Feature Request: Adaptive Model and Reasoning Allocation for ChatGPT Work
Proposal: Introduce an optional, usage-aware mode that automatically selects and changes AI models and reasoning effort throughout a multi-step Work project.
Current problem
Users working on complex, long-running projects must frequently decide which model and reasoning level to use before knowing how demanding the next step will be.
Choosing too little reasoning can cause failed attempts, unnecessary retries, and repeated analysis. Choosing too much can waste usage allowances on simple operations.
The need to adjust settings manually also interrupts otherwise continuous workflows.
Proposed functionality
Allow Work to assess the task and dynamically allocate the appropriate model and reasoning effort.
-
Initial assessment: Use a lightweight model to inspect the task and estimate its complexity.
-
Predictive escalation: Increase reasoning effort or select a more capable model before difficult operations when justified.
-
Automatic de-escalation: Return to less resource-intensive models for routine tasks, summaries, documentation, and follow-up operations.
-
Context preservation: Transfer relevant conclusions, verified findings, and working state between model changes without unnecessary repeated analysis.
-
Budget awareness: Account for remaining usage allowances and user-selected resource priorities.
-
User controls: Offer Economy, Balanced, and Quality preferences, with optional approval before expensive model escalation.
-
Transparency: Provide a concise record of model changes, their purpose, and resulting resource usage where measurable.
Real-world example
During a ComfyUI/Ollama image-processing development project, Work may need to inspect a ZIP archive, diagnose a Python bridge failure, implement a correction, review test results, and update a technical checkpoint.
These steps require different levels of reasoning.
An adaptive system could employ Sol at lower effort for routine inspection and documentation, increase reasoning for difficult diagnostics, invoke Astra when specialized analysis is warranted, and return to Sol for subsequent routine operations.
The entire process would remain within a single Work task without requiring the user to manually select the model for every stage.
Expected benefits
-
Reduce unnecessary model usage and repeated work.
-
Improve first-attempt task success.
-
Preserve continuity across complex operations.
-
Lower the amount of manual supervision required.
-
Give users control over cost, quality, and speed.
Important implementation consideration: Routing overhead must be included when evaluating resource savings. The system should avoid expensive preliminary analysis or unnecessary model switching that costs more than it saves.
Recommendation
Introduce this as an optional adaptive routing mode alongside existing manual model controls, initially as an experimental feature with transparent usage reporting and measurable comparisons against manual selection.