How your workspace chooses among them
With Auto on, Geyser picks the AI for each kind of work from the companies your workspace’s AI limits allow. When Auto chooses your Agents’ main model, it only picks companies whose terms say they never train on what you send; if a company’s terms don’t say, Auto leaves it out (an Agent that runs on Mistral Vibe stays on Mistral). To keep your work away from a company altogether, an Owner or Admin can tick it under “Providers to avoid” in the AI limits.
If you connect your own AI account or key, the work goes to that company under your own agreement with it. Models running on your own computers send nothing to any of these companies.
Anthropic
- Who
- Anthropic, PBC
- What it does for Geyser
- Claude models: chat, coding, helpers and background work such as memory.
- Where it processes
- Anthropic’s global processing
- Trains on your content
- No. Anthropic doesn’t train on API content.
- How long it keeps it
- Deleted within 30 days. Content flagged under Anthropic’s usage policy can be kept up to 2 years. Batch results are kept 29 days.
OpenAI
- Who
- OpenAI OpCo, LLC
- What it does for Geyser
- GPT models: chat, coding with Codex, helpers and background work. Also spoken replies, and the embeddings that help Agents find things in their memory.
- Where it processes
- OpenAI’s global processing
- Trains on your content
- No. OpenAI doesn’t train on API content unless the account owner opts in to share it.
- How long it keeps it
- Abuse-monitoring logs are kept up to 30 days. Batch output files are deleted 30 days after the batch finishes.
- Who
- Google LLC
- What it does for Geyser
- Gemini models, Gemini speech for talking with your Agents, and making images and video.
- Where it processes
- Google’s global processing
- Trains on your content
- No. On the paid Gemini API, which Geyser uses, Google doesn’t use prompts or answers to improve its products.
- How long it keeps it
- Prompts and answers are kept up to 55 days for abuse monitoring.
Together AI
- Who
- Together Computer, Inc.
- What it does for Geyser
- Open-weight models such as Qwen, GLM and DeepSeek, run on Together’s servers.
- Where it processes
- Together’s processing locations
- Trains on your content
- No. Sharing data for training is off unless the account owner turns it on.
- How long it keeps it
- Together stores prompts and answers by default and doesn’t state for how long. Batch files are kept 7 days. Some models are served by their original maker through Together.
Mistral AI
- Who
- Mistral AI SAS
- What it does for Geyser
- Mistral models, for Agents that run on Mistral Vibe or when someone in your workspace picks one.
- Where it processes
- Mistral’s global processing
- Trains on your content
- Not under Geyser’s account settings. Mistral’s terms allow training in some cases, such as preview models or when its opt-out isn’t used.
- How long it keeps it
- Mistral runs abuse monitoring and doesn’t state for how long it keeps the content.
Meta
- Who
- Meta Platforms, Inc.
- What it does for Geyser
- Meta’s Muse models and Muse voice transcription, when someone in your workspace picks one.
- Where it processes
- Meta’s processing locations
- Trains on your content
- Meta’s standard models don’t train on what they receive. Its “Contributor” models may, so pick a standard Meta model for anything private.
- How long it keeps it
- Meta keeps content as long as it needs and flagged content up to 2 years.
Thinking Machines
- Who
- Thinking Machines Lab (Tinker)
- What it does for Geyser
- Models Tinker serves (its own Inkling, and open models such as Qwen, Kimi, DeepSeek and GPT-OSS) when you pick one for an Agent, and models trained for your own workspace.
- Where it processes
- Thinking Machines’ processing locations
- Trains on your content
- No. It doesn’t train its own models on customer content.
- How long it keeps it
- It keeps your content only as long as it needs to provide the service.
Hugging Face
- Who
- Hugging Face, Inc.
- What it does for Geyser
- Open models through Hugging Face Inference Providers when you pick one for an Agent, and a private Inference Endpoint that can serve a model trained for your own workspace.
- Where it processes
- The region chosen for an endpoint. Inference Providers pass each request to the company that runs the model (listed below).
- Trains on your content
- No. It doesn’t train on your content, and it doesn’t store requests it passes to a provider.
- How long it keeps it
- Request logs without content are kept up to 30 days. Hugging Face’s terms set how long it keeps other data.
DeepInfra
- Who
- DeepInfra
- What it does for Geyser
- Runs open models picked through Hugging Face Inference Providers.
- Where it processes
- DeepInfra’s processing locations
- Trains on your content
- No. DeepInfra doesn’t train on content sent to its API.
- How long it keeps it
- Prompts stay in memory only while a request runs, and answers are deleted once they’re sent back.
Cerebras
- Who
- Cerebras Systems Inc.
- What it does for Geyser
- Runs open models picked through Hugging Face Inference Providers.
- Where it processes
- The United States and other countries
- Trains on your content
- No. Cerebras’ terms don’t let it train on your content.
- How long it keeps it
- It doesn’t keep prompts or answers from its inference service.
Changes to this list
We update this page before Geyser starts sending work to a new AI company. Questions go to hello@geyserlabs.ai.