Credit Budget & Model Selection for Production Teams
ForgeEcho credit planning guide—when to use nano-banana-fast vs nano-banana-2 vs 4K, Chat/Image/Voice cost tables, and weekly budgets for Amazon listings, Meta ads, and faceless shorts.
Why credit planning beats “generate until it looks good”
ForgeEcho bills by action: AI Chat 0.5 credits/reply, AI Image 3 / 4 / 8 by resolution (1K / 2K / 4K), AI Voice 1 credit per 500 characters. Without a budget, teams burn the month’s plan on 4K re-rolls and never finish the SKU list.
This guide turns the credit meter into a production tool: pick the right model and resolution for each stage, and reserve headroom for winners—not exploration.
Unit costs (current product)
| Module | Unit | Cost |
|---|---|---|
| AI Chat | 1 reply | 0.5 |
| AI Image | 1 image @ 1K | 3 |
| AI Image | 1 image @ 2K | 4 |
| AI Image | 1 image @ 4K | 8 |
| AI Voice | ≤500 characters | 1 |
| AI Voice | 501–1000 characters | 2 |
Image model choice does not change the resolution price table above—nano-banana-fast at 1K still costs 3 credits. Model choice changes quality, reference strength, and how many re-rolls you need.
Model selection matrix
| Stage | Model | Resolution | Role |
|---|---|---|---|
| Prompt / script draft | Chat (Gemini family) | — | Structure constraints before any image spend |
| Direction filter | nano-banana-fast | 1K | 4–6 cheap variants; pick composition |
| Everyday finals | nano-banana-2 | 2K | Listings, feed ads, social covers |
| Strong reference edit | nano-banana-2 / nano-banana-2-cl | 1K→2K | Product shape / face likeness locked |
| Hero / print / large display | nano-banana-2 or nano-banana-pro | 4K | Only after a 2K winner exists |
| Voice tone test | ElevenLabs Turbo 2.5 | short clip | 10–20s before full read |
Rule: Never open at 4K for a new prompt family. Filter composition at 1K, lock the prompt, then upscale once.
Budget templates (copy into your ops sheet)
A) 20-SKU Amazon main refresh
Goal: white-background mains, one final per SKU.
| Step | Count | Unit cost | Subtotal |
|---|---|---|---|
| Chat field-swap / tweak | 20 × 0.5 | 0.5 | 10 |
| 1K screen × 2 per SKU | 40 × 3 | 3 | 120 |
| 2K final × 1 per SKU | 20 × 4 | 4 | 80 |
| Total | ~210 |
If you skip 1K and re-roll 2K × 3 per SKU, expect ~240+ with worse consistency. See Ecommerce Image Optimization.
B) Meta/TikTok creative fatigue pack (1 week)
Goal: 5 templates × 3 variants + 5 short voice hooks.
| Step | Count | Subtotal (approx.) |
|---|---|---|
| Chat: template variants | ~15 replies | 7.5 |
| Image: 1K screen | 5 × 4 × 3 = 60 | 180 |
| Image: 2K promote top 30% | ~15 | 60 |
| Voice: 15–40 word hooks | 5 × 1 | 5 |
| Total | ~250 |
Promote only assets that beat account CTR average; retire the bottom 20% monthly (Social Media Batch Creative).
C) Faceless Shorts batch (3 clips)
| Step | Subtotal (approx.) |
|---|---|
| Chat scripts + cover prompts | 3–5 |
| Image 1K filter + 2K finals (3 covers) | 30–40 |
| Voice samples + finals | 6–10 |
| Total inside ForgeEcho | ~40–55 |
Assembly (CapCut / Premiere) is outside credits. Full steps: Faceless Short Video Pipeline.
Decision tree: where is the money going?
New brief?
├─ No gold template → Chat × 2–3 first (cheap)
├─ Composition unknown → 1K × 4–6, never 4K
├─ Winner locked → 2K × 1–2
└─ Platform demands print/zoom → 4K × 1 only
Failing at 2K?
├─ Structure wrong → back to Chat (one lever)
├─ Shape/identity drift → add reference, stay at 1K until stable
└─ Soft detail only → one 4K pass on the winnerSoft vs hard budget controls
| Control | How to run it |
|---|---|
| Soft weekly cap | e.g. 300 credits/week for “listings + ads”; stop new exploration at 80% |
| Hard per-SKU cap | Max 12 credits/SKU before escalating to a human review |
| Voice sample rule | Never generate full 90s scripts until a 15s sample passes |
| Library promotion | Only save prompts that shipped or beat CTR—noise templates cost future credits |
Pairing with Prompt Library
Credits stretch when winners are reusable. Tag library entries with:
channel(amazon-main / meta-4x5 / tiktok-9x16)model+resolutioncredit-band(e.g. “~10.5 / SKU”)
Next month you change fields, not structure—see Brand Visual System.
Common mistakes
- Opening every idea at 4K “to save time”
- Using Chat as a long essay writer instead of a constraint layer
- Batching voiceovers before locking the visual hook still
- Mixing compliance mains and lifestyle ads in one credit bucket with no labels
FAQ
Does switching from fast → 2 change the credit price?
No. Resolution sets image credits. Model choice changes how many generations you need.
Should agencies bill clients in credits?
Better to bill in deliverables (e.g. 20 mains + 6 ad stills) and keep an internal credit estimate like the tables above.
Where do I diagnose bad outputs without more spend?
Use AI Image Troubleshooting before another re-roll.