Automatic Model Selection: How AI Model Routing Works
Direct answer: Automatic model selection removes the need to pick an AI model for every message. A router evaluates the current request, applies a defined policy, and sends it to a concrete answer model. It is most useful when the model pool, capability rules, fallback behavior, billing, and manual override are visible.
Fact-checked September 23, 2026. This is the canonical guide to automatic model selection. For exact product behavior, read how Magicdoor Auto chooses DeepSeek, Luna, Sol, or Opus. For the user choice, see Auto versus manual model selection.
The problem a model router solves
A multi-model workspace gives you more choice, but choice creates work. Should a short rewrite use the same model as a business diagnosis? Does a CSV request need code execution? Automatic selection moves that repeated decision into a policy while preserving manual control.
The routing loop
| Step | Router responsibility | Trust question |
|---|---|---|
| Read bounded context | Inspect enough recent text to understand the latest request | What leaves the main chat path? |
| Evaluate signals | Ask task-specific probability questions | What do the scores mean? |
| Apply policy | Use thresholds and capability rules | Why was this model selected? |
| Choose a model | Send the full answer request to one concrete model | Which model writes the answer? |
| Handle failure | Use a deliberate fallback or stop on cancellation | Does failure route cheap or strong? |
| Record usage | Attribute usage to the answering model | What am I billed for? |
The evaluator and answering model do different jobs. A decision model can choose a route without writing the response.
A concrete Magicdoor example
Magicdoor's Experimental Auto setting uses Jev from TypeSafe AI to evaluate substantive judgment and code/data execution. Magicdoor applies its evaluated quality-first-v7 policy.
| Latest request | Policy result | Answer model |
|---|---|---|
| "Rewrite this paragraph in a warmer tone." | Straightforward expression | DeepSeek V4.1 Flash |
| "Run Python on this CSV and plot churn." | Code/data execution signal | GPT-6.1 Sol, if Jev recognizes the execution need |
| "Summarize this PDF briefly." | Routine OpenAI-required context | GPT-6 Luna |
| "Diagnose the figures in this PDF." | Substantive OpenAI-required context | GPT-6.1 Sol |
| "Compare these launch options and recommend one." | Substantive inference or judgment | Claude Opus 5.5 |
| "That ignores our cash constraint. Reconsider it." | Challenged answer requiring judgment | Claude Opus 5.5 |
These examples explain policy behavior, not guaranteed outcomes.
What a trustworthy router should disclose
Model pool and capabilities
Magicdoor Auto can select only DeepSeek V4.1 Flash, GPT-6 Luna, GPT-6.1 Sol, or Claude Opus 5.5, not every model on the models page. Capability checks consider effective saved history, including prior files and container context. Code execution, PDFs, non-image files, and hard OpenAI requirements use Luna when routine or Sol when quality or execution signals warrant it; incompatible history prevents an Opus handoff. Sol selection does not guarantee a code tool will run.
Failure and cancellation
Missing, invalid, or oversized routing context, service errors, and the 2.5-second deadline fail closed to Sol for incompatible effective context or Opus 5.5 for compatible text. User cancellation stops the request and does not launch another model.
Accounting and override
Magicdoor records and bills answer usage against each concrete answer-model attempt. Separately metered provider-recovery attempts are recorded separately. Jev routing usage is operational telemetry, not added to answer-model usage. Manual model selection bypasses Jev.
Limits
- A probability is not a universal difficulty score or promised quality improvement.
- Bounded context can miss information; oversized context can trigger fallback.
- Capability and fail-closed rules can select a more expensive model.
- Policies can become outdated as models and needs change.
Magicdoor replayed 164 synthetic scenarios and made 324 live TypeSafe requests during discovery. That validates policy behavior against recorded labels. It did not generate and compare final answers or prove real answer-quality superiority.
Use Auto for mixed work. Choose manually when provider identity, a specific model behavior, or predictable cost control matters more. Read the decision guide for the tradeoffs.
Compare the categories directly: OpenRouter Auto vs Magicdoor Auto, ChatGPT automatic selection vs Magicdoor Auto, the published router ranking methodology, and when a multi-model workspace beats separate subscriptions.
Before trusting a router, compare it against a direct-model baseline and read what Magicdoor actually tested.
FAQ
What is automatic model selection?
Automatic model selection evaluates each request against a routing policy and chooses an answer model before generation. A useful router considers task demands, required capabilities, fallback behavior, and usage accounting.
Does an AI model router always choose the best model?
No. Routing is a policy decision under uncertainty, not a guarantee of the best answer. Strong systems disclose their model pool, rules, fallbacks, limitations, and manual override.
Which models can Magicdoor Auto choose?
Magicdoor Auto chooses fresh for every message among DeepSeek V4.1 Flash, GPT-6 Luna, GPT-6.1 Sol, and Claude Opus 5.5. Auto is an Experimental chat setting, not a separate priced model.
Sources
Accessed September 23, 2026.
- Current Magicdoor routing code and production notes for model pool, policy, payload, fallbacks, cancellation, and accounting.
- TypeSafe AI's Jev launch.
- TypeSafe API reference and Noul documentation.
- Vercel's Jev guide.
Related Resources
Automatic vs Manual AI Model Selection: Which Should You Use?
Decide when automatic model selection saves time and when choosing an AI model manually gives you better provider, capability, and cost control.
Best AI Model Routers and Automatic Model Selectors: A Transparent Guide
Compare automatic model selectors by category with a published rubric for consumer workspaces, developer API routers, and browser extensions.
AI Model Costs on magicdoor.ai: Current Chat Rates
Compare current chat-model input, cached-input and output prices, including Claude Haiku 5.5 long-context tiers and Mistral Large 4 guarded sale rates.
How to Evaluate an AI Model Router: Quality, Cost, Latency, and Policy
A practical, vendor-neutral router evaluation rubric: compare a direct-model baseline, blind answer quality, full billed cost, p95 latency, and failures on representative requests.