model selector
in English: model router / model picker
The model selector decides which AI model will answer your question: either you choose from a menu, or the system chooses automatically for you.
In plain terms
Behind an AI service, there is rarely a single model. Some are fast and cheap, others slower, more expensive, but more thorough. The model selector decides which model your query goes to.
It is like a general practitioner's patient triage system. It does not send a common cold to a professor, but does not keep a complex complaint at the nurse's desk. This ensures the specialist's time is spent where it is truly needed.
It has two forms. The manual version (model picker) is a dropdown menu where you decide which model to use. The automatic version (model router) operates in the background, making decisions per query, often without you seeing anything of it.
The goal in both cases is the same: a fast, cheap response for easy questions, and a thorough one for hard ones. The provider saves money and computing capacity, while you save time.
How it works
- Your question arrives, and before any large model begins working on it, a pre-filter examines it. This can be a small, fast model or a simple rule system.
- A pre-filter estimates task difficulty: whether it is a simple fact question, a long analysis, programming, or a problem requiring multi-step reasoning.
- It also takes circumstances into account: what you explicitly requested (for example, 'think carefully'), what subscription you have, how loaded the servers are, and the cost of running each model.
- Based on this, it forwards the question to the selected model. Some systems work in stages: first, the cheaper model attempts, and only if the answer is insufficient does the stronger one step in.
- Better systems learn from feedback: they track when you manually switched to another model or which answer you rated as good.
An everyday example
You ask the chatbot at what temperature to bake pogácsa, and the answer arrives immediately. Then you ask it to compare two mortgage offers. At this point, the 'thinking' label appears, and it works for half a minute because the model selector routed the query in the background to a stronger, more deliberative model.
Who is affected
A common misconception
Limits and risks
- The selector can be wrong: behind a short, seemingly simple question, there can be a difficult problem, and in such cases, a weaker model gives a confident but superficial answer.
- You often do not see which model responded, so you may receive answers of different quality to the same question on different days.
- On free and cheaper plans, use of the more powerful model is capped within tighter limits, and once the quota runs out, the system often switches to a smaller model.
- If the router directs traffic between models from multiple companies, your data may reach multiple providers, so check the terms for sensitive content.
Similar, but not the same
- agent orchestration – Orchestration breaks a complex task into steps and distributes them among multiple agents, while the model selector decides only which model responds to a given request.
- mixture of experts – There, different sub-networks activate within a single model, whereas the model selector chooses between separate, full models.
- thinking mode – Thinking mode determines how much time a model spends reasoning over a response, while the model selector determines which model receives the question.
Related concepts
model cascade, automatic model selection, manual model selector
reasoning model, token, agent orchestration, API gateway, small language model
A brief history
One early, widely cited piece of research was the 2023 FrugalGPT paper by Stanford researchers. In it, models work in a cascade: a cheap model tries first, and a more expensive one is called only if necessary. In 2024, the RouteLLM paper (LMSYS, Berkeley) presented a router that learns from human preference data when a weaker model is enough. The general public met the idea at scale in August 2025, when OpenAI launched GPT-5 and switched on an automatic model selector in ChatGPT.
Where things stand
By the end of 2026, automatic model selection alongside the manual menu is commonplace in major chat applications and coding agents. In corporate systems it is a basic cost-reduction tool, but its decisions remain hard for users to follow.
In the news
- VS Code 1.143 makes Copilot the default, adds AI model bake-off and local AI without creditsOct 19, 2026 · Code
- GitHub Copilot auto model selection gains cost and quality modesSep 14, 2026 · Code
- Google releases Gemini 3.8 Flash, now in the Copilot model pickerSep 7, 2026 · Code
Also known as
model selector · model controller · model router · model picker · LLM router · model routing · model routing · automatic model selection
Sources
- Chen, Zaharia, Zou: FrugalGPT – How to Use Large Language Models While Reducing Cost and Improving Performance (arXiv, 2023)
- Ong et al.: RouteLLM – Learning to Route LLMs with Preference Data (arXiv, 2024)
This entry was machine-edited and has undergone machine expert review; the “In the news” section updates automatically from Agent Age stories.