Skip to main content
Point Optexity at any LLM — Gemini, Claude, GPT, or a self-hosted endpoint — by setting one environment variable. Every LLM call routes through LiteLLM, so any model LiteLLM supports works without code changes.

​
Environment Variables

The primary and fallback models can live on different providers, each with its own key. Omit a key and LiteLLM reads the provider’s own environment variable instead (GEMINI_API_KEY / GOOGLE_API_KEY, ANTHROPIC_API_KEY, OPENAI_API_KEY, …).

​
Model Strings

Write models as provider/model: See the LiteLLM provider list for the full set of supported prefixes.
Always include the provider prefix. A bare name like gemini-2.5-flash is passed to LiteLLM as-is and may resolve to a different provider than you expect.

​
Resolution Order

The model for any given LLM call is resolved from the most specific setting available:

​
Task-Level Override

Set the model once for the whole automation:

​
Action-Level Override

Give a single action a stronger (or cheaper) model than the rest of the automation:
Actions that accept llm_model_name: extraction_action.llm, extraction_action.pdf, extraction_action.locator, misc_action.llm_query, and interaction_action.captcha.
Captcha solving is the one exception to the resolution order — interaction_action.captcha defaults to gemini/gemini-2.5-pro rather than falling through to LLM_MODEL, because captcha grids need a stronger vision model. Override llm_model_name on the action to change it.

​
Fallbacks

When LLM_MODEL_FALLBACK is set, a failed primary call is retried once on the fallback model before the action errors. Each model uses its own key, so the fallback can be on a completely different provider — useful for surviving a single provider’s rate limits or outages.
Fallbacks apply to the environment-level models only. A per-task or per-action llm_model_name overrides the primary model; the same LLM_MODEL_FALLBACK still backs it up.

​
Cost Tracking

Token usage and cost are reported per task from LiteLLM’s pricing data. Reasoning and tool-use tokens are already counted inside completion tokens, so they are reported but never billed twice. If a model has no pricing entry in LiteLLM, tokens are still tracked and cost is reported as 0.

​
Migrating from llm_provider

llm_provider is deprecated. Existing automations that set it keep working — the provider and model are joined into a LiteLLM string — but new automations should use a single prefixed llm_model_name. The provider is no longer restricted to gemini, anthropic, or openai — any LiteLLM prefix is accepted.