Log inStart Free Trial
Meta Tech Provider

Which AI model should answer your WhatsApp customers?

Pick the fastest model that follows your rules in the languages your customers write. WhatsApp replies are short, so a smaller, quicker model is often enough; a larger one earns its cost on hard questions. Compare two or three models on the same 10 real questions, checking the reply, the time it took and the tokens used, before you switch one on.

easyCoexistence lets you put any model on your WhatsApp Business number, on your own Anthropic, OpenAI or OpenRouter key, and compare models with Test before customers see a reply: real answers, nothing sent, tokens shown.

45 seconds

How long a Reply with AI step waits for the model before the run fails.

600 tokens

Default Longest reply, adjustable from 50 to 2,000; it caps thinking plus text on models that think.

Fastest

Anthropic's comparative latency label for Claude Haiku 5.5; Sonnet 5.5 is Fast, Opus 5.5 Moderate.

max_tokens

Thinking tokens count toward it, so a long thinking pass can use the budget before the reply finishes.

2027

Earliest retirement year Anthropic commits to for its current Claude models, after which their ids stop working.

What matters most for a WhatsApp reply?

Speed first, then whether the model sticks to your facts. A customer on WhatsApp expects an answer in seconds, and the step gives up after 45 seconds, so a model that thinks for a long time on simple questions is a poor match for a busy number. Second, the model must follow your Instructions and the rule easyCoexistence adds under them, never inventing prices, stock, dates or policies. Third, it must write naturally in your customers' languages. Cost comes after those three, because WhatsApp replies are short and the input is capped: Instructions up to 20,000 characters and at most 20 earlier messages.

What to judgeWhere you see it
SpeedSeconds the AI step tookRuns tab: each step shows its time in seconds
Sticking to factsNo invented prices or promisesThe reply in Test, on questions your Instructions do not cover
LanguagesNatural replies in each languageTest with messages in Portuguese, Spanish and English
CostTokens in and out per reply"{in} in, {out} out tokens" on the test result and each run
LengthShort, complete messagesReplies cut off mid-sentence mean Longest reply is too low

Small and fast, or large and careful?

Most WhatsApp questions are routine: opening hours, prices in your Instructions, delivery areas, how to order. A small model answers those quickly and for few tokens. A larger model is worth it when replies need judgment, such as long product catalogs in the Instructions, mixed questions in one message, or careful wording on complaints, and when tests show the small one inventing details.

You do not have to choose one model for the whole number. Provider and model are set per Reply with AI step, and per agent, so a keyword automation for "price" can run a fast model while the catch-all runs a stronger one. Providers describe their models by speed: Anthropic, for example, labels Claude Haiku 5.5 Fastest, Claude Sonnet 5.5 Fast and Claude Opus 5.5 Moderate in comparative latency.

How do you compare two models fairly?

Keep everything else fixed and change only the model. Save the automation switched Off, open Test, and run the same set of questions one by one; then change Model, click Save, and run the set again. Write the questions from real chats, including two the Instructions cannot answer, so you see which model says a person will reply instead of guessing.

One limit of the test matters here: it sends only the sample message, without earlier messages of a chat, so follow-up questions such as "and on Sunday?" cannot be judged in a test. Check those in the Runs tab during the first days live. Tests appear there too, marked Test, with the seconds and tokens of each step, and they spend real tokens on your key.

A test set for comparing models (example)
1. Are you open on Sunday?
2. How much is delivery to the city centre?
3. Quanto custa o bolo de chocolate?
4. ¿Hacen envíos hoy?
5. I want 3 croissants for tomorrow at 9
6. Do you have vegan options?          (not in the Instructions)
7. Can I get a 50% discount?           (must not be promised)
8. My order arrived damaged            (should go to a person)
9. hi                                  (one-word message)
10. A long message with three questions in it

How should you set Longest reply for the model you chose?

