Model chooser  /  Run a customer-facing chat assistant  /  One-off

To prototype a support assistant, start with Claude Opus 5.

A pilot has one job: prove the assistant can answer your real questions well. Build it on Opus 5 first, find the quality ceiling, then test whether a cheaper model holds that line before launch.

Why Claude Opus 5

The reasons it wins for this job.

  • If the strongest practical model cannot answer your questions from your docs, a cheaper one will not either. Starting at the top separates "the model is not good enough" from "our docs are missing the answer".
  • Anthropic positions Opus 5 as its default for most workloads, and it holds a top-ten place on LMArena, where answers are judged by blind human preference.
  • Its 1M-token context lets a pilot skip retrieval entirely: put the whole help centre in the prompt, cache it, and test answers the same afternoon.
  • Opus 5 supports mid-conversation system messages, so the pilot can switch modes (for example, hand-off to a human) without rebuilding the cached prompt.

How to use it

4 steps to a first result.

  1. 1Export your help centre and top ticketsTake the docs plus 50 real questions customers asked last month, with the answers your team gave.
  2. 2Build the pilot in a Claude ProjectUpload the docs and write the assistant's instructions. No code yet.
  3. 3Run the 50 questionsGrade each answer: right, wrong, or should have handed off. Fix the docs where the answer was missing.
  4. 4Then test the cheaper modelRun the same 50 on Sonnet 5. If the grades hold, that is your production model.

Instructions for the pilot

Start from this.

Edit the parts in capitals, then run it.

You are the support assistant for COMPANY. You answer questions from customers about PRODUCT.

Rules:
- Answer only from the attached documentation. If the answer is not there, say you are not sure and offer to connect them to a person.
- Never promise refunds, discounts, delivery dates or features. Hand those to a person.
- Link to the help article you used.
- Two short paragraphs at most. Plain words.
- If the customer is angry or mentions cancelling, apologise once and offer a person immediately.

Hand-off message: "I'll pass this to the team. You'll hear back at EMAIL within TIME."
Nothing is sent anywhere. It copies to your clipboard.

Alternatives that also work

If Claude Opus 5 is not an option.

Claude Sonnet 5Official page →

Anthropic · closed · cost: medium

Pick it when you already know from a previous pilot that a mid-tier model is enough.

OpenAI · closed · cost: high

Pick it when your company is standardised on ChatGPT Enterprise and the pilot has to live there.

Gemini 3.8 FlashOfficial page →

Google · closed · cost: low

Pick it when the assistant needs to handle voice or images, such as a customer sending a photo of an error.

Watch out for

Doing this again and again? For a support assistant in production, run Claude Sonnet 5. Live chat is judged on answer quality, response time and cost per conversation, all at once. Sonnet 5 is Anthropic's speed-and-intelligence balance, handles tool calls for order lookups reliably, and caches your knowledge base between turns.

Sources

Checked . Models change monthly; we re-check this page when they do.

Picking the model is the easy part.

Wiring it into a workflow that runs every week, with evals, fallbacks and a cost you can predict, is the work. Fifteen minutes, no deck.

Book fifteen minutes →