Free AI models for agents — and what free actually costs
You can run agents on Choreo using genuinely free frontier models — including GPT-5-class ones. "Free" always has a mechanism, though, and most platforms make you find it in the terms. Here it is up front: the three classes of free model we offer, their exact daily caps, and the verbatim warning we show before you pick one.
Three classes of free
- OpenAI flagship — gpt-5.4, gpt-5, gpt-4.1, gpt-4o, o3. Capped at 50,000 tokens per user per day, drawn from a shared platform-wide pool of 250,000 a day.
- OpenAI mini — gpt-5-mini, gpt-5-nano, gpt-4.1-mini, gpt-4o-mini, o3-mini, o4-mini. Roomier: 500,000 tokens per user per day from a 2.5M shared pool. For most everyday agent work, this class is plenty.
- Stealth — an experimental model from an unknown publisher, served via OpenRouter. 25 requests per user per day. It may disappear without notice, and we label it exactly that bluntly in the picker. (Update, October 2026: the last one did disappear, and none is offered at the moment.)
What "free" costs you
The OpenAI classes are free because of a data deal, and we put the warning on the selection screen rather than in a policy page. Quoting it verbatim: "Free because your prompts AND the model's outputs are shared with OpenAI and used to train their models. Do not send anything confidential." The stealth model's warning is blunter still: the publisher is anonymous, your prompts and outputs are retained by them, and their no-training claim cannot be verified. That is the trade. It is a fine trade for public research, drafts, experiments and learning the platform — and the wrong trade for anything you wouldn't paste into a stranger's form.
The shared pool runs dry — and we say so
Free models have per-user caps and a platform-wide daily pool. When the pool is exhausted — by everyone's usage collectively — free models go unavailable for all users until midnight UTC. We could hide that behind vague errors; instead the product tells you to configure fallback models, and the agent runtime does the rest: when a model is rate-limited or out of quota, the agent switches to the next model you allowed, says it did, and carries on. Free usage never debits your balance, and it still shows up on your usage page like everything else — visible even when it's $0.00.
The free tier around the models
The models sit inside a plan that is also actually free: $0, one project, up to five agents, 100 runs a month, and $2 of platform credits once you confirm your email, with no card — so your first agent can run on a paid model before you've decided anything. When free stops being enough, the upgrade paths are all visible on the pricing page: monthly credits, your own API key at no markup, or the Claude Code / ChatGPT subscription you already pay for. The sane way to choose is empirical — start on the free class, watch the usage page for a week, and pay only for what the numbers say you need. The full cost breakdown walks the four routes with the actual figures.
See it for yourself — free, $2 of credits once you confirm your email, no card.
Start free