The market is pricing tokens, but it should be pricing work
Frontier token prices collapsed in July, and the coverage has the story backwards. Buyers do not purchase tokens; they purchase completed tasks, and once you price the work rather than the tokens, the ranking inverts. The near-priciest model on the sheet becomes the cheapest to run, and the models winning the price-cut headlines become the most expensive way to get an agentic job done. The best model for the money is currently Claude Opus 4.8, but Google’s Gemini 3.1 Pro, a frontier model at a fraction of the price, closes the gap to a whisker. This note shows why, and it argues that the price war is a trap that transfers the appearance of savings to the buyer and the real cost of failure right back to them.
Mythos 5, the priciest sticker on the sheet at $50, runs near cheapest, while the cheapest sticker, Muse Spark, runs dearest. Cost per completed enterprise task at a fixed quality: The near-top sticker finishes the work for less than half what the cheapest model on the sheet costs to run.
Table of contents
| Key takeaways | 1 |
| The technology behind a token price | 2 |
| The unit of account: Value per token | 3 |
| Margin mechanics: Who can subsidize, and is this a war? | 5 |
| The barbell and the squeezed middle | 7 |
| Market implications | 7 |
| Consumer implications | 8 |
| The surface decides | 9 |
| The verdict: Best model, best value | 9 |
| What would change the call | 10 |
| References | 11 |