Pick the right AI model for each agent
A support agent, a Flow node and a Deep mission do not need the same model. Choose by role: latency and cost for chat, reliable JSON for extraction, tool calling and long context for missions.
ReAct agent
Short answers, tool calling and advisor handoff. Optimized for latency and cost per conversation.
Recommended models
gpt-4.1-minigemini-2.5-flashclaude-haiku-4-5Avoid · gpt-4o-mini on complex journeys: less reliable with tools.
Flow agent
Classification, field extraction and deterministic decisions in AI nodes. The cheapest model that holds the schema.
Recommended models
gemini-2.5-flash-litegpt-4.1-minimistral-small-latestAvoid · Reasoning models: unnecessary latency and cost for a bounded task.
Deep Agent
Planning, multi-step tool use, synthesis and deliverables. Optimized for reliability and long context.
Recommended models
claude-sonnet-5-5gpt-4.1gemini-2.5-proAvoid · gpt-4o-mini beyond 10 steps: insufficient budget and tool calling.
Embeddings
Document vectorization and knowledge base. Pick one model per workspace, then do not change it without reindexing.
Recommended models
text-embedding-3-smallgemini-embedding-001mistral-embedAvoid · Mixing two models in the same index: vectors are not comparable.
Key takeaway
The effective model follows the workspace configuration: platform mode, a provider assigned to the bot, or a BYOK key. ReAct, Flow and Deep Agent share the same resolution; a bot without a configured model fails explicitly instead of borrowing the global key.
One model per role: fast and cheap for chat, more capable for long missions.
Check tool calling and context window before assigning a model to a Deep Agent.
Update model-pricing.ts before using a new model: maxCostCents limits rely on those prices.