AI Media Cost Framework

Portal: https://projects.jbnx.io/framework/ai-media-cost-framework Markdown for agents: https://projects.jbnx.io/framework/ai-media-cost-framework?format=md Source: AI_MEDIA_COST_FRAMEWORK.md · dated examples in AI_MEDIA_COST_RATES.json Updated: 2026-10-25 · version 3

Choose the lowest effective cost that passes the required realism, identity, motion and audio checks. A cheap rejected render is waste. A more expensive model is justified only when it measurably improves the accepted result.

When to use

Use for TTS and voice-agent budgets, consistent fictional presenters, narration and lip-sync, image-to-video production, API-versus-subscription comparisons, and provider routing. Applies to providers such as xAI, ElevenLabs, HeyGen, Google, OpenAI, Kling, MiniMax, FLUX and LTX, including authorized hosting platforms.

1. Define the deliverable before selecting a model

Record the finished duration, aspect ratio, delivery resolution, frame rate, required realism, allowed cuts, speech/lip-sync needs, deadline, intended use and hard spending cap. Separate the cost of reusable reference assets from the marginal cost of each finished video. Set explicit acceptance checks for face, body, hands, feet, motion, voice, lip-sync and frame continuity.

Use the smallest architecture that satisfies the brief:

Architecture	Default design	What to compare
Narration over images or footage	One pinned stock TTS voice and edited visuals	TTS, editing and asset reuse; no presenter engine unless needed
Talking presenter	Locked face/body references, pinned voice track, compatible lip-sync or avatar endpoint	Identity, mouth timing, speech quality, framing and complete rendered-video cost
Cinematic or physical action	Approved scene first frame and a reference-capable video endpoint	Motion, identity through turns, anatomy and usable seconds
Realtime voice agent	Native speech-to-speech or a measured STT/LLM/TTS stack	Total metered session cost, reasoning/tool fees, latency and reliability

Do not compare a silent motion draft with a lip-synced presenter as if they were equivalent products.

2. Lock identity and voice with reusable assets

Keep one versioned character identity authority: neutral face close-up, neutral full-body anchor, supporting angles, expression references, hairstyle and designed body proportions. Use feet flat and a relaxed neutral stance for the baseline; keep stylized poses separate. Record fictional height as a design target, never as a measurement recovered from an image.

Attach the exact anchor or matching approved scene image to each supported request. Preserve reference file hashes and exact endpoint/model revision. A seed is useful for replay only when the provider supports it; it does not lock identity across prompts, resolutions, models, versions or scenes. Independent generated angles remain interpretations, not a calibrated 3D scan.

For speech, pin provider, stock voice_id, model/version, language, accent, speaking rate, relevant settings and a reference audio sample. Verify the voice's commercial-use terms. Generate the approved master script audio once, cache it, then split the same track at edit boundaries. Regenerate only changed lines with matched context when a full retake is unnecessary. A voice ID alone does not prevent changes after provider model updates.

Use the pinned track to drive compatible presenter/lip-sync endpoints. Native generated dialogue is acceptable for a one-off clip if continuity is not required; for a recurring character, verify actual voice-reference support or replace the generated speech. Do not promise that a text description such as “warm British voice” establishes cross-clip voice identity.

3. Follow a gated draft ladder

1. Plan: complete the shot list, script, camera, continuity and cost estimate before generation. Reuse existing approved references. Do not regenerate the whole reference pack for a small correction.
2. Approve the still: produce or reuse the exact scene first frame. Check identity, proportions, composition, wardrobe and lighting before paying for motion. For image-to-video, give a short motion/camera instruction rather than redundantly re-describing the image.
3. Test motion: use the shortest supported clip at the cheapest resolution that still exposes identity and anatomy failures. Keep the same model family where practical. A different draft engine tests blocking, not final-engine identity or quality.
4. Review: inspect representative frames and playback for face/body drift, hands/feet, hair, cuts, speech and lip-sync. Log acceptance, rejection reason and usable seconds. Make one targeted correction. Default maximum: two submitted attempts per shot, then stop and revise the plan.
5. Finish only accepted work: keep an accepted draft if it already meets delivery requirements. Otherwise render one final pass at the required resolution. Changing resolution, model or endpoint requires another quality check; the draft does not guarantee the final result.
6. Edit and deliver: assemble clips, reuse the locked audio, mix once and export once. Upscale only accepted footage when the tested result meets the brief and total cost is lower. Upscaling cannot repair facial drift, anatomy, motion or missing detail. Do not default to 4K, frame interpolation or repeated enhancement passes.

4. Compare routes using complete unit economics

Compare first-party APIs, authorized hosted APIs, existing subscriptions and self-hosting under the same deliverable and quality criteria. Direct APIs are not always cheaper; hosted providers may have different models, discounts, infrastructure or minimums. Do not assume a universal aggregator markup.

