How Magicdoor Auto Chooses DeepSeek, Luna, or Sol

Direct answer: Magicdoor's Experimental Auto setting chooses before every message. Substantive judgment or reconsidering a challenged answer goes to GPT-5.6 Sol. Otherwise, code/data execution or a hard OpenAI capability requirement goes to GPT-5.6 Luna. Remaining requests go to DeepSeek V4.1 Flash.

Fact-checked September 21, 2026. This page owns Magicdoor's exact product behavior. See the automatic model selection pillar for the category and Auto versus manual selection for control tradeoffs.

The exact decision order

Jev returns probabilities for two typed yes/no questions. Magicdoor uses a 0.30 quality threshold and 0.50 execution threshold. These are evaluated policy thresholds, not universal difficulty scores or percentages of expected quality improvement.

Worked examples

Routing is fresh per turn. A Sol answer can be followed by a simple DeepSeek transformation. Selecting any model manually from the model lineup bypasses Jev.

What Jev sees

The routing state contains up to six recent user/assistant text messages, including the latest request, plus the latest request and a capability flag. It excludes file parts and file URLs, tool payloads, system messages, separate reasoning parts, and embedded <think> blocks.

The serialized state has a 16,000 UTF-8-byte cap. Oversized state is not truncated to justify a cheaper route. User-written text can still contain sensitive strings or URLs, so this boundary is not full redaction.

Failure, cancellation, and billing

Jev routing usage is operational telemetry and is not added to user answer-model usage. If answer-provider recovery runs after a failed attempt, each metered concrete model attempt is recorded separately. Auto has no invented fifteenth-model rate and is not a public API model.

Evidence and limitations

Magicdoor tested 164 synthetic scenarios with 324 live TypeSafe requests. This validates routing policy behavior, not real answer-quality superiority. The study did not generate and compare final answers. Auto also has only three possible answer models, bounded context may miss relevant facts, and fail-closed rules can select Sol for a simple-looking request.

The Jev explainer separates Magicdoor's routing evidence from TypeSafe's vendor claims.

Try Experimental Auto in chat

FAQ

Is Magicdoor Auto a separate AI model?

No. Auto is an Experimental chat setting. It chooses a concrete answer model for each message. Usage is recorded against each concrete answer-model attempt, including any separately metered provider-recovery attempt, never against Auto.

Does Magicdoor Auto reconsider the model on every message?

Yes. Each Auto message is routed fresh. The previous answer model does not lock the next turn to the same model.

What happens if Jev cannot return a valid decision in time?

Missing, invalid, or oversized context, service errors, and the 2.5-second deadline fail closed to GPT-5.6 Sol. User cancellation stops the request without launching another model.

Sources

Accessed September 21, 2026.

• Current Magicdoor routing policy, router, model registry, tests, and production notes.
• Magicdoor's recorded discovery report for methodology and limitations.
• TypeSafe API reference and Noul documentation.