WhiteLine Daily: a reflection of the WuShuo team’s thinking, providing readers with the most valuable information and analysis for the day, and catching the trend shifts in the AI era.

A one-sentence takeaway

AI is now going through a change that looks a lot like the early days of the internet: per-unit intelligence is getting cheaper, but agent work time and token consumption are growing even faster. Next, the market will look not only at whose model is stronger, but also at usage, distribution efficiency, and gross margin.

01｜Codex starts to move beyond the programmer circle

In its latest disclosure, OpenAI reports that Codex’s weekly active users have exceeded 5 million; of that, about 20% are knowledge workers, and this segment’s growth rate is more than 3 times that of developers. Beyond writing code, users are also using Codex to create reports and spreadsheets, organize contracts, complete research, and process internal data.

Beyond the number of users, what’s even more worth noting is usage depth. Among users in external organizations who used Codex in the past month, the share was 17.3%, but Codex has already contributed 63.3% of the total output tokens from Codex and ChatGPT combined. In other words, the agent penetration rate isn’t high yet, but once people start using it, token consumption increases noticeably.

Because the chatbot mainly answers questions, while the agent receives a job that requires continuously reading files, calling tools, modifying content, and checking results. More than 10% of Codex users run three or more agents in the same week. Based on estimates of the manual time required to complete similar tasks, the share of individual users giving Codex work that takes more than eight hours has grown nearly tenfold since the beginning of this year.

Token records are no longer just counts of conversations—they’re starting to get close to actual work volume.

02｜The cheaper the tokens, the higher the revenue might be instead

Since mid-July, the real-world usage price of top U.S. models has dropped by nearly a quarter. After OpenAI cut the price of Luna by 80%, usage on OpenRouter grew 14x in the short term, while revenue still increased by 34%. Terra’s usage grew by about 5x, with revenue up 45%.

However, this set of data only covers about two weeks after the price cuts. Whether those levels can be sustained long-term remains to be seen. Price declines can lead to more calls, but it’s still unclear whether the new demand comes from stable users or from developers temporarily switching between different models.

DeepSeek has taken another approach: the peak output price of V4 Pro was raised from $0.87 per million tokens to $3.96, while also introducing peak-and-off-peak pricing. Off-peak periods offer price incentives, and peak periods raise prices—using pricing to manage limited compute capacity.

Although each firm’s strategy differs, they’re calculating the same equation: revenue depends on the unit price and the amount of usage. As long as usage grows faster than prices fall, price cuts can still generate more revenue. If usage can’t keep up, the price war will directly compress gross margins.

Usage growth > price drop → token revenue growth

03｜Does making so many tokens actually turn a profit?

Anthropic said its year-over-year annualized revenue growth for May has already exceeded $47 billion. Note that annualized figures simply project the current run rate over a full year—they are not the same as audited full-year revenue.

Anthropic and OpenAI each submitted confidential S-1 filings on June 1 and June 8, respectively. A confidential filing doesn’t mean a public listing date has already been decided. But once the public version of the prospectus is disclosed, the market can see the key sets of data that are currently most lacking: model gross margin, long-term compute commitments, the cost of a single agent task, and how much additional compute is needed to drive revenue growth.

Rapid growth in token usage can support revenue, but if inference costs and compute investment grow even faster, the larger the scale, the greater the chance that cash burn will increase as well. In the next phase, revenue growth rates remain important; unit economics and cash flow will determine whether these companies’ valuations can hold up.