The Constraint Atlas · AI compute · status: measuring
The same AI job can cost 1,500 times more depending on the model you pick
The limit: What builders can afford to automate is capped by price per token, and most teams pay far more than the cheapest model that does the job.
Evidence
| Measure | Value | Source |
|---|---|---|
| Models tracked with dated prices | 71 | Tensorust LLM price table, 2026-10-23 |
| Cheapest output price (gpt-5-nano, gpt-4.1-nano, gemini-2.5-flash-lite) | $0.40 per 1M tokens | provider pricing pages, 2026-10-23 |
| Highest output price (o1-pro) | $600 per 1M tokens | OpenAI pricing, 2026-10-23 |
What would move it
- Route each task to the cheapest model that passes your own test set, instead of one default model for everything.
- Batch and cached-input pricing for work that is not interactive.
- Open-weight models on rented GPUs when volume is high enough.
The experiment we are running
We record every price change daily and publish the history. The AI Cost Audit applies it to one team's real usage and keeps watching for cheaper options as prices move.
Measured progress
| Date | Reading |
|---|---|
| 2026-10-23 | 71 models, $0.40 to $600 per 1M output tokens |
Next actions (what you can do today)
- Price your workload across every model:
https://api.tensorust.com/v1/llm/cheapest?input_tokens=...&output_tokens=...(free). - Check the LLM API cost calculator before you pick a default model.
Tools that act on it: LLM API cost calculator · AI Cost Audit · LLM price history API
Data: /data/v1/atlas.json (CC BY 4.0). Updated 2026-10-27.
Art: Nicholas Roerich, Himalayas, Sikkim (c. 1928-29). Public domain, via Wikimedia Commons.
