Report · updated 2026-08-28 · always live
The same AI model can cost you up to 14× more — depending on where you run it.
An analysis of 1,502 models from 86 vendors across 38 hosting providers, priced in USD per 1M tokens (blended 3:1 input:output). Every figure is computed live from the Verna One model index — so this page updates itself.
Key findings
- Identical models, wildly different bills. Of the 137 models available from more than one provider, 65 vary by more than 2×, 12 by more than 5×, and 2 by more than 10×.
- The widest gap is 14×. DeepSeek V3.2 — the same model — is hosted by 16 providers, from $0.241/1M at the low end.
- Open weights are where the money leaks. Only 135 of 1,502 models (9%) ship open weights — but because anyone can host them, that is where nearly all the price arbitrage lives.
- No single cheapest provider. OpenRouter (default route) is the cheapest option for 91 models; a different provider wins for the rest.
- Some prices can't be taken at face value. For 3 models, first-party sources disagree with the cheapest host — a single quoted price is not the market price.
The same model, 14× the price
When a model has open weights, several providers host the exact same weights and price them independently. That is not a rounding difference — it is a different budget:
| Model | Price spread | Hosts | Cheapest |
|---|---|---|---|
| DeepSeek V3.2 | 14.0× | 16 | StreamLake — $0.241/1M |
| Llama 3.2 1B Instruct | 11.3× | 3 | DeepInfra — $0.0063/1M |
| Qwen2.5 Coder 32B Instruct | 9.0× | 2 | DeepInfra — $0.083/1M |
| Llama 3.1 8B Instruct | 8.8× | 6 | DeepInfra — $0.025/1M |
| Gemma 4 31B | 7.6× | 24 | OpenInference — $0.147/1M |
| MythoMax 13B | 7.5× | 8 | OpenRouter (default route) — $0.060/1M |
| Nano Banana Pro (Gemini 3 Pro Image) | 7.0× | 2 | OpenRouter (default route) — $4.50/1M |
| gpt-oss-120b | 6.9× | 24 | CoreWeave — $0.065/1M |
If you deployed on your provider's default and never re-checked, you are very likely paying multiples of the going rate for the identical output.
Who's actually cheapest
Across every model where we can compare, the cheapest option is:
| Provider | Models it's cheapest for |
|---|---|
| OpenRouter (default route) | 91 |
| OpenAI | 43 |
| DeepInfra | 35 |
| Alibaba | 23 |
| Anthropic | 16 |
| Mistral | 12 |
The pattern: closed models are usually cheapest from the vendor itself; open models are cheapest from specialist hosts or an aggregator route. There is no single "cheapest provider" — it is per model.
Best value right now
Quality per dollar (capability index ÷ price), from the live index:
| Model | Tier | Cheapest |
|---|---|---|
| Phi 4 | mid | DeepInfra — $0.087/1M |
| MiMo-V2.5 | frontier | Xiaomi — $0.175/1M |
| Granite 4.0 Micro | economy | OpenRouter (default route) — $0.041/1M |
| GPT-5.6 Luna | frontier | OpenAI — $0.225/1M |
| Nova Lite 1.0 | mid | OpenRouter (default route) — $0.105/1M |
| Mistral Small 3 | economy | OpenRouter (default route) — $0.058/1M |
What this means
- For engineering leaders: your model choice and your provider choice are two separate cost decisions. The second is worth 2–14× and almost nobody owns it.
- For finance & boards: AI cost scales with usage, so a wrong default compounds. Cost-per-1M by model and provider is basic unit-economics hygiene — and a hedge against lock-in.
- For everyone: re-check regularly. Hosts add, drop and reprice models constantly.
Check your own models
The full, live data — every provider for every model — is free, no signup.
Open the model price index →Methodology
Prices are normalised to USD per 1M tokens, blended 3:1 input:output, sourced from provider APIs and pricing pages and verified with dates. "Spread" is the ratio of the most- to least-expensive host of the same model, after reducing each provider to its own best price. Figures are as of 2026-08-28 and regenerate with the index. Full data: verna.one/models.
Frequently asked
How much can the same AI model vary in price?
For open-weight models hosted by several providers, the cost per million tokens can differ by as much as 14× for the identical model. Across the 137 multi-hosted models in the index, 65 vary by more than 2×, 12 by more than 5×, and 2 by more than 10×.
Which provider is cheapest?
There is no single cheapest provider — it is per model. Closed models are usually cheapest from the vendor itself; open-weight models are cheapest from specialist hosts or via an aggregator route. The live index shows the cheapest option for each model.
Is the data free and current?
Yes. The full index is free at verna.one/models, covers 1,502 models across 86 vendors and 38 hosting providers, and this report is regenerated from it — figures shown are as of 2026-08-28.