Review pricing formulas for all 22 models against actual provider rates

[SecurityQQ]

Context

The dynamic pricing system (PR #209) added ProviderPricing formulas to all 22 model definitions. During implementation, we already found one inaccuracy — Kling's formula ignored the generate_audio flag ($0.112/sec vs $0.14/sec), which was fixed in 37b3875.

There may be similar issues with other models. Each formula should be verified against the current provider pricing page.

What needs to be done

For each model definition in src/definitions/models/*.ts, verify:

1. Rate accuracy — does the per-unit cost match the provider's current pricing?
2. Parameter coverage — are all price-affecting parameters accounted for? (e.g. resolution tiers, audio toggles, model variants)
3. Bounds correctness — do minUsd and maxUsd reflect the actual min/max possible costs?
4. Description clarity — does the description string accurately describe the pricing without mentioning features that aren't modeled?

Models to review

Model	Provider	Pricing type	Verify against
kling	fal	per-second + audio toggle	fal.ai/kling-video
wan	fal	per-second	fal.ai/wan
reve	fal	per-second	fal.ai pricing page
ltx-a2v	fal	per-second	fal.ai pricing page
veed-fabric	fal	per-second	fal.ai pricing page
omnihuman	fal	per-second	fal.ai pricing page
seedance	piapi	per-second	piapi pricing
flux	fal	per-image	fal.ai pricing page
flux-schnell	fal	per-image (flat)	fal.ai pricing page
nano-banana-2	fal	per-image (varies by mode)	fal.ai pricing page
nano-banana-pro	fal	per-image	fal.ai pricing page
phota	fal	per-image (varies by variant)	fal.ai pricing page
qwen-image-2	fal	per-image	fal.ai pricing page
recraft-v4	fal	per-image	fal.ai pricing page
soul	higgsfield	per-image	higgsfield pricing
llama	fal	per-image	fal.ai pricing page
elevenlabs-tts	elevenlabs	per-character	elevenlabs.io/pricing
heygen-avatar	heygen	per-second	heygen pricing
sonauto	elevenlabs	flat	elevenlabs pricing
whisper	groq, fireworks	per-minute	groq/fireworks pricing

Known gaps

• Some models have multiple provider variants with different pricing tiers (e.g. Kling v2, v2.5, v2.6, v3, O3 all map to the same formula). These may actually have different per-second rates.
• The gateway maps ~50 model IDs to SDK model definitions in gateway/src/services/pricing/dynamic.ts. If provider rates differ per variant, we may need per-variant pricing in the SDK.
• HeyGen pricing uses an estimated $0.10/sec and speech-rate heuristic (12.5 chars/sec) — needs verification against actual HeyGen invoices.

Related

• SDK PR: feat: provider cost formulas and render pricing callback #209
• Gateway PR: vargHQ/gateway#101
• Render PR: vargHQ/render#72