Our dataset for July 2026 covers 58 verified price records across 9 providers and 6 production video models. Every number below was pulled directly from provider pricing pages and verified on July 30, 2026. This is not a review — it is a price-structure analysis.
The spread is 28x — and it hides inside a single output format
The spread from cheapest to most expensive AI video API call is 28x in normalized USD-per-second terms. The cheapest verified rate in our dataset is $0.025/s (Kling 2.1 at Kie.ai, per-clip pricing). The most expensive is $0.70/s (Sora 2 Pro at 1080p, official OpenAI API, source).
That spread exists within a single output format: both are delivering video at roughly 720p–1080p. The gap is not about quality tier alone. It reflects provider markup, billing scheme, and whether you understand what you are actually buying.
Six models dominate the API market right now:
| Model | Developer | Min $/s | Max $/s | Spread | Providers |
|---|---|---|---|---|---|
| Veo 3.1 | Google DeepMind | $0.03 | $0.60 | 20x | 3 |
| Kling 3.0 | Kuaishou | $0.084 | $0.42 | 5.0x | 5 |
| Seedance 2.0 | ByteDance | $0.052 | $0.682 | 13.1x | 5 |
| Sora 2 | OpenAI | $0.10 | $0.70 | 7x | 1 |
| Hailuo 2.3 | MiniMax | $0.027 | $0.067 | 2.5x | 1 |
| Wan 2.6 | Alibaba | $0.10 | $0.15 | 1.5x | 1 |
The Kling 3.0 Min ($0.084) reflects verified Kling 3.0 rates only. A cheaper $0.025/s point exists at Kie.ai, but it is a Kling 2.1 clip recorded as a proxy (no Kling 3.0 price is published there), so it is excluded from the Kling 3.0 spread above and treated separately below.
Veo 3.1 has the largest within-model spread of any model in the dataset: 20x between Google's own Lite tier (video-only, 720p, $0.03/s) and the Standard 4K+audio tier ($0.60/s). Both are from the same developer on the same official API (source).
Cheapest Verified Rate per Model at ~720p
These are the floor prices from our dataset for the most common production resolution, verified July 30, 2026:
| Model | Provider | Tier | $/s | Notes |
|---|---|---|---|---|
| Veo 3.1 | Google Vertex (official) | Lite | $0.03 | Video-only, 720p |
| Hailuo 2.3 | PiAPI | Fast | $0.027 | 768p, per-clip normalized |
| Kling 3.0 | fal.ai / WaveSpeedAI | Standard | $0.084 | 720p, no audio |
| Wan 2.6 | WaveSpeedAI | Standard | $0.10 | 720p |
| Sora 2 | OpenAI (official) | Standard | $0.10 | 720p |
| Seedance 2.0 | Novita.ai | Standard | $0.052 | 720p (proxy rate, reported confidence) |
| Seedance 2.0 | WaveSpeedAI | Standard | $0.10 | 720p, per-clip verified |
For Seedance, the Novita.ai rate ($0.052/s) is reported confidence, based on the Seedance 1.5 Pro published rate rather than a separately confirmed 2.0 price. The WaveSpeedAI rate ($0.10/s) is fully verified from their pricing page.
See /cheapest/seedance-2.0 and /cheapest/veo-3.1 for live-updated floor prices.
Official API vs Resellers: Google Flips the Script
The conventional wisdom is that official developer APIs are more expensive than resellers who aggregate and compete on price. Veo 3.1 breaks that assumption.
Google Vertex AI official pricing:
- Veo 3.1 Fast, 720p + audio: $0.10/s (source)
- Veo 3.1 Lite, 720p + audio: $0.05/s (official only)
- Veo 3.1 Lite, 720p, video-only: $0.03/s (official only)
Reseller pricing for the same Fast tier with audio:
fal.ai and WaveSpeed charge 50% more than Google's own API for Veo 3.1 Fast with audio. The Lite tier, Google's cheapest at $0.05/s with audio or $0.03/s video-only, has no reseller equivalent in our dataset. If you want Lite pricing, you go direct to Vertex AI.
This is an unusual market dynamic. It exists because resellers provide a convenience layer (one API key, unified billing, simpler onboarding), but the margin is visible and measurable.
