Video API batch budget: cost per usable clip and retry planning

Budget for the clips you can use

Divide the pilot’s total recorded generation charges by the number of clips accepted for use. This is the observed cost per usable clip. Keep cost per attempt as a separate measure: one describes the generation bill, while the other helps you plan the deliverable.

A completed generation can still miss the brief because of composition, motion, timing or product details. Define acceptance rules before testing. Otherwise, reviewers can judge the same batch differently and produce incompatible estimates.

If no clips are accepted, cost per usable clip is undefined. Revise the prompt, inputs or model choice and test again before projecting a larger batch.

Fix the settings before comparing prices

Record the exact model ID, endpoint, duration, resolution and reference inputs. Include output count and audio settings when the model exposes them. Change one setting at a time when investigating differences in result or cost.

Read the billing unit. Video generation may be priced by tokens, generated duration, request or another unit. An estimated price for one short clip is not a flat rate for every request. On EasyAI, inspect the scenario assumptions on the model page and confirm your account rate in the console.

Compare candidates on the same creative task and show each model’s supported settings. If one cannot produce the requested duration or resolution, record the mismatch rather than treating the prices as equivalent.

Use a pilot to estimate the batch

Consider a hypothetical pilot of 20 attempts with $12.00 in total recorded charges. If 12 clips meet the brief, acceptance is 60%, average cost per attempt is $0.60 and observed cost per usable clip is $1.00. These numbers are illustrative, not EasyAI rates or a measured model benchmark.

At the same average cost and acceptance rate, 100 usable clips gives a first estimate of $100.00. A whole-attempt plan rounds 100 / 0.60 up to 167 attempts, or $100.20 at $0.60 each. Neither calculation guarantees 100 accepted clips; acceptance varies by prompt and source asset.

Test a less favorable assumption too. At 40% acceptance and $0.60 per attempt, 100 usable clips implies 250 attempts and $150.00. Substitute your own pilot results. Keep editing, storage and delivery expenses in separate budget rows.

Model ID and rate checked on:
Endpoint and billing unit:
Duration / resolution / reference inputs:
Acceptance rules:

Pilot attempts:
Pilot total recorded charges:
Pilot accepted clips:
Average cost per attempt = charges / attempts
Observed cost per usable clip = charges / accepted clips

Target usable clips:
Expected acceptance rate:
Planned attempts = round up(target / acceptance rate)
Estimated budget = planned attempts * average cost per attempt
Maximum approved spend:
Conditions for pausing the batch:

Track the task before submitting another attempt

If the chosen endpoint returns an asynchronous task ID, store it with your own job ID, model and settings. Use that endpoint’s documented status and result operations to determine the final outcome.

A network timeout does not establish that generation failed. Check the existing task where the API supports it before submitting another request. If its state is unknown, record that uncertainty and investigate; duplicate submissions can create more work and charges.

fal’s queue documentation illustrates separate submission, status and result operations. Its endpoints, state names, webhooks and retry guarantees belong to fal. For EasyAI, verify the workflow supported by the selected model; do not send fal-specific requests to an EasyAI endpoint.

Set stop conditions in your application

Submit a small batch first. Track pending jobs as well as completed charges so work already in progress remains visible when deciding whether to submit more. Reserve a conservative estimate for jobs whose final charge is unknown until they are reconciled.

Choose a maximum spend and attempt limit. Pause new submissions when a threshold is reached, error rates rise or acceptance falls below the budget assumption. These are controls for your own worker; this guide does not claim that EasyAI enforces an automatic hard cap for every endpoint.

Give each attempt an identifier and link it to its parent creative task. Record why a result was rejected and whether the next attempt changed the prompt or settings. This helps distinguish a model-quality issue from a broken integration or an unsuitable brief.

Bring measured requirements to a volume discussion

Share the model, expected usable clips per month, pilot acceptance rate, generation settings and purchasing period. If billing uses tokens or duration, include those measurements. They provide a better basis for a quote than a clip count alone.

Begin with the public scenario estimate and a small paid test. Confirm any enterprise rate, eligible models and purchasing terms before applying it to your forecast. Revisit the budget when a model version or major setting changes.