Base44: InvokeLLM Per-Call Model Override

Per-Call Model Selection in InvokeLLM

Base44 integrations.Core.InvokeLLM method is the primary SDK entry point for generating AI responses from within backend functions. Before March 12, 2026, every InvokeLLM call used whichever model had been configured at the application level — a single setting that applied uniformly across all AI calls in the app.

Base44 changed this by adding an optional model parameter that overrides the app-level setting for a specific call. The parameter accepts a string identifier from the following list of supported models:

• gpt_5 — OpenAI flagship model
• gpt_5_mini — OpenAI lighter, faster variant
• gemini_3_pro — Google high-capability Gemini model
• gemini_3_flash — Google speed-optimized Gemini model
• claude_sonnet_4_6 — Anthropic balanced Claude model
• claude_opus_4_6 — Anthropic highest-capability Claude model

Why Per-Call Model Selection Matters

Different AI tasks inside an application often have very different requirements. A function that classifies a user intent from a short string has different latency and cost needs than one generating a detailed structured report. With a single app-level model setting, developers previously had to either accept a suboptimal model for some tasks or build custom abstraction layers outside of the SDK.

The model parameter resolves this mismatch directly. Developers can now route lightweight classification or extraction tasks to gpt_5_mini or gemini_3_flash for speed and cost efficiency, while directing complex reasoning or long-form generation tasks to claude_opus_4_6 or gemini_3_pro — all within the same application, without changing the default model for the rest of the app.

Backward-Compatible and Incremental

The parameter is optional and additive. Existing InvokeLLM calls with no model argument continue to behave exactly as before, falling back to the app-level model setting. Only calls that explicitly pass a model string are routed differently. This makes the change fully backward-compatible and safe to adopt incrementally — developers can introduce per-call overrides one function at a time without any migration risk.

The addition responds directly to developer feedback on the Base44 feedback board, where requests for dynamic LLM routing and per-call model control had accumulated significant support. The implementation covers models from all three major frontier AI providers, ensuring Base44 apps are not locked to any single vendor capabilities.