Separate consumer subscription access from developer API entitlements. Confirm current quotas, rate limits, commercial rights, automation permissions, credit expiry and overages. Use the subscription only through a supported authorized workflow; do not assume a web plan includes API credits or bypass a native-login requirement.

Self-hosted open weights have no provider inference charge only when the relevant license allows it. Include GPU rental or depreciation, idle time, electricity, storage, setup, maintenance and operator time. Marginal API charge of zero is not total cost of zero.

Calculate:

project cost = reference setup + voice + all submitted video attempts
             + lip-sync + editing/upscale compute + storage/egress
             + applicable subscription allocation + taxes/fees

effective cost per accepted second = actual project cost / accepted seconds

forecast = remaining attempts × verified billable units × current rate
         + remaining fixed and finishing costs + explicit contingency

Use exact script character counts for character-priced TTS. Convert to cost per audio minute only with a stated character-count and spoken-duration assumption. Include minimum clip duration, billed rounding, request count, resolution/audio multipliers, aspect constraints, rejected-but-billed work and paid retries. Show low/base/high estimates with their attempt and acceptance assumptions; do not present an estimated acceptance rate as measured performance.

5. Keep billing, replay and credentials under control

• Verify rates at decision time. A usable rate record needs provider, exact endpoint/model/tier, currency, region, unit, amount, resolution, audio mode, minimum/rounding rules, date, source URL and any discount expiry. Unknown values remain unknown. An unverified rate cannot authorize a paid run.
• Reserve the worst-case cost of each allowed request against the remaining job cap before submission. If the estimate, wallet access or billing unit is unresolved, keep the job pending; do not top up, change plans or silently switch endpoints.
• Cache by input hashes, prompt, model/revision and generation settings. Persist provider request ID immediately. Poll the same request after an uncertain response; never blindly re-submit because a client timed out. If provider idempotency is undocumented, use a durable local job lock and reconcile before retrying.
• Count each submitted attempt. Distinguish estimated, reserved, provider-reported and reconciled charges. A failed or cancelled request is free only when the provider documents it or the receipt confirms it. Stop at the attempt or spending cap.
• For realtime agents, verify whether billing follows connected time, active audio, tokens or other units. Use appropriate VAD, idle timeout or push-to-talk for the product's needs. Do not disable turn detection blindly or assume silence is unbilled.
• Count pass-through LLM/tool fees, concurrency limits, burst/overage charges and minimums only when documented for the chosen plan. Disable unused audio generation when supported and actually cheaper.
• Keep provider keys server-side in existing credential stores. Never copy secrets into prompts, public documents, client code, manifests or logs. Record credential locations and authenticated capability checks, not values. Keep safety checks enabled.

6. Required recommendation output

Every recommendation must contain:

1. Recommended stack: exact currently available provider, model/endpoint and billing tier for the anchor, stock voice, draft motion and production video. Identify an existing reusable asset instead of proposing paid regeneration. Mark unavailable components and untested quality honestly.
2. Unit economics table: voice cost per minute under an explicit script assumption, and/or total cost per 30/60-second finished deliverable. Compare the locked/reused path with a genuinely comparable premium or repeated-render path. Include setup, drafts, accepted/failed attempts, finishing and taxes/unknown exclusions; show assumptions and no unsupported savings percentage.
3. Architecture guardrails: reference hashes, stock voice ID and model, exact video endpoint, framing/resolution/audio, attempt and spending caps, cached assets, request-ID reconciliation, QA criteria and fallback conditions. No fallback may exceed authorization or budget.
4. Execution evidence: verified rate sources and check date, planned versus actual spend, request IDs, assets used, usable seconds and quality result. If no video has run, report “planned” rather than “validated.”

Rate examples and application

AI_MEDIA_COST_RATES.json is a dated evidence snapshot, not permanent pricing policy. Refresh it before a paid decision. The original supplied examples are not default tariffs: xAI's published TTS rate checked on 2026-10-25 is $15 per million characters, and ElevenLabs offers USD-priced API TTS. Provider products and billing tiers must not be mixed.

For Kayla model 001, apply this framework to the reference pack, stock voice sample, shortest four motion tests and shot ledger at https://jbnx.io/design/m/001. Use an explicit small test budget; production quality remains pending until actual clips pass inspection.

Applied Kayla benchmark · 2026-10-25

Six provider requests used 266 credits ($1.33 at listed rates; cash unreconciled) within a $1.65 allowance, including the rejected turn. Four curated motion baselines (18.6667s) and a 7.4667s voice proof passed scoped checks. Voice waveform and transcript matched; full-clip SyncNet passed 177 windows (−40ms offset; confidence 7.5897). Human listening was not performed. These short references do not guarantee future clips. Setup and local compute costs remain unknown. Evidence: https://jbnx.io/design/m/001.