On the 17th at early morning, the release of Kimi K3 went viral across the internet, marking the second "DeepSeek moment" for open-source AI.
Foreign media Axios reported that K3 is priced far lower than the high-end models it challenges. How long can the high-price strategies of US AI companies last? Coincidentally, around the same time, two giants were engaged in a customer acquisition war. Previously, Altman posted on X, not mentioning a new model, but starting with an admission of fault:
Our performance over the past 12 months has not been ideal, and that's largely my fault.
For someone who has been leading OpenAI in a head-to-head battle with Anthropic, publicly characterizing the past year as "not good enough" and taking the blame is unusual in itself.
What truly shocked the entire internet was the second half of his statement.
Altman then pivoted, stating that OpenAI is about to experience its "best 12 months ever," the team is performing excellently, and they are "seeing results that satisfy them."
What exactly is Altman betting on with "the best 12 months ever"?
The internet's first reaction was almost unanimous: Is GPT-6 coming?
Added 3 Million Active Users
in a Few Days
"Close Twitter and go do your real work!"
Echoing Altman's tweet was OpenAI's Codex lead, Tibo, who dropped a new number of 9 million active users on the 16th.
Looking at the active user numbers stacked together: In February, Codex had less than 1 million; by July 12th, 6 million; July 14th, 8 million; and by the 16th, Codex and ChatGPT Work combined broke 9 million.
Four days, an increase of 3 million people.
Tibo said in his post that he wanted to restore quotas earlier but was held up by the "millions of tasks" the team was busy with to keep the system from crashing and ensure stable operation.
He ended the post reminding everyone that quotas would be restored in a few minutes and to focus on their own work instead of constantly refreshing Twitter.
Simply put, user growth was too rapid, and the engineering team was working around the clock to plug the holes.
This brings to mind Anthropic CEO Dario's "humble brag" a few months ago.
In May, Dario said at a developer conference that the company had originally planned for 10x annual growth, but Q1 revenue and usage, annualized, skyrocketed by 80x: that's where the compute tension came from.
He half-jokingly complained: He really hoped the 80x wouldn't continue, it's too hard to handle, and looked forward to returning to a more normal number, "a mere 10x" would be fine.
Two AI giants, one working non-stop to plug holes, the other wishing its growth would slow down. The fire of agents is burning faster than anyone planned.
In the Same Week
Two Rivals Compete to Give You More Quota
Altman's apology tweet landed right in the middle of a fierce battle of attrition between OpenAI and Anthropic.
Not long after ChatGPT Work's release, OpenAI temporarily removed the 5-hour usage limit for Plus, Pro, and Business plans, and repeatedly reset user quotas: first to about 500,000 users, then rolling out to 7 million users, replenishing everyone once.
Anthropic didn't back down either.
It extended paid access to Claude Fable 5 again and increased the weekly quota for Claude Code by 50%, both valid until July 19th.
One side is removing limits and frantically giving away quota, the other is increasing access quotas and raising limits.
On the surface, both are pampering users; in reality, they are fighting for users.
Looking deeper, this is an arms race in the age of agents: Whoever can retain users during this frenzy will grasp more real long-task data.
The most valuable thing is the real usage data generated when an agent works for you for several hours. Whoever has looser quotas gets more users and usage volume, and thus more data.
And OpenAI's CFO has already applied this calculus to its competitor.
According to OpenAI's statement, on the long-cycle engineering task benchmark DeepSWE v1.1, GPT-5.6 Sol at its highest reasoning tier scored 72.7%, surpassing Claude Fable 5's 69.9%; while estimated API costs are 36.2% lower.
Higher scores, yet cheaper—this is exactly what "how much work per dollar" aims to prove. Simultaneously, output tokens are reduced by 54%, and estimated API costs are lowered by 36.2%.
Interestingly, the named Anthropic did something else almost simultaneously: extended Fable 5 paid access again, increased Claude Code weekly quota by 50%, both until July 19th.
One uses charts to prove it's more cost-effective, the other just opens the floodgates wider for you to try.
This also explains why both companies would rather bear system pressure than push usage volume up. High-intensity real-world usage itself is a moat.
However, some have picked up a different scent from this rapid surge.
Economist Jeremy Nguyen, commenting on "Codex and Claude Code resetting quotas on the same day," said that perhaps many years from now, we'll tell young people about the early "agent token war," how crazy the token subsidies were back then.
He dug up old stories from the dot-com bubble: back then, some startups paid you hundreds of dollars a month just to keep an ad bar on your computer.
His question was pragmatic: If such a window only comes once, how can one make the most of Codex and Claude Code now to get their money's worth?