Longest reply (tokens) is a hard cap, not a target: 600 by default, from 50 to 2,000. On a model that answers directly, 600 is plenty for a WhatsApp message, and length is better controlled in the Instructions with a line such as "answer in 1 to 3 short sentences".

On a model that thinks before answering, the cap covers thinking and text together. Anthropic documents that thinking tokens count toward the limit, so a long thinking pass can leave the reply cut off or empty, and an empty reply fails the run. The step sends no thinking or effort setting, so the model's default applies. For those models, raise the cap to between 1,000 and 2,000 and watch the out tokens in the first runs.

What does the model choice change on your bill?

Only the AI provider's part. Each reply costs input tokens (your Instructions, the rules added under them, the earlier messages and the new one) plus output tokens (the reply, and thinking where the model thinks), at the rates your provider lists for that model. Multiply the average tokens from your test runs by those rates and by the replies you expect in a month. Trimming the Instructions and lowering "Earlier messages it reads" reduce input on any model. easyCoexistence's per-number price and Meta's message billing to your own WhatsApp Business Account do not change with the model.

When should you switch models again?

When a provider announces a retirement, when replies drift from your rules, or when a newer model in the same family tests faster. Anthropic's model ids are pinned snapshots, so replies stay stable until you change them, while OpenRouter ids that start with ~ follow the newest model of a family. A retired id fails its runs with an error and an email to you. Switching is a one-field change: pick the new id in Model, click Save, and run your test set again. The automation's name, trigger, other steps, settings and run history stay as they are.

How to set it up

  1. Write 10 test questionsTake them from real chats, in your customers' languages, including two the Instructions cannot answer and one that must go to a person.
  2. Save the automation OffFlip the switch in the editor bar to Off and click Save, so tests run without customers receiving anything.
  3. Run the set on the first modelOpen Test and use Run test for the questions one at a time. Note the reply and the tokens in and out.
  4. Change only the model and repeatPick another id in Model, click Save, and run the same questions. Compare seconds per step in the Runs tab.
  5. Pick, set the cap and go liveKeep the model that follows the rules fastest, set Longest reply for it, then flip the switch On and save.

Common mistakes

  • Comparing models on questions the Instructions answer word for word. Every model passes those; the difference shows on questions outside the Instructions.
  • Judging follow-up handling from Test. A test sends only the sample message, so check multi-message conversations in the Runs tab after going live.
  • Choosing a thinking model and keeping Longest reply at 600. Thinking shares the cap, and a reply cut to nothing fails the run.
  • Picking the largest model for a number that mostly gets routine questions. Replies get slower and costlier with no visible gain on hours and prices.
Doing this with EasyCoexistence

Switching models is one field. We never charge for AI or mark it up. US$ 9 down to US$ 0.75 per number per month, falling as you add numbers, with the first 3 days free.

easyCoexistence is a multichannel API and MCP server that connects AI agents such as Claude and ChatGPT to WhatsApp, Instagram, TikTok, Facebook and more. On WhatsApp it uses Coexistence: the business keeps using its number in the WhatsApp Business app while the AI works through Meta's official API.

Frequently asked questions

Is the most capable model the best choice for WhatsApp?

Not by default. WhatsApp replies are short and customers expect speed; a larger model pays off on complex questions or long catalogs, and tests decide it. Many numbers run a fast model for routine replies.

Do I lose anything when I switch models?

No. The automation keeps its name, trigger, steps, settings and run history; only the Model field changes. Save, then run your test questions again to confirm the new replies.

Do open models handle Portuguese and Spanish?

Many do, but quality varies by model. The added rules already ask for the customer's language, so test with real messages in each language you receive before choosing.

Does a longer chat history make replies better?

Up to a point. 10 earlier messages, the default, covers most conversations; up to 20 helps long ones but adds input tokens to every reply. It reaches back through the messages kept, 90 days by default.

Keep reading

Ready to get started?

Set up WhatsApp Coexistence in minutes, not months. The app keeps working on the phone.

Start Free Trial

Verified on