Models
The models available inside Meerkats, how to choose one, and why the numbers never depend on it.
Meerkats runs its own agents on large language models from Google and Anthropic, and lets outside AI tools such as Claude connect to your business context over MCP. Which model writes the answer changes the wording and the speed. It never changes the numbers: every figure comes from a recorded query on the semantic layer, whichever model asked for it.
Models inside Meerkats#
The model picker sits next to the chat box on Cockpit. The same choice is available per agent.
| Model | Provider | Use it for |
|---|---|---|
gemini-3.1-flash-lite | The default. Fast and inexpensive; fine for monitors, reports and most questions. | |
gemini-3.5-flash | Faster reasoning for investigations with many branches. | |
gemini-3-flash-preview | Preview of the next Flash model. | |
gemini-2.5-flash-lite | Previous-generation light model. | |
gemini-3.1-pro-preview | Deeper reasoning for audits and monthly reviews. Slower. | |
gemini-2.5-pro | Previous-generation Pro model. | |
gemini-2.5-flash-image | Image generation and editing, for creative work in Assets. | |
claude-sonnet-4-5 | Anthropic | Strong writing and long-form reasoning; good for client-facing summaries. |
Models are grouped by provider in the picker, with a search box. The default for new workspaces and built-in agents is gemini-3.1-flash-lite, chosen because continuous monitoring has to be cheap enough to run all day.
The model never does the maths
Thresholds, feasibility, rankings and every metric are computed by deterministic code. The model requests them and renders the result. So a cheaper model is not a less accurate one; it is a plainer writer. Pick a larger model when the task is judgement over many branches (an audit, a monthly review), not when you want the numbers to be right.Choosing a model#
| Task | Suggested model | Why |
|---|---|---|
| Monitors and rules | gemini-3.1-flash-lite | Runs hundreds of times a day; the logic is in the rule, not the model |
| Scheduled reports | gemini-3.1-flash-lite or claude-sonnet-4-5 | Flash for internal; Sonnet when the report goes to a client |
| Questions in the chat | gemini-3.1-flash-lite | Fast; switch up if an answer needs more reasoning |
| Investigations | gemini-3.5-flash | Many branches through the relationship graph |
| Audits and monthly reviews | gemini-3.1-pro-preview or claude-sonnet-4-5 | Long, structured judgement |
| Creative assets | gemini-2.5-flash-image | Image generation and edits in Assets |
Credits are counted per run and per call, not per model, so switching model doesn’t change what a run costs you.
Bring your own AI tool#
Claude and other MCP clients connect to Meerkats over the MCP server and read the same business context, with the same metric definitions and the same provenance on every number. Your subscription to that tool is separate; Meerkats counts only the calls it makes to your workspace. Your own code uses the API.
What models see, and don't#
- A model sees only the context the task needs: the metrics, relationships, state and documents the run loads. It does not see the whole workspace, and never another workspace.
- Your data is never used to train any model, ours or a provider’s. Provider calls are made with training opt-out, under a data processing agreement.
- Connection tokens, API keys and passwords are never passed to a model.
- A model’s own judgement over raw data (review themes, creative scores) is always labelled AI-judged and never blended with facts.
Troubleshooting#
Two models gave different numbers
They can’t; the numbers come from the same query. Check the date ranges and the attribution basis of the two answers. If the numbers really differ for the same question and range, open the sources on each and send both to support.
A model is missing from the picker
Models are added and retired as providers change them. Preview models can disappear; an agent set to a retired model falls back to the default and says so in its run.
Answers are slow
Pro models are slower. For quick questions, switch back to a Flash model.