Models & routing
Auto-routing across providers, quality bias, and pinning a specific model.
The model chip in the agent input footer controls which model answers. The default is auto: Mel routes each turn to the best model for the job across every configured provider — Anthropic, OpenAI, Google Gemini, xAI, Groq, DeepSeek, Moonshot, Z.ai, MiniMax, Alibaba and Xiaomi — without you thinking about it. See Models & pricing for the model families.
How auto routing works
auto is an instant heuristic router, not another model call. It classifies your turn in microseconds from cheap signals — prompt length, code and agentic cues, reasoning cues, context size — into a tier:
| Tier | Typical turn |
|---|---|
| Trivial | Quick questions, one-liners |
| Simple coding | Small edits, straightforward scripts |
| Complex coding | Multi-file changes, real builds |
| Reasoning | Hard analysis, tricky debugging |
| Huge context | Very large conversations or inputs |
Each tier maps to the best available model for it, so there's no routing latency before your answer starts streaming. The chip updates to auto → <model> once a turn routes, and the usage summary records the routed model plus a "why this model" line.
Bias: a ceiling, not a pin
In the picker's auto submenu you can bias the router:
- Balanced (called Responsive before v0.1.49) — the default: the quality each step needs, leaning a little toward faster models. It is not a speed guarantee. A hands-off or orchestrated run keeps the mode you picked and is routed as long, multi-step agent work.
- Cost-efficient — prefers the cheapest capable model; escalates only when it has to.
- Genius — escalates to the strongest model when the task is hard, while still using fast, cheap models for simple turns.
- Open-weights — routes only to open-weight models (GPT-OSS, Gemma, Qwen, GLM, DeepSeek, Kimi, MiniMax).
The bias raises or lowers the router's *ceiling* — it never pins one model for every turn. Genius does not mean "always the biggest model"; it means the hard turns are allowed to reach it.
The Auto mode you pick is remembered for new conversations, and the chip names it when it isn't the default — auto (genius). You can also set it in Settings → AI → Default models. Picking a specific model stays with that one conversation.
Picking a model explicitly
Click the model chip (or type /model in the conversation) to open the picker. Models are grouped by provider, each with a brand monogram, and hovering a model shows its Model Specs card — intelligence, speed, and cost bars plus its context window. Select a model to pin it for the conversation; pick an auto tier to go back to routing. Your selection persists.
When a model is unavailable
If the model Mel picked can't answer (a provider outage, a rate limit, a model the provider stopped serving), Mel moves to the next one and the conversation says which model answered instead.
Reasoning levels
Models that support an adjustable reasoning effort get a Reasoning level submenu. Higher levels think for longer and use more tokens. The level is only sent to models that actually support it.
Vision routing
When your turn includes image attachments, routing guarantees a model that can see them:
autoroutes to a vision-capable model, and any failover stays vision-capable — attachments are never silently dropped.- A pinned model without vision is upgraded to a vision model for that turn, announced in the transcript's router info.
- If no vision-capable provider is available, the turn fails with a clear error instead of ignoring your images.
For the wire-level routing contract, see Agent streaming.