The right model for every turn.

We used Luminair to build Luminair.

Everything in this post was researched, written and tested inside the app it describes. We split the work across sessions and models, each doing what it is best at, and kept the findings in the Luminair Journal so no session had to start from zero.

Perplexity SonarPulled current rate cards and model docs, with citations, for every provider on the roster.

GPT-6 AstraWorked through the routing policy and the edge cases: follow-ups, images, stale allowance readings.

Luminair JournalHeld the research doc: sources, prices, test results and what is still unmeasured.

Luminair itselfRan the sessions side by side, then the tests, then the signed 1.0.234 build.

Luminair · three sessions in panes

Most turns are small.

A real coding session is not one long, hard question. It is a hard question followed by dozens of little ones:

• “Change the button background to the design token.”
• “Fix the typo in the header title.”
• “Add 8px of space under the card.”

If your session is pinned to a flagship model, every one of those little turns is billed at flagship rates. If it is pinned to a cheap model, the first hard question falls over. Switching the model picker by hand fifty times an hour is not a workflow.

Four steps, one receipt.

Here is what happens between pressing Enter and a model starting to work, in plain terms. Fig 01 at the top of this page is the same flow as a diagram.

The details that matter.

1 · The fastest decision is no decision

We measured Laya on one Mac with six synthetic prompts. A warm classification took 610 to 674 ms; the very first call took 1,444 ms. That is fine for a refactor. For “make the button blue” it is more time deciding than the change deserves.

So small visual edits never reach the classifier. The simple-edit rule is plain, inspectable code, and it hands the turn to a light model with zero classifier wait.

2 · Precision comes before price

Cheap is only good if it is good enough. The policy first asks how precise the work must be. Small edits may use light models. Clear security, payment or migration work requires a heavy model, and with Protect high precision work on, it never downgrades a stronger model you already picked.

3 · Switching providers has a hidden cost

A naive router compares price tables and jumps to whichever is a few cents cheaper. But a long session builds up a warm prompt cache with its provider. Move the conversation to a different provider and that cache is gone; the history has to be bridged over and paid for again.

So the scorer adds a penalty for changing provider and a small bonus for the model you already have. Out of the box, Stay with the same provider is on, so Automatic only chooses within your current provider unless you allow otherwise.

// Switching providers can lose warm caches and requires a history bridge.
if (current && current.engine !== candidate.engine) score += 1.5;
if (candidate.id === context.model) score -= 0.4;

Lower scores win. A model has to be meaningfully better suited before it is worth leaving a warm cache behind.

4 · No surprise bills, no made-up numbers

Many people run Luminair on subscriptions, but some models are pay-per-token API lanes. A turn that starts on a subscription will never move into a metered API lane unless you turn on Allow switching into paid API models. It ships off.

Allowance works the same careful way. The router reads your account meters, keeps separate pools separate, and treats any reading older than three minutes as unknown. It never turns a percentage into a guessed token count.

Where the controls live.

1. 1Open Settings › Models and scroll to Decision models.
2. 2Pick Laya to download it, or Jev to connect your Cloudflare account.
3. 3Leave each task strength on Automatic — best fit, or pin your own model per strength.
4. 4Send a prompt. Last decision shows which model ran and why.

Who does what.

Each model has a researched profile: its published price, its strength level, and a source link. Strength levels are our routing policy, not benchmark scores. Prices are USD per million input / output tokens at standard API rates.

Local and unknown models (Ollama, LM Studio, downloaded weights) stay available as manual picks; they get no invented score. The profiles carry a review date: on 27 October 2026 Automatic stops using this snapshot until it is refreshed.

Checked, and not yet checked.

We also clicked each of the four new switches off and on in a real app window and confirmed each setting was saved, and routed a payment refactor from Luna to Sol at high effort.

When those numbers exist, they will be in a follow-up post here. Until then, we would rather show you the method than a made-up headline.