Pro guide · model control

Asking for the model you want

You already know the catalog. This page is about the other half — making sure the model you named is the one that ran, and knowing what to do when it wasn't. It's all plain conversation; there's nothing to configure and no mode to switch on.

Just name it

Mention the model in your request and it gets used. No special syntax.

Make a brass compass on weathered oak — use flux-schnell. Animate this frame with kling-o3-i2v, five seconds.

If you're not sure of a name, ask before you render — it's free and instant:

Which video models can I use, and what do they cost?

And when you don't care which model runs, say nothing about models. You'll get the default fast path, with no model chatter in the reply.

You'll be told what actually ran

Whenever you name a model, the reply ends by telling you which one did the work. Three things you might see:

It used your pick

Generated image — a brass compass on weathered oak · ~$0.003 · ran on flux-schnell

Nothing to do. That's a confirmation, not a warning.

It used something else, and says why

Generated image — an origami crane · ~$0.03 · ran on flux-dev, not sora-2 (not a registered model name)

The reason is always there. The next section covers what each one means.

You didn't name anything

Generated image — a ripe pear on a white plate · ~$0.03

No mention of models at all. Naming one is what opts you into the extra detail.

You can always just ask. If a reply ever leaves you unsure, say "what did that actually run on?" — the answer is always available, including for long renders you came back to later.

Four reasons your pick might not run

Three of these are the network protecting you. Only the first one is your typo.

What you'll be toldWhat it meansWhat to say next
not a registered model name That name isn't live here, so the closest available model was used instead. "List the video models" — then use the exact name from that list.
that model is temporarily disabled It's known to be broken and charges you even when it fails, so a working replacement ran. Nothing. You were kept from paying for a dead model.
no source frame You named an image-to-video model but didn't give it a starting image. That request gets rejected, so its text-to-video sibling ran. Give it a starting frame — "animate this image with …" — and your pick will hold.
the model you asked for failed Your pick went first, hit an error it could recover from, and a similar model was tried so you'd still get something. "Try that again", or demand exactness (next section) to see the real failure.

Near-miss names are the one real trap

An unfamiliar name isn't rejected — it's matched to the closest live model. That's deliberate: it rescues typos and keeps long pipelines from dying on a spelling slip. The cost is that a wrong name gets you a wrong model rather than an error. Real cases:

You ask forYou getWhy it stings
sora-2flux-devYou wanted video and got a still image.
runway-gen4flux-devSame — no video model matched that name.
wan-2.5wan-22-i2v-loraA shared name prefix lands you on an unrelated image-to-video style model.
Two seconds of habit prevents every case above. Ask for the model list rather than typing a name from memory, and glance at the "ran on …" line on your first render of a session. If it says what you expected, the rest of the session is safe.

When you want it exact or not at all

Sometimes a substitute is worse than nothing — you're comparing two models, or a client signed off on one specifically. Say so, plainly:

Use exactly kling-o3-i2v. Don't substitute — fail instead. Render this on flux-pro only. No fallbacks.

Phrased like that, your name goes straight through untouched. If that model isn't available, you get a clear failure instead of a quiet stand-in.

Know the trade-off before you ask for it. Exact mode is deliberately bare. It skips the conveniences the normal path gives you: a voice or trained style you've attached to a project, your brand colours and fonts, automatic quality checks, and saving the result into a project's scene list. Your spending is still tracked and still counts against your cap, and you're still told the cost. Everything else, you handle.

So: reach for exactness when identity of the model matters more than convenience, and stay on the normal path the rest of the time.

Things worth saying

When you want to…Say something like
Use a specific model"…using kling-o3-i2v"
Insist on it, no substitutes"use exactly kling-o3-i2v — fail rather than substitute"
See what's available"which image models can I use?"
Check a price first"what does kling-o3-i2v cost for 5 seconds?"
Understand one model properly"tell me about kling-o3-i2v — inputs, limits, price"
Go fast, or go premium, without picking"make it fast" · "do this one in high quality"
Keep an image-to-video model on your promptgive it a starting frame — "animate this image…"
Know what a finished render used"what did that run on?"
See your spending"what have I spent today?"
Put a ceiling on it"cap my spend at $5 a day"

If you're scripting against the raw response

Everything above is also machine-readable. The reply carries the model that ran, the model you asked for, and — only when those differ — a note holding both names, the reason, and the underlying error. Long-running jobs report the same three when you check back on them.