Wanted to know what different models actually cost across the whole market. Numbers turned out really interesting.
The spread.
Cheapest output on the platform is Mistral Nemo, $0.03 per million tokens. Most expensive is o1-pro at $600. I re-ran that twice because it looked like a units bug. Median paid model is about $2, so most of the catalog sits down near the floor and there's a thin little line of stuff way up at the top.
Provider averages, with a caveat.
- OpenAI: $47.63
- Anthropic: $44.79
- Google: $5.58
- Mistral: $3.68
- Qwen: $2.86
- Meta: $0.74
Caveat first because someone will say it anyway: these are averages over each provider's catalog, not weighted by what people actually run. OpenAI's number is dragged way up by o1-pro, which I doubt anyone is using at volume. Blended is 3:1 input to output, which is roughly what my own usage looks like.
Even so, Meta at $0.74 against OpenAI at $47.63 is a 64x gap. For the stuff I use models for (mostly code and summarizing), I don't get 64x anything.
Output tokens are where reasoning models get you.
Input and output are priced separately, and on the thinking models the ratio gets silly. Qwen3's thinking variants are $0.20/1M in and $2.40/1M out, so 12x. Gemini 2.5 Flash is 8.3x. Fine if you're sending one question. Less fine if you've got an agent looping thirty times and every step is paying the output rate. I got a bill like that once and it took me an embarrassingly long time to work out why.
19 free models, and a few are usable.
Not trial-credit free, actually free on the API:
- NVIDIA Nemotron 3 Ultra, 1M context
- Google Gemma 4, the 26B and 31B, multimodal, takes video, 262K context
- Poolside Laguna S and XS, 262K
- gpt-oss-20b, 131K (an OpenAI model, on the free list)
There are rate limits obviously. But for messing around or something low volume it's a lot better than it used to be.
Context went up 63x, price didn't really move.
| Year | Avg context | Avg cost/1M |
|---|---|---|
| 2023 | 10.5K | $22 |
| 2024 | 140K | $12 |
| 2025 | 357K | $21 |
| 2026 | 662K | $16 |
Price per token is roughly flat across three years. Context is up 63x. Whatever you think about everything else going on, that part is real.
Feels like two separate products now.
One side is $0.03 to $2 per million with big context windows, Mistral and Meta and Qwen and DeepSeek. The other is $30 to $600, OpenAI and Anthropic up top. They're not really pitching the same buyer anymore. Down at the bottom price stops being a thing you think about at all, and up top you're paying because the output quality moves some number in the business.
Data's from the OpenRouter API on Aug 16.
Link to full dashboard: https://app.vetros.dev/dash/eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJ0eXAiOiJzaGFyZSIsInBpZCI6IjEyMmZmNTk1IiwiZGFzaCI6ImRfODdmNDU3MzkiLCJ2ZXIiOjIsImlhdCI6MTc4NzA4NDc5MH0.V8uCPZtnzJ-djAXAv3HEmmZUHPkhO2NfhSgG2zGMYqw
[link] [comments]