tokens used × the rates below = your charge.
Tokens in one minute
AI models read and write text in tokens (roughly ¾ of a word each). Rates are quoted per million tokens (MTok) — the industry standard. Four kinds of work are metered:Rate tables
The standard per-model rates:Atomos — most capable model
Atomos is 1.67× Muranyi 3.5 on output, input and cache reads. Blended across real usage
it works out to ~132 credits per million AI tokens. Typical prompt: 275–3,600
credits.
Muranyi 3.5 — standard model
Blended across real usage, Muranyi 3.5 works out to ~95 credits per million AI tokens,
because the overwhelming majority of tokens in a typical prompt are cheap cache reads.
Typical prompt: 45–550 credits.
Muranyi Fast — fast & cheap edits
Muranyi Fast is for iterating on an existing game — it can’t start new games. Its
output rate is 2.5× cheaper than Muranyi 3.5 and its fresh-input rate a third cheaper,
but note that its cache-read rate is identical (cached context is already as cheap as it
gets). Since cache reads are most of any prompt, the blended difference is much smaller
than the headline rates suggest: ~110 credits per million AI tokens. Where Fast really
saves you credits is that its prompts are small — a focused edit rather than a whole build.
Typical edit: 35–65 credits.
Long prompts on Muranyi Fast cost double. Above 200,000 tokens in a single prompt, the
underlying model doubles every rate — output 1,848, input 616, cache write 616, cache read
92 credits/MTok. Mature games can cross that line on a single edit. Muranyi 3.5 and Atomos
have no such threshold, which is worth knowing if you are iterating on a large project.
Why cache reads dominate
The key insight for understanding any charge: on every prompt, the AI re-reads your
whole project to know what it’s working with. On a mature game that context is often
1–3 million tokens — but nearly all of it hits the cached rate (46 credits/MTok on
Muranyi), not the fresh-input rate. That’s why a prompt on a big game can process millions
of tokens yet cost a couple of hundred credits.
A worked example
A medium edit on an established Muranyi 3.5 game might meter like this:The same rates for every plan
Each model has its own published rates, above. What does not vary is who you are: there is no per-plan pricing, no volume curve, and no surcharge for picking the capable model. A prompt costs the same on a free account as on Ultra — plans differ in how many credits they grant, never in what a prompt costs. Picking a cheaper model simply makes your credits go further on the same work, which is why the model picker is worth using rather than leaving on the default. Credits are purchased at $1 = 100 credits, with bonus credits on the larger top-up bundles (see Plans & top-ups).Where do my credits go?
Audit any charge — per-prompt tokens, cache hit rate, and balance after