Introducing Kilo.Flux
A classified router model that picks the right model for every request, and bills one rate whichever model answers. Available today in Superbrain.
Until now, using an AI coding agent meant choosing a model first and hoping it suited the work. Pick a large one and you pay for reasoning a one-line fix never needed. Pick a small one and it runs out of depth on the bug that mattered. Kilo.Flux removes the choice. In Superbrain it appears as a single model called Flux, and everything behind it is the router’s job.
How it routes
Every request is read before it is sent anywhere. Flux classifies the work, weighs the reasoning it needs against the effort you asked for, and picks the model best placed to answer it from the pool below. Once a conversation has its model, it keeps it for as long as it runs, so the context stays cached and the reasoning stays coherent across turns.
Flux writes the prompt cache itself and keeps a model’s thinking in context where the model supports it, so the long tail of a session is read at the cached rate rather than in full. If a provider stalls before the first byte, the request is retried elsewhere, and a provider that keeps failing is rested until it recovers. None of this needs anything from you.
Four classes of work
Not every request deserves the same amount of thinking. Flux sorts work into four classes and matches each to the model and the depth of reasoning it warrants. The class changes what answers. It never changes the price.
| Class | Built for | Typical work |
|---|---|---|
Swift | Answers in a breath | Lookups, explanations, small edits, tests |
Standard | The everyday work | Multi-file changes, debugging, reviews, refactors |
Deep | Thinks before it moves | Architecture, hard bugs, changes that touch everything |
Frontier | Maximum reasoning | The problems you would hand to your best engineer |
The model pool
Flux routes across 13 models from 7 providers, from the fastest small models to the deepest frontier ones. A Swift request lands on a model that answers in a breath; a Frontier one lands on a model built to reason. Every model in the pool is measured on the same benchmarks before it serves a turn, and it stays only while it earns its place.
- Claude Haiku
- Claude Sonnet 5
- Claude Opus 4.8
- Claude Opus 5
- GPT-5.6 Luna
- GPT-5.6 Terra
- GPT-5.6 Sol
- Gemini 3.8 Flash
- DeepSeek V4.1 Flash
- DeepSeek V4
- Kimi K3
- GLM 5.3 Flash (self-hosted)
- Grok 4.6
Pricing
A turn on Flux costs tokens times one published rate, whichever model answered. The rate is the same on every plan; a plan decides how much Flux usage is included, never what it costs.
| Per 1M tokens | USD |
|---|---|
Input Everything the model reads fresh | $0.80 |
Output Everything it writes back, thinking included | $2.40 |
Cached input What it has read before and kept | $0.08 |
Availability
Kilo.Flux is live today in Superbrain on Free and Prevue, with usage included on every plan. Forge runs every model on your own API keys instead. Download Superbrain, open a project, and leave the model set to Flux.