"Free AI agent" is one of the most searched phrases around agents, and one of the most misleading. In 2026, an AI agent that works on its own for hours costs real money to run: a computer, a browser, and a lot of model usage. So "free" always means something is limited. The question is what.
What "free" usually means
Across the market, free tiers combine some of these limits:
- A daily or weekly quota (credits, tokens or messages) that runs out, sometimes mid-task.
- A lighter model than paid plans.
- One task at a time.
- Fewer tools: no connected accounts, no scheduled tasks.
- Your data may train the model, unless you opt out.
- A payment card required anyway.
None of this is a scandal. It's just worth knowing before you build your routine on a free plan.
The free tiers, one by one
Manus
New accounts get a starting credit allowance, then 300 credits a day (no rollover), capped monthly. Free users get the "Lite" model and one task at a time. It's good for trying the concept, but heavy tasks use up credits quickly.
Meta Muse
A free tier with a weekly token allowance, reported at around 100 million tokens. It requires a payment card even on the free tier, and is available in the US only.
ChatGPT
The free plan is a chatbot. OpenAI's agents ("dots") start with Pro plans at $100/month, and they are not available on personal plans in the EU. See Baguette Bot vs OpenAI Dots.
Gemini
Since May 2026, Gemini's free limits are compute-based, refreshing every few hours within a weekly cap. Its agent features in Chrome are reserved for paid plans and, according to July 2026 reports, exclude the EEA.
Mistral (Le Chat, now Vibe)
Mistral's assistant has a generous free tier for chat, with connectors and memories. Its autonomous "Work" mode appears to be a paid feature. It's European, which is a plus. See French AI alternatives to ChatGPT.
Open-source agents (Hermes, OpenClaw)
The software is free. The agent is not, because you pay for a server and model usage. You can run local models on your own hardware, but they need a strong machine, and Hermes requires at least 64k tokens of context. See Baguette Bot vs Hermes Agent.
Baguette Bot
Our free plan, « Découverte », is designed to let you build real bots, not just chat:
- Up to 6 bots, each with its own job and memory.
- A cloud computer hosted in Paris (2 vCPU, 3 GB RAM, 20 GB), with a browser and a terminal. It is a shared machine on the free plan.
- Trial credits, plus models marked « Gratuits » (free).
- Routines and watchers, so bots can work on a schedule or react to a change.
- The apps: web, Mac, Linux, Android and iPhone (beta).
- Export of your bots to an open
.botfile. - No payment card. You just verify your email.
The limits, stated plainly:
- connected accounts (Gmail, Outlook, WhatsApp…) need a paid plan;
- the most expensive models are reserved for paid plans;
- answers are capped at a shorter length.
When you need more, Pro costs €29/month.
Free tiers compared
| Starting point | Agent works in background | Main limits | Available in France | |
|---|---|---|---|---|
| Manus | Daily credits | Yes | Credits, light model, 1 task | No restriction found |
| Meta Muse | Weekly tokens | Yes | Card required | No (US only) |
| ChatGPT Free | Chat only | No agent | No dot below Pro | Yes (chat) |
| Gemini Free | Compute quota | Limited | Agent features paid | Yes (chat) |
| Mistral Free | Messages quota | Work mode paid | Message caps | Yes |
| Hermes / OpenClaw | Free software | If your server is on | You pay server and API | Yes |
| Baguette Bot | 6 bots + computer | Yes (routines, watchers) | Trial credits, no connected accounts | Yes |
How to choose a free AI agent
- Decide what you want it to do. Chat? A weekly report? Monitoring a price? Free tiers differ most on background work.
- Check where it runs and who sees your data. A free tier is often where data use for training is on by default.
- Check that you can leave. If you build a good agent on a free plan, can you export it?
- Start small, then pay only for what you use regularly.
Sources
- usecarly, "Manus pricing", 2026: https://usecarly.com/blog/manus-pricing/
- The Neuron, "How to get started with Meta Muse", 9 September 2026: https://www.theneuron.ai/explainer-articles/how-to-get-started-with-meta-muse/
- eesel, "OpenAI Dots pricing", 30 September 2026: https://www.eesel.ai/blog/openai-dots-pricing
- aiplanfinder, Gemini usage limits, July 2026: https://aiplanfinder.com/gemini-usage-limits
- Mistral AI, pricing: https://mistral.ai/pricing
- Hermes Agent docs, providers: https://hermes-agent.nousresearch.com/docs/integrations/providers