DocsModels
FutureX models: Fast, Pro, Light, Eco
FutureX ships four models. Fast is the default, Pro thinks hardest, Light runs with reasoning on, and Eco keeps the bill small. Here is what each one allows, and when to reach for it. New to the CLI? Start with the quickstart.
All documentation pages
Getting started
Reference
The four models
Every model carries three things: the effort it starts at, the efforts it will accept, and a hard ceiling on how many tokens it can write back. The CLI help calls the value you pass to --tier a module — the four module ids are the ones in the first column.
| Module | Name | Default effort | Allowed efforts | Output cap | Rate |
|---|---|---|---|---|---|
| fx-pro | Pro | max | off · high · max | 64,000 | $2 per 1M tokens |
| fx-fast | FastDefault | off | off · high | 16,000 | $1 per 1M tokens |
| fx-mini | Light | high | off · high · max | 32,000 | No separate rate published |
| fx-eco | Eco | off | off | 8,192 | No separate rate published |
Two rates are published: Fast at $1 per 1M tokens and Pro at $2 per 1M tokens. See pricing for the full terms.
When to use which
Start on Fast. Move up only when the work asks for it — that is the whole heuristic, and it is usually enough.
Fast — the everyday model
Fast is the default and it is the right default. Ordinary edits, small features, questions about the repo, and most of the agent loop — the reads, the searches, the one-file changes — all land here. It accepts effort off or high, so you can switch reasoning on for one hard step without changing model.
Pro — hard reasoning and large refactors
Reach for Pro when the thinking is the bottleneck: a refactor spanning many files, a design decision with real trade-offs, a bug that has already survived two attempts. Pro starts at max effort and carries the largest output ceiling, so long plans and long diffs do not get cut off mid-sentence. It costs twice as much per token as Fast, and it earns that on the work above.
Light — reasoning without moving to Pro
Light sits in between. It starts at effort high and will also accept max, so you get a thinking pass without switching to Pro. Good for a tricky function, a test failing for a non-obvious reason, or a plan you want reasoned through before a cheaper model executes it.
Eco — cost discipline in a flag
Eco is the strict one. Reasoning is off and cannot be turned on, and the output ceiling is the smallest of the four. Use it for high-volume, low-stakes runs — bulk renames, one-line answers, scripted jobs where the total matters more than any single reply.
Picking a model with --tier
--tier picks the module for the commands that call the model — chiefly ask and the interactive session. The value is one of the four module ids; pass nothing and you get fx-fast. The CLI's own help adds the caveat worth knowing: the agent executor runs fast by design, so reach for a heavier module on the thinking commands rather than expecting it to slow the executor down. Commands that never call the model — index, verify, login and friends — accept the flag and ignore it.
futurex ask "what does the auth guard do?"
futurex agent "extract the billing client" --tier fx-pro
futurex plan "split the monolith" --pro
futurex ask "rename these constants" --tier fx-ecoTo stop typing the flag, set a default in settings. FutureX reads ~/.futurex/settings.json first, then <repo>/.futurex/settings.json, so a single repository can run on a different model from the rest of your machine.
{
"defaultTier": "fx-pro"
}The full flag list lives in commands.
Setting reasoning with --effort
--effort takes off, high, or max. Each model starts at its own default from the table above; the flag overrides that default for one run.
futurex ask "why is this test flaky?" --effort high
futurex ask "why is this test flaky?" -r
futurex build "migrate the queue" --maxMore effort means the model spends longer thinking before it answers. That is worth paying for on a hard problem and wasted on a rename, which is why the default sits at off for Fast and at max for Pro.
Effort is enforced against the model
A model does not accept every effort. Ask for one it does not allow and the run stops with an error rather than quietly dropping to something cheaper — so you always know what you asked for and what you paid for. Eco is the strictest case: it allows off and nothing else.
Model
Effort
futurex agent "add rate limiting" --tier fx-fast --effort highok — Fast runs at effort high, output capped at 16,000 tokens
The default effort for Fast is off. Try Fast at max to watch the check reject a combination.
The check runs in the CLI before the request leaves your machine, and again on the server, so a flag can never buy more reasoning than the model allows.
Output-token caps
Each model has a hard ceiling on the tokens it can write back: 64,000 for Pro, 32,000 for Light, 16,000 for Fast, 8,192 for Eco. On a reasoning run the thinking and the answer share that budget, which is why Pro — the model that thinks the most — gets the largest ceiling.
--max-tokens lowers the ceiling for a run. It cannot raise it: a value above the model’s cap is clamped back down to the cap.
futurex ask "summarize this module" --max-tokens 2000What it costs
Fast is $1 per 1M tokens. Pro is $2 per 1M tokens. Billing is usage-based, there is no per-seat subscription, and new accounts get a free first month with no card required.
Caps are ceilings, not charges. You pay for the tokens a run actually uses, so a two-line answer from Pro is billed as two lines — the 64,000-token ceiling costs nothing until you use it.
Full terms, including how the studio is quoted, are on pricing. For what the models remember between runs, see project memory.