Choose AI Model
Choose the appropriate AI model for a task based on use case, cost, and capability requirements. This skill is the single source of truth for model selection and model-specific quirk documentation across Forge.
When to Use
- Selecting a model for a bot, agent, or automated task
- Resolving which model powers a given use case (routing, reasoning, execution, evaluation, indexing)
- Looking up model-specific quirks (parameter requirements, rejections, pricing, context windows)
Use Case Table
| Use Case | Recommended Model | Notes |
|---|---|---|
| forge-router | glm-5.2 | Lightweight routing/classification; cost-efficient for large combined transcripts |
| forge-thinker | gpt-5.4 | Deep reasoning/planning; large context |
| forge-executor | gpt-5.4 | Code generation/execution |
| forge-evaluator | gpt-5.5 | Evaluation/review; highest capability |
| forge-indexing | gpt-5.4-nano | Indexing/embedding generation; fast, cheap. Requires max_completion_tokens >= 300; do not pass temperature |
Decision Procedure
- Identify the use case. Classify the task type: routing, reasoning, execution, evaluation, or indexing.
- Check constraints. Consider cost budget, latency tolerance, context-window requirements, and required reasoning depth.
- Select the model from the Use Case Table above that matches the identified use case and satisfies the constraints.
- Verify availability. Confirm the model is accessible via the configured provider and API key.
- Consult the model reference file. Before invoking the model programmatically, check
Models/<model>.mdfor model-specific quirks (required parameters, rejected parameters, pricing tiers).
Convention
In knowledge files and skills, refer to use case names (e.g., “the indexing use case”), not specific model names (e.g., “gpt-5.4-nano”). The Choose_AI_Model skill maps use case names to current models. When a model is swapped, only this skill’s Use Case Table needs updating — all other docs remain valid because they reference the use case, not the model.
Model Reference Files
Per-model quirk documentation lives in the Models/ subfolder. These files are the single source of truth for model-specific behavior (provider, context window, cost per 1M tokens, parameter requirements, rejections, strengths). Model-specific quirks must not be duplicated inline in other skills or knowledge files — point here instead.