The Token-Formula Trap: Seedance 4K vs 480p
BytePlus (the official Seedance API) bills by tokens, not by seconds. The formula: width × height × fps × seconds ÷ 1024 = tokens. Because cost scales with pixel count, every resolution step compounds. The table below applies that formula directly at a flat $0.014/1k-token rate. These are modeled projections, not separately verified BytePlus records — our dataset holds one verified BytePlus point ($0.1512/s at 720p), which sits below what the raw formula produces (likely a different effective token rate at that tier). Treat every row here as the formula's own arithmetic, not a quoted price:
| Resolution | Pixels vs 480p | Tokens/s | $/s (formula) |
|---|---|---|---|
| 480p (854×480) | 1.00x | ~9,600 | ~$0.13/s |
| 720p (1280×720) | 2.25x | ~21,600 | ~$0.30/s |
| 1080p (1920×1080) | 5.06x | ~48,600 | ~$0.68/s |
| 4K (3840×2160) | 20.23x | ~194,400 | ~$2.72/s |
Because the rate is flat per token, the dollar column scales exactly with pixel count: a 5-second 4K clip runs ~$13.60 on the formula versus ~$0.67 at 480p, a 20x gap that equals the pixel ratio. Note the mismatch with our one verified BytePlus record: the formula puts 720p at ~$0.30/s, but reapi.ai verified BytePlus at $0.1512/s at 720p — roughly half. The raw $0.014/1k formula lands about 2x higher than the verified point, so BytePlus almost certainly applies a lower effective token rate at this tier. Budget off the verified $0.1512/s, not the formula rows.
We can show the non-linearity with a same-provider pair (reported confidence, sourced to reapi.ai): on Replicate, Seedance 2.0 goes $0.08/s at 480p to $0.18/s at 720p, a 2.25x jump that matches the pixel ratio to two decimals (source). Per-second aggregators flatten the top end. On fal.ai, 720p is $0.3034/s and 1080p is $0.682/s (2.25x), softer than the token math because fal prices on its own grid. See /models/seedance-2.0 for the full resolution breakdown.
Audio Surcharges: Kling's Consistent 50% Penalty
Kling 3.0 applies a uniform 50% audio surcharge across all tiers and resolutions. This holds on both fal.ai and WaveSpeedAI:
| Tier / Resolution | No Audio | With Audio | Surcharge |
|---|---|---|---|
| Standard, 720p | $0.084/s | $0.126/s | +50% |
| Pro, 1080p | $0.112/s | $0.168/s | +50% |
Sources: fal.ai, WaveSpeedAI
If you do not need audio (and many production pipelines add audio in post), you can cut Kling costs by a third. Contrast this with Seedance 2.0 on fal.ai, which includes audio at no extra charge at the same per-second rate.
Veo 3.1 on Google Vertex charges explicitly for the audio flag: Standard 720p is $0.20/s video-only or $0.40/s with audio. If you only need muted clips, that halves the bill on the official API.
Sora 2: API Sunset September 24, 2026
OpenAI has confirmed the Sora 2 API is discontinued on September 24, 2026 (source). Current official pricing: $0.10/s (720p standard) and $0.70/s (1080p Pro). If you are building on Sora today, migration is not optional — it is a deadline.
The closest functional replacements, on a per-second cost basis:
- Veo 3.1 Fast via Google Vertex: $0.10/s with audio (same price, adds native audio)
- Kling 3.0 Standard via fal.ai or WaveSpeed: $0.084/s without audio, $0.126/s with audio
Neither is a drop-in replacement for instruction following or style, but both are at comparable price points. Use /calculator to run a migration cost comparison based on your current clip volume.
Market Structure: Who Holds What
Provider roles matter because they determine pricing floors and reliability:
- Official APIs (Google Vertex, OpenAI, BytePlus): Set the price floor for their own models. Currently Google undercuts its resellers on Veo 3.1.
- Aggregators (fal.ai, WaveSpeedAI, Replicate, Novita.ai): Resell multiple models through unified billing. Margin visible in dataset — fal charges 3x WaveSpeed for Seedance 2.0 at 720p standard ($0.3034/s vs $0.10/s).
- Relay providers (PiAPI, Kie.ai): Use host-your-account or relay models. Cheaper in some cases (Kie.ai Kling 2.1 at $0.025/s) but with different trust profiles.
Our /methodology page documents confidence levels for each record. The 58 records in this dataset are weighted: 52 carry "verified" confidence (rate directly published on a provider pricing page), 6 carry "reported" (cited in secondary sources cross-referencing official pricing — the Novita Seedance proxy, the Kie Kling 2.1 proxy, Replicate's Seedance rates, and the BytePlus token normalization among them).
What to Watch
Watch for a reseller Lite-tier response to Google's $0.03/s floor by Q4 2026: right now no aggregator in our dataset touches Veo's Lite pricing, and that gap is the clearest arbitrage on the board. GenRates normalizes every scheme (per-second, per-clip, per-token) to USD per second so you can compare on one axis. Check /models for live rates.
Prices are accurate as of July 30, 2026 and may change without notice. Check the live tracker for current rates. Some provider links on this site may be referral links.