API credits are the currency you spend on the CoreReflex platform, and the model behind them is deliberately lopsided in your favor: you are only charged when the system actually generates media. Planning a film, scoring its shots, and editing the timeline cost nothing, so you can design, validate, and price an entire render before a single credit ever moves. This guide breaks down exactly what consumes API credits, what is free by design, and the patterns that let you put a hard ceiling on cost.
What an API credit actually buys
A credit maps to one thing: generation work. The moment a model on the stack turns a prompt and a set of parameters into new pixels, audio, or video, that call is metered. A Veo or Kling video shot, an Imagen still, a Lyria music bed, a TTS voiceover read, a super-resolution upscale pass. Each is a discrete, billable generation.
Everything that happens around generation is structural, not generative, and it does not touch your balance. The agentic Director can plan a film, board every shot, score a draft, and let you rearrange the timeline as many times as you like. You spend a credit only when you ask a model to render something new.
That distinction matters because most of the iteration in a real project is structural. You will reorder shots, rewrite a prompt, change a camera move, tighten a cut, and swap a voice far more often than you will commit to a fresh generation. On a metered-per-generation model, all of that planning friction is free.
Free by design: plan, score, and edit
Three categories of work never cost credits, and they happen to be the three you do most:
- Planning. The Director loop runs PLAN, where it decomposes your sentence into a shot list, each shot carrying a role, a camera move, and a prompt. Boarding, re-boarding, and refining that plan is free.
- Scoring. Every shot is scored against concrete checks (prompt match, sharpness, motion coherence, on-screen text legibility, on-brand, claims-risk). Reading those scores, inspecting why a shot failed, and deciding what to regenerate costs nothing.
- Editing. Trimming, reordering, retiming, swapping clips, and adjusting the manifest on the timeline are all structural operations on data, not new generations.
Why free planning changes how you budget
Because planning is free, you can treat it as a quoting tool. Board the full film, look at how many shots the Director proposes, and you know the upper bound on generation before you commit. If the plan calls for twelve shots, twelve generations is your worst case, and selective regeneration usually keeps you well under that, you are pricing the render from a finished blueprint, not guessing and hoping.
This is also why the JSON engine that lets you edit video as data pairs so well with the credit model: you can manipulate the entire structure of a project programmatically, validate it, and only spend when the structure is final. If you are new to the API surface, the developer quickstart walks through authenticating and issuing your first plan call.
What spends credits
Generation is the line item. In practice that means:
- Producing a shot. When the Director moves from PLAN to PRODUCE, each shot it generates is one charge. This is where the bulk of spend on a typical film lands.
- Selective regeneration. When a shot fails the quality gate, the system regenerates that shot, not the whole film. A regeneration is still a generation, so it costs, but the unit is one failed shot, not a full restart. Selective regeneration is the difference between paying for one retry and paying to rebuild everything.
- Audio generation. A Lyria music bed or a TTS voiceover read is generated media and is metered like any other shot.
- Upscaling. Running the native 4K/8K super-resolution seam over a clip is a generation pass and consumes credits.
- Images. Imagen stills produced for an Image Studio composition or a thumbnail are metered per image.
The useful mental model: if a model had to create something that did not exist before, it cost a credit. If you only rearranged or evaluated what already exists, it did not.
Cost-control patterns for the API
The credit model gives you natural levers. Here are the ones worth wiring into your integration from the start.
Quote from the plan before you generate
Issue the plan call, read back the shot count, and you have a hard ceiling. For a programmatic pipeline, gate the PRODUCE step behind a check: if the plan exceeds an expected shot budget, pause for review instead of generating. Because programmatic editor control through JSON exposes the full project structure, this gate is a few lines of logic, not a separate approval system.
Set a budget ceiling with Workflow governance
For scheduled or hands-off runs, the Workflow pillar adds budget and checkpoint governance on top of the credit model, you define a spend ceiling and checkpoints, and the pipeline stops rather than blowing past the limit. This is the right home for any Autopilot job a company runs unattended, the recipe enforces the budget so nobody has to babysit the run.
Reuse provenance instead of regenerating
Every generation carries a portable trace: the model, the prompt, the parameters, and the score. Because that trace is reproducible, you can replay a known-good generation rather than rolling the dice on a new one. When a shot already passed the gate, anchor on its provenance and keep it, spending a credit to regenerate something that already works is the most common avoidable cost.
Separate cheap structure from expensive pixels
Do all your prompt tuning, camera-move tweaking, and timeline edits before you generate. Tuning a camera move in the JSON shot spec is free; discovering you wanted a different move after a generation means paying twice. Front-load the free work.
Reading credit usage in a render job
Generation and final rendering are distinct stages. The Director generates the shots (metered), and the deterministic render-worker assembles the approved manifest into a finished file. Polling a job tells you where spend has already happened and what remains. The mechanics of watching a job through to completion are covered in polling render job status via the API, and the same trace that drives billing transparency is what makes a job reproducible: the same manifest renders the same cut.
If you want the higher-level walkthrough of turning a prompt into a finished film over the API, the call sequence that ultimately spends credits, see generating a film from one sentence. For the full reference, including current rate behavior and endpoints, the API documentation is the source of truth, and the pricing page lays out how credits map to plans. More developer-focused walkthroughs live in the developers section of the blog.
Frequently asked questions
What actions cost credits on the CoreReflex API?
Only generation, that includes producing a video shot, generating a still image, creating a music bed or a TTS voiceover, regenerating a failed shot, and running an upscale pass. Each of those asks a model to create new media, and each is metered as a discrete unit.
Is planning free if I never generate a shot?
Yes. You can run the PLAN step, board a complete shot list, score a draft, and edit the timeline indefinitely without spending a credit. Credits move only when you trigger generation. If you plan a film and never produce it, your balance is untouched.
Does a failed shot cost me twice?
A regeneration is a new generation, so it does carry a charge, but the system regenerates only the shot that failed the quality gate, not the entire film. That selective regeneration keeps the cost of a retry scoped to a single shot rather than a full restart.
Can I cap spend on an automated run?
Yes. Workflow recipes add budget and checkpoint governance, so a scheduled or Autopilot run halts at a ceiling you define instead of generating past it. Combined with quoting from the free plan, this gives unattended pipelines a hard upper bound.
Do edits and re-renders of the same manifest cost credits?
Editing the manifest is free. Re-rendering an already-generated manifest through the render-worker assembles existing approved media into a file, it does not re-run the generative models, so the expensive, credit-spending step is the original generation, not the assembly.
Build cost control in from the first call
The credit model rewards a simple discipline: do all the free work first. Plan the film, read the shot count, tune the structure, and lean on provenance, then spend credits once, on generation you have already validated. That is how teams keep AI video predictable instead of open-ended.
Ready to see it on a real project? Start free, no credit card required, board a film, and watch how far you get before a single credit moves.