Claude Haiku 5.5: Cheaper AI Models and What They Change for Swiss SMEs

On October 7, 2026, Anthropic released Claude Haiku 5.5, the newest and smallest model in the Claude 5 family. The headline is price: $0.10 per million input tokens and $0.50 per million output tokens for prompts up to 100K tokens. For a company that runs classification, extraction or support triage at volume, that changes which workflows are worth automating. It also comes with migration traps that can break existing code. Here is what is solid, what is vendor-reported, and what to do about it.

What Anthropic Released

Based on Anthropic's release notes and third-party coverage:

• Model ID: claude-haiku-5-5, a fixed ID with no date suffix.
• Context: a 1M-token context window and up to 128K output tokens.
• Price: $0.10 / $0.50 per million input/output tokens up to 100K tokens of prompt. Above that threshold the rates rise to $0.50 / $2.50.
• Effort levels: the first Haiku with an adjustable effort setting (low to max, medium by default), so you can trade intelligence against cost and latency per request.
• Headline claim: roughly 75% lower cost than Haiku 4.5 on average.

Read the Numbers Carefully

Most benchmark figures in circulation come from Anthropic itself. Vendor-reported results include 72.4% on the offline subset of OSWorld 2.1 (computer use) versus 15.7% for Haiku 4.5. Independent testing by Artificial Analysis reports an Intelligence Index of 43 at max effort, slightly ahead of Gemini 3.8 Flash (41) and GPT-6 Luna (38). Treat these as indicators, not guarantees; none of them measures your documents, your language mix or your edge cases.

Two caveats matter for your budget:

• The tokenizer. Third-party analysis reports the same text counts as roughly 30% more tokens than on Haiku 4.5. A "75% cheaper" list price therefore does not translate one-to-one into a 75% lower invoice.
• Max effort is not free. Reports show max effort consuming far more tokens and time without improving results on a coding test. Start at the default and raise effort only where an evaluation shows a gain.

Where a Cheap Model Pays Off

A small, low-cost model fits tasks that are high-volume, well-defined and easy to check:

1. Triage and classification: routing support emails, tagging invoices, flagging urgent requests.
2. Structured extraction: pulling fields from forms, delivery notes and contracts into JSON.
3. Drafting first responses that a human reviews before sending.
4. Pre-filtering for a larger model: let Haiku handle the easy 80% and escalate the rest. This is the idea behind the advisor strategy.

It is a poor fit for open-ended reasoning, high-stakes legal or financial judgement, and long agentic chains where one wrong step compounds. For those, keep a larger model such as Sonnet 5 or Opus 5. A cheaper model also does not remove the need for production-grade agent design.

Migration Traps From Haiku 4.5

Anthropic's release notes warn that code written for Haiku 4.5 can fail on Haiku 5.5:

• The sampling parameters temperature, top_p and top_k return a 400 error.
• Thinking tokens count against max_tokens, so a request can return an empty response with stop_reason: max_tokens. Raise the limit. Related: check finish reasons before parsing truncated JSON.
• Assistant prefill and the old computer-use tool are also reported as unsupported.

If your application sets a temperature on every call, a model swap is not a one-line change.

A Practical Test Plan

1. Pick one workflow with measurable output, such as ticket routing.
2. Build an evaluation set of 100 real examples with expected answers, including German, French and Italian cases if you serve all language regions.
3. Run Haiku 5.5 at default effort against your current model and compare accuracy, latency and real token usage, not list price.
4. Compute cost per correct result, including human review time for errors.
5. Roll out behind a switch so you can revert. This also keeps you portable, a point we made in our piece on AI vendor risk.

Also confirm data processing locations and your obligations under Swiss and EU rules before moving personal data to any new model.

The Bottom Line

Haiku 5.5 is not a breakthrough in capability; it is a shift in economics. Tasks that were too expensive to automate at scale may now pencil out, as long as you measure real cost per correct result, handle the migration changes, and keep a stronger model for hard cases. If you want help identifying which of your workflows are candidates, talk to our team.

Sources: Anthropic: Introducing Claude Haiku 5.5 · Claude Platform release notes · AI Weekly · Computing for Geeks · Let's Data Science