setting, which lets a developer choose between a cheaper, quicker answer and a more carefully reasoned one.
Two other changes came with it. Anthropic halved the price of cache reads on Claude Sonnet 5.5, a charge for re-using material the model has already been sent, and says this makes Sonnet 5.5 around 20 per cent cheaper on most agent tasks. It also introduced a monthly API credit for Max and Team plans, for subscribers who want to build their own tools on its developer platform.
Key details
| Detail | Claude Haiku 5.5 | Claude Haiku 4.5 |
|---|---|---|
| Input price per million tokens | US$0.10 for prompts up to 100,000 tokens, US$0.50 above | US$1.00 |
| Output price per million tokens | US$0.50 for prompts up to 100,000 tokens, US$2.50 above | US$5.00 |
| Computer use (OSWorld 2.1, offline subset) | 72.4 per cent | 15.7 per cent |
| Knowledge work (GDPval-AA v2.1 rating) | 1620 | 735 |
| Reasoning (Humanity's Last Exam, no tools) | 45.9 per cent | 10.2 per cent |
| Context window | 1 million tokens | 200,000 tokens |
A token is the unit of text these services bill by, roughly a short word or part of a longer one. The prices and benchmark figures are from Anthropic's announcement, and the benchmark results are the company's own. The context window, meaning how much material the model can take in at once, is from Anthropic's developer page on what is new in Claude Haiku 5.5. The model is available on Anthropic's own platform and through Amazon Web Services, Google Cloud and Microsoft Azure.
Anthropic's headline figure needs one footnote, which the company supplies itself. The list price is 90 per cent lower for requests up to 100,000 tokens and 50 per cent lower above that. The new model also counts text differently, so the same document uses more tokens than before. The average of around 75 per cent takes both into account.
Why it matters
Most AI in business software is small-model work. Reading an incoming email and deciding which queue it belongs in, pulling the supplier name and total from an invoice, summarising a long thread, answering "where is my order": none of this needs the largest model, and all of it happens constantly. The cost of the small model sets the cost of those features.
The capability jump is the bigger story. On OSWorld 2.1, which tests whether a model can operate a computer to finish multi-step tasks, Anthropic reports 72.4 per cent for Haiku 5.5 against 15.7 per cent for Haiku 4.5. On that test the small, cheap model has gone from rarely finishing the work to finishing most of it. Anthropic's customer examples point the same way. HubSpot's Ze'ev Klapow, a distinguished software engineer, is quoted saying that on an audit task in its CRM (customer relationship management) system "Haiku 5.5 was fastest to complete the task".
The limits are stated plainly. Anthropic says its larger Sonnet and Opus models remain the better choice for complex coding work, and that Haiku 5.5 suits narrowly scoped tasks. On Terminal-Bench 4.0, a hard test of command-line work, Haiku 5.5 scores 39.2 per cent against 70.6 per cent for Sonnet 5.5.
Independent testing broadly agrees, with caveats. Artificial Analysis, a benchmarking firm, published its evaluation of Claude Haiku 5.5 on the day of release. It scored the model 43 on its Intelligence Index, 26 points above the previous Haiku, and notes that the lower price tier matches GPT-6 Luna, a competing model from OpenAI. It also found that Haiku 5.5 uses considerably more output tokens per task than Luna at its highest effort setting, which eats into the saving, and that it scored lower on factual knowledge than competing small models from Google and OpenAI, partly because it is more willing to admit when it does not know something.
What this means for businesses
Most small and medium businesses will never choose Haiku 5.5 from a menu. They will meet it inside products: a help desk tool that drafts replies, a CRM that tidies records, a document system that answers questions about a contract. Over the coming months it is reasonable to ask your software vendors whether lower model costs will show up as better features, higher usage limits or lower AI add-on charges. None of that is automatic.
If your business has its own software or automation that calls Claude Haiku 4.5, this is not a one-line change. Anthropic's migration guide for Claude Haiku 5.5 lists settings that worked on the old model and now return errors, says the same text counts as approximately 30 per cent more tokens, and notes that the model's safety systems can decline a request, which the calling software has to handle. Budgets and limits worked out for the old model need to be recalculated, and long prompts cost more per token than short ones.
Cheaper processing also makes some jobs worth doing for the first time, such as summarising every support ticket or checking every supplier invoice against its purchase order. A lower price does not change the need to check results, or your privacy obligations for the information being processed.
A short checklist:
- Ask your main software vendors which AI models their features use and whether pricing or limits will change.
- If you have custom integrations on Claude Haiku 4.5, schedule a tested migration instead of a quick swap.
- Recalculate usage budgets, because the new model counts tokens differently and prices long prompts higher.
- List two or three repetitive, high-volume tasks that were too costly to automate and cost them again.
- Keep a person reviewing a sample of outputs, particularly anything sent to a customer.
Comingwave is a technology company that provides custom software, business systems and integrations and IT consulting to small and medium businesses. If you would like an independent view of what your current systems cost to run and where automation would pay for itself, request a free first consultation. We reply within one business day.
Key takeaways
- Anthropic released Claude Haiku 5.5 on 7 October 2026 as its new small, fast model.
- Anthropic says it costs around 75 per cent less to run on average than Claude Haiku 4.5.
- Its benchmark results are far above the previous Haiku, especially on computer use and knowledge work.
- Larger models are still the better choice for complex coding and long, difficult tasks.
- Businesses with software built on the older Haiku need a planned migration, because several settings and token counts have changed.
Frequently asked questions
What is Claude Haiku 5.5?
Claude Haiku 5.5 is the small model in Anthropic's Claude family, released on 7 October 2026. It is designed for quick, repetitive, high-volume tasks such as summarising, classifying and answering simple questions.
How much cheaper is Claude Haiku 5.5 than Haiku 4.5?
Anthropic says it costs around 75 per cent less to run on average. The list price is 90 per cent lower for requests up to 100,000 tokens and 50 per cent lower for larger ones, and the new model counts the same text as more tokens, which the average takes into account.
Is Claude Haiku 5.5 as capable as the larger Claude models?
No. Anthropic says Sonnet 5.5 and Opus 5.5 remain better choices for complex agentic coding, and its own benchmark table shows Sonnet 5.5 ahead of Haiku 5.5 on every test listed.
Will my software subscriptions get cheaper?
Not necessarily. The price cut applies to what software makers pay Anthropic. Whether a vendor passes that on as lower prices, higher limits or new features is the vendor's decision.
Do apps built on Claude Haiku 4.5 switch over automatically?
No. Haiku 5.5 has a different model name and several changed settings. Anthropic publishes a migration guide, and a developer needs to update and test the integration before moving to the new model.