On this track, Anthropic's Claude Code has been constantly active recently, with enterprise agents for finance and engineering already deployed.
The agent war between the two has just begun.
Hours After the Apology
The CFO Presented a Ledger
Coincidentally, on the same day Altman posted, another article went live on OpenAI's official website.
The author wasn't Altman, but CFO Sarah Friar.
Friar opened by saying that everywhere she goes, CFOs are asking the same question: How to get more value for money from AI. Her answer is a new yardstick—Useful Intelligence per Dollar.
In the past, measuring software success looked at adoption rates: how many seats bought, how many active users, renewal rates. Friar said AI needs a tougher metric: look at how much work is actually accomplished.
She even directly debunked the illusion of "token unit price": Lowest token price ≠ Lowest cost per outcome.
Cheaper model tokens are cheaper, but may require more attempts, more time, and more human review; expensive models get it right the first time. What should really be calculated is the total cost per qualified completed task, the real calculation is the
Full Cost Formula = (Model call cost + Compute resources + Human review time + Retry attempts + Rework cost) ÷ Number of successfully qualified tasks.
What truly determines cost-effectiveness is never the single token price, but the complete cost of one successful task.
The GPT-5.6 family (Sol flagship, Terra balanced, Luna fast and low-cost) was born for this purpose.
A support team's "completion" is a customer issue resolved; an engineering team's "completion" is a code change passing tests; a legal team's "completion" is a contract reviewed without errors.
Reading this and looking back at the earlier quota war, the flavor changes.
OpenAI removing the 5-hour limit and repeatedly resetting quotas, Anthropic increasing Claude Code weekly quota by 50%—this isn't just subsidizing users to grab data. When the unit of measurement shifts from "seats" and "tokens" to "completed work," giving away quota isn't a concession, it's changing the metric.
Whoever gets everyone used to calculating based on "how much work is done" first redefines how this market competes.
Capability wins the first use, reliability makes AI become the work process itself.
As usage grows, is every dollar of AI creating more value for you?
The answer lies in the compute infrastructure flywheel: Better infrastructure → Stronger models → Better products → Higher adoption → More revenue → Sustained investment in next-gen research and compute.
When, over time within the same workflow, the number of successful tasks grows faster than the total cost, and quality improves rather than declines, "useful intelligence per dollar" achieves positive compounding.
What the Duopoly is Fighting Over
is Who Becomes Your AI Colleague
Taking a broader perspective makes things clearer.
Anthropic struck first earlier this year. Claude Code became legendary among developers, and the subsequent Cowork pushed agents from programmers to general knowledge workers, taking the lead in "AI office work."
OpenAI's ChatGPT Work, released on July 9th, almost directly targeted Cowork's selling points, built-in Codex, and even launched the latest model GPT-5.6 Sol released the same day.
A command like "Help me create a project tracking table" results in a Gantt chart with 18 projects and 29 tasks: This isn't chatting, it's delivering work.
What changed this time isn't parameters, it's form.
In the past, using ChatGPT meant you asked, it answered; now you just give one goal, and it handles the rest: breaking down tasks, calling tools, searching your files and apps, working for hours straight until the job is actually done.
The official statement clarifies the positioning: ChatGPT is no longer just a machine that answers questions, but a partner for tackling complex work.
OpenAI also experimented on its own company: Today, nearly 100% of internal teams, from finance to sales, use ChatGPT Work and Codex to get work done.
Furthermore, compared to Anthropic, OpenAI holds a card that's hard to beat short-term: distribution.
Cowork requires users to actively download a desktop client; but integrating agents into ChatGPT means over 900 million weekly active users are already using agents the moment they open that familiar app.
Altman himself stated that the usage of agentic products increased 2.5x week-over-week. This curve is part of the foundation for his "strongest next year" confidence.
Returning to Altman's prediction.
He's likely betting not on a new model called GPT-6, but on a more fundamental form shift: transitioning AI from "answering questions" to officially "doing work for you." And Friar's yardstick is precisely prepared for this shift—when AI starts doing work, what measures it is no longer parameters, but how much work is done and how much it costs.
One makes statements on the front stage, the other changes the accounting ledger backstage.
The final outcome of this competition may not depend on whose model is stronger, but on who truly moves into everyone's workflow first, becoming your AI "colleague."
References:
https://x.com/sama/status/2077817060068057493?s=20
https://x.com/JeremyNguyenPhD/status/2077719990116258162
This article is from the WeChat public account "New Zhiyuan", author: ASI Apocalypse, editor: Yuanyu Aeneas
