Tokens on sale at OpenAI

OpenAI recently reduced the price for tokens from the Terra and Luna versions of GPT 5.6.

These models were only released on July 9th, so the original pricing was in place for just one month. That’s a quick timeline for a pricing change. So what could be the positive spin from this move for OpenAI? This change does appear to be driving revenue. Take a look at the OpenRouter chart of OpenAI tokens from this timeframe.

From the OpenRouter chart, we see that the token usage has increased significantly after the price drop. Volumes for Terra have approximately tripled after the price change and Luna volumes have increased by 10x.

As long as this price level is greater than the cost to host the models, OpenAI’s revenue growth story will get a boost from this pricing.

Stratechery made a Youtube video titled Who’s afraid of Chinese models. The video makes two arguments against the common opinion that Chinese models will drive down token prices in a race to the bottom. This will pressure the leading AI labs. Stratechery counters:

1. Tokens are not a true commodity. Models behave differently. Some are smarter than others and some are better at different skills than others.
2. The leading labs may have a cost advantage. In a commodity market, the winner will be the firm that has the lowest costs since they have the best margins. Firms producing at breakeven will eventually go out of business.

Both of these arguments makes sense to me, but I’m not sure the recent trend points to an easy victory for either Anthropic or OpenAI.

Re: No1 on the question of whether tokens are a true commodity or not, it’s obvious different models are producing different outcomes. But this is also true for separate runs of the same model! The coding harnesses like Claude Code and Opencode are meant to address this concern by looping model output until the code is working. The frontier labs do enjoy a price advantage while they sit on the frontier of intelligence. This is why OpenAI maintained pricing for the Sol tier. Over the past year the gap between the frontier models and the rest of the pack has been shrinking. One year ago Claude was clearly the answer to coding challenges. Today, many Chinese models are competent. If OpenAI and Anthropic always produce the smartest models, they will maintain pricing power, but it’s hard to continue this pace of innovation.

Re: No2 on whether the leading labs have a cost advantage? They don’t today enjoy an advanage on hardware costs. GPUs are the largest expense for a data center and Nvidia is available to all the players. (Maybe Google’s TPU is the low cost leader?) That leaves intelligence per token as the main driver of cost efficiency. The leading labs are hoarding some of the best researchers in the field, but there is also a ton of competition. Even excluding the Chinese competitors, Google, SpaceXAi, and Facebook are all trying to win market share with the advantage of a strong cash business to lean on. If OpenAI can be the lowest cost producer, that still isn’t great for margins. They may “win”, but the spoils won’t be as great. The business dynamics will be closer to traditional companies than some of the winner take all champions from Silicon Valley’s previous generation.

So even while higher revenues are a great outcome from OpenAIs price cuts, I think this is ultimately a bad signal from them. If they enjoyed a huge edge in the AI space, they wouldn’t have to cut prices to raise volumes. They were uncompetitive on the low end and had to lower prices to gain share against cheaper models. On the high end, Anthropic and OpenAI still have pricing power today, but how long will they enjoy that position?