# Claude Haiku 5.5: what changes for people building products and automations with AI Author: Carlos Imlau, AI Product Builder · Certified trainer (CCP) · AI applied to work. Creator of Tapinfy, FalaTuga and Sourmize. Published on 2026-10-20 · Updated on 2026-10-20 URL: https://www.carlosimlau.com/blog/claude-haiku-5-5-o-que-muda ## Summary Claude Haiku 5.5, released by Anthropic on 7 October 2026, is the company's cheapest and fastest small model and costs up to 90% less than Haiku 4.5. It is built for short, repetitive tasks like classifying, summarising and answering; for building software, a larger model is still worth it. ## Key points - Released on 7 October 2026, it costs $0.10 per million input tokens and $0.50 per million output tokens for requests up to 100k tokens - Anthropic says it is about 75% cheaper than Haiku 4.5 on average - The same text counts roughly 30% more tokens, so redo your maths - For serious coding, Sonnet 5.5 and Opus 5.5 are still ahead - Max and Team plans now include monthly API credits **What changes with Claude Haiku 5.5?** Mostly the cost of small, repetitive tasks. Anthropic [released Claude Haiku 5.5 on 7 October 2026](https://www.anthropic.com/claude-haiku-5-5) as its cheapest, fastest and most capable small model, priced at $0.10 per million input tokens and $0.50 per million output tokens for requests up to 100k tokens. If you classify emails, summarise documents or answer simple questions inside an app or an automation, the bill gets much smaller. For building the software itself, a larger model still makes more sense. I build products with AI every day, using Claude Code and Codex, and the question I hear most from beginners is "which model should I use?". This launch is a good excuse to answer it with numbers. ## What is Claude Haiku 5.5? It is the smallest model in the Claude 5.5 family, alongside [Sonnet 5.5](https://www.anthropic.com/claude-sonnet-5-5), released on 28 September 2026, and Opus 5.5. According to the [official page](https://www.anthropic.com/claude-haiku-5-5): - **It is available in the API** as `claude-haiku-5-5`, and also on AWS, Google Cloud and Microsoft Azure. - **It is the first Haiku with an adjustable effort level**, from low to max. Low effort answers faster and spends less; high effort thinks more before answering. - **It is designed for high-volume work**: summaries, classification, database queries, live customer support and "subagents" that handle simple tasks next to a bigger model. Anthropic itself notes that for complex agentic coding, Sonnet 5.5 and Opus 5.5 remain better. On Terminal-Bench 4.0, for example, Haiku 5.5 scores 39.2% and Sonnet 5.5 scores 70.6%. These are the company's own numbers, not independent tests, so read them as a signal, not a guarantee. ## How much does Claude Haiku 5.5 cost? Prices are on the [official API pricing page](https://claude.com/pricing#api), per million tokens: | Model | Input | Output | |---|---|---| | Haiku 5.5 (requests up to 100k tokens) | $0.10 | $0.50 | | Haiku 5.5 (requests over 100k tokens) | $0.50 | $2.50 | | Haiku 4.5 | $1 | $5 | | Sonnet 5.5 | $2 | $10 | | Opus 5.5 | $4 | $20 | Anthropic says Haiku 5.5 is about 75% cheaper than Haiku 4.5 on average: 90% cheaper for requests up to 100k tokens and 50% above that. Batch processing (requests that don't need an instant answer) gets a further 50% off, according to the same page. There is a detail many people will miss. The [migration guide](https://platform.claude.com/docs/en/models/haiku-5-5/migration-guide) explains that Haiku 5.5 uses a new tokenizer and that the same text produces roughly 30% more tokens than on Haiku 4.5. The savings are real, just not exactly what the table suggests at first glance. ## In practice, how much do you save on an automation? This is my own calculation, using only the official prices, to give you an order of magnitude. Picture an automation that reads 10,000 emails a month and sorts each one (quote request, complaint, invoice, other). Say each email plus instructions is 1,000 input tokens and the answer is 100 output tokens, already counted with the new tokenizer. - **Haiku 5.5:** 10 million input tokens ($1) plus 1 million output tokens ($0.50) comes to **about $1.50 a month**. - **Sonnet 5.5:** the same job is $20 plus $10, **about $30 a month**. - **Haiku 4.5:** with roughly 30% fewer tokens on the old tokenizer, it lands **around $11 to $12 a month**. If the model thinks before answering, those tokens count as output and the bill goes up. For classification, low effort is usually enough. Always measure with your own data before deciding: these figures are for comparison, not an invoice. > Video from the channel (in Portuguese): [Agentes de IA fizeram o que ninguém mandou](https://www.youtube.com/watch?v=H9SKYEolTQQ) ## When is a small model enough, and when do you need a big one? My rule is simple: the small model does the repetitive work; the big one does the work that needs long reasoning. **Use a small model like Haiku 5.5 to:** - Classify messages, requests or leads. - Summarise texts, meeting notes or support conversations. - Pull data from a document into a spreadsheet. - Generate short, standardised text. In Sourmize, for example, AI generates UTMs for links: that is the kind of short, repetitive task where a small model makes sense. **Use a larger model (Sonnet or Opus) to:** - Build or change your product's code with Claude Code. - Plan a feature, review architecture or debug a hard error. - Run long tasks where one early mistake ruins everything after it. A common pattern is to combine both: the big model decides and the small one carries out the simple parts. Anthropic cites exactly this use, with Haiku 5.5 as a "subagent" for a larger model. ## What should you test before switching models? Switching the model in an app or automation that already works is more than changing a name. Before you touch anything in production: 1. **Pick 20 to 50 real examples** of what the automation does and save the current answers. 2. **Run the same examples on Haiku 5.5** and compare. If it gets more wrong than you can accept, don't switch. 3. **Check your token limits and cost estimates.** The migration guide points out that the same text counts more tokens and that the model's reasoning also counts towards the output limit. 4. **Review the technical changes.** The guide lists parameters that are no longer accepted, such as `temperature`. If you use Claude Code, it can do the migration for you and leave a list of things to check. 5. **Switch gradually.** Part of the volume first, then the rest. If you use tools like n8n or Make, the logic is the same: swap the model in a test workflow, compare, and only then move the main workflow. I went through this in more detail in [how to automate tasks with AI without coding](https://www.carlosimlau.com/blog/automatizar-tarefas-com-ia-sem-programar). ## Is there anything new for Claude Max or Team subscribers? Yes. Around the same time, Anthropic started including monthly API credits in the Max and Team plans. According to the [official help article](https://support.claude.com/en/articles/17154008), that is $100 a month on Max 5x, $200 on Max 20x and, on Team, $20 per Standard seat and $100 per Premium seat, up to $500 a month. Credits don't roll over, and the Free, Pro and Enterprise plans are not included. If you already pay for one of these plans, it is a way to try the API and build your first automation without opening a separate account. ## Frequently asked questions ### Is Claude Haiku 5.5 good for coding? It works for simple coding tasks and as a helper to a bigger model. For building a product end to end, Anthropic itself says Sonnet 5.5 and Opus 5.5 are better. ### Can I use Haiku 5.5 without knowing how to code? Yes, through automation tools that connect to the Claude API, or by asking Claude Code to build the automation for you. You need an account on Anthropic's platform and a spending limit set up. ### Will Haiku 4.5 stop working? Not yet. As of 20 October 2026, Anthropic's [model deprecations page](https://platform.claude.com/docs/en/about-claude/model-deprecations) lists Haiku 4.5 as active, and the company commits to at least 60 days' notice before retiring a model. Even so, it is worth testing Haiku 5.5 in good time. ### How do I know if my business needs this? If you have repetitive text tasks (emails, requests, reports), there is almost always a cheap automation waiting to be built. If you are in Portugal, my [AI training for teams](https://www.carlosimlau.com/formacao) works on your company's real cases. If you want to launch a product with AI behind it, take a look at the [SaaS Starter](https://www.carlosimlau.com/curso) or talk to me about your [MVP](https://www.carlosimlau.com/mvp).