AI models

Sawala Cloud runs your AI features on a curated set of third-party models. You pick which one your assistant uses from a dropdown; you never manage API keys, servers, or capacity, and you never sign a separate contract with a model maker.

Models do four things on the platform: they answer questions and write replies, they transcribe recordings into text, they generate images from a description, and they speak text aloud. Every model on this page is listed with its maker, so you always know whose model is answering your customers.

The three places a model runs

Every availability statement on this page is written in terms of three surfaces, so it is worth fixing what they mean:

• Crew — the assistant your own team chats with inside Ajena, grounded in the documents you have uploaded to your knowledge base. Low volume, quality first.
• Connect — the assistant that replies to your customers on messaging channels such as WhatsApp. High volume, and it needs to be fast, affordable, and good at your customers' language.
• Flow — Ajena's automation builder. A flow is a sequence of steps, and several kinds of step call a model: one writes text, one transcribes a voice note, one generates an image, one speaks a reply.

Not every model is available on all three. Check Which models work where before you plan around a particular model — it is the difference between choosing a model for your team and choosing one your customers will ever see.

Chat and reasoning models at a glance

Ten models answer questions and write replies. They are listed cheapest first, so the table reads as a price ladder.

Each model is explained in full on Chat & reasoning models — what it is good at, how well it handles Bahasa Indonesia, and why it is offered where it is.

The context window is how much text the model can consider at once: your instructions, the passages retrieved from your knowledge base, and the conversation so far, all together. A bigger window lets the assistant hold more of your material in view. Note the spread: the tightest windows on the platform are 24K and 33K tokens, while every newer model offers 128K or more.

Default models

If nobody has chosen a model, each surface falls back to a platform default:

Treat this as a snapshot rather than a promise: defaults move as models improve, and the Model page for each surface always marks the one in force for your organization. Once you choose a model yourself, your choice takes precedence — a later change to the platform default leaves you where you are.

Which models work where

Some models are offered on every surface, some only where they make sense. This is the full picture, all sixteen models on the platform:

Available everywhere — general purpose (6)

Crew only — premium reasoning (3)

Connect and Flow only — ultra-cheap (1)

Flow only — audio and images (6)

The pattern behind those groups:

The six general-purpose models are the middle of the range: capable enough for your team's internal work, cheap and fast enough for customer volume. If you are unsure, choose from these.

The three Crew-only models are the platform's most capable and most expensive. You cannot point your customer-facing Connect assistant at them, by design: they are shaped for deep work over long material — where a handful of high-quality answers a day is the usage pattern — and on a busy WhatsApp channel they would multiply the cost of every reply without making short answers any better.

Granite 4.0 Micro is the mirror image. It is the cheapest model on the platform and is offered exactly where high volumes of small, cheap lookups happen, but it is deliberately withheld from Crew: a model this small is the wrong tool for grounded research across your internal documents.

The audio and image models appear only inside flows. A chat surface never transcribes a recording, draws a picture, or speaks, so these six never show up in a Crew or Connect model picker — only in a flow step's own Model field.

Models for audio and images

These are covered on Transcription models, Image generation models, and Speech synthesis models.

How to choose

Work through these in order — the first question usually eliminates most of the list.

Which surface is this for — your team, your customers, or an automation? Answer that and the matrix above narrows the list before cost or context window enters into it.

Are you answering customers at volume? Start with your surface's default and change it only when you have a reason — it is a cheap, capable, well-exercised starting point.

Do you need the assistant to read long documents in one pass? Then look for a 128K-or-larger window — a long contract plus a conversation will not fit in the 24K and 33K windows.

Do you mostly reply in Bahasa Indonesia? Prefer the models marked as strong multilingual choices, GLM-4.7 Flash and Qwen3 in particular. Be aware that the GPT-OSS models tend to answer in English even in an Indonesian conversation.

Is cost the constraint? The $ tier is genuinely usable — for looking things up and writing short replies it is often indistinguishable from the tiers above it.

Are you doing hard multi-step reasoning over long material, for your own team? That is exactly what the Crew-only models exist for. Use them there and leave your customer channel on something cheaper.

Where you choose a model

A change to your Connect model applies to the very next incoming message — there is nothing to deploy and nothing to wait for.

What it costs

The $ to $$$$ marks on this page are a relative guide to AI-compute cost, not a price. They rank the models against each other for one representative reply, and they are the same marks the model picker shows, so the docs and the product always agree.

Two things drive what a reply actually costs: the size of the prompt (your instructions plus the passages retrieved from your knowledge base) and the length of the reply. A model with cheap output can still be expensive on a surface that sends it a large prompt every turn — which is why the tiers here blend both rather than ranking on output alone.

Your actual usage is recorded per surface and per model, and you can read it in Ajena under Settings → Usage. What that usage costs on your plan is covered by your Sawala Cloud pricing.

Where your data goes

Next: read the full entries for the models that answer your questions in Chat & reasoning models.

Model list last verified: 2026-08-07.