If you call hosted LLMs through more than one provider, you already know the prices don't line up. What I didn't appreciate until I normalized them into one dataset is how much they don't line up.
I scraped the public pricing pages of the major inference providers into a single index — 472 models, 657 price rows, every row carrying the source URL it came from. Same normalized model key, input price per 1M tokens:
| Model | Cheapest $/1M in | Most expensive $/1M in |
|---|---|---|
| Llama-3.3-70B-Instruct-Turbo | $0.10 (OpenRouter) | $1.04 (Together) |
| Meta-Llama-3.1-8B-Instruct-Turbo | $0.02 (DeepInfra) | $0.80 (OpenRouter) |
| Qwen3.8-27B | $0.04 (OpenRouter) | $0.80 (Groq) |
| DeepSeek-V4.1-Flash | $0.009 (OpenRouter) | $1.32 (Together) |
That's a 10x spread on the Llama-3.3-70B, 40x on the 8B, and over 140x end-to-end on DeepSeek-V4.1-Flash. Some of the top end is provider variants and quantization differences — which is exactly why every row links its source page, so you can check what you're actually getting before you route traffic on price alone.
The point isn't that the expensive listings are scams; it's that "the price of a model" isn't a number, it's a range across providers, and the range is wide enough to matter at any real token volume. If your router or your FinOps sheet hardcodes one provider's price, you're probably leaving money on the table.
The index is live on RapidAPI as LLM Token Price Index with a free tier (1,000 requests/month): https://rapidapi.com/vangogh/api/llm-token-price-index
/summary gives you the cheapest listing per model in one call; /prices filters by model or provider when you want the full range.
Which providers am I missing? That's the main way this gets better.
Top comments (0)