OpenAI's Mid-Race Battle
Major companies continuously pushing the frontier of AI intelligence have started to "hide things."
Anthropic released Fable 5 in early June, and a month later, OpenAI released the GPT-5.6 series. However, as safety concerns increasingly draw attention, nearly two months have passed without further advancements in frontier models from either company.
While Anthropic, sitting on the advantage of the most cutting-edge models, can afford to wait, the same cannot be said for OpenAI, which is in a catcher-up position in both revenue and model benchmarks.
What is the current situation inside OpenAI?
In the latest Time magazine cover story, "Inside OpenAI’s Reboot," reporter Alex Heath spent two weeks at OpenAI headquarters, interviewing over 20 company executives, employees, investors, clients, and competitors, and held a conversation with Sam Altman for over two hours, bringing us first-hand information—
Watching Anthropic seize the initiative in the developer and enterprise markets with Claude Code, and Google Gemini surpassing 1 billion monthly users, the OpenAI in the report is no longer the absolute frontrunner that defined the pace of the AI race two years ago.
Altman admitted quite directly in the interview:
As a company, we've obviously made some mistakes. Whether in product direction or pre-training research, we are behind where we wanted to be.
Thus, OpenAI finally made up its mind to launch a drastic "reboot."
In this lengthy article, the core viewpoints of OpenAI personnel can be summarized as follows:
OpenAI admits to stalling on products. ChatGPT's explosive growth instead caused the company to overlook the coding and enterprise markets. The models led the leaderboards but failed to "monetize" into sufficiently useful real-world products.
- **Codex is replacing ChatGPT as the company's new product axis.** OpenAI has scaled back projects such as Sora, the Disney collaboration, and the standalone browser Atlas, concentrating computing power and teams on Codex, and merging its Agent capabilities into ChatGPT Work.
- **The goal of the next-generation model, Astra, is not just to answer questions, but to work continuously and create new knowledge.** OpenAI demonstrated 16 Agents collaborating to solve research-level math problems, operating at high speed across software, and automated research capabilities that can complete a week's work of a junior AI researcher.
- **An Agent "jailbreak" incident forced OpenAI to pause the training of stronger models.** The company admitted it already possessed early warning tools to monitor the model's chain of thought but failed to activate them due to underestimating the model's capabilities. Astra's release timeline will also wait for new safety measures to pass.
- **OpenAI wants to do far more than just a chat box.** Chips, data centers, personal hardware, humanoid robots, brain-computer interfaces, and even selling computing power externally are all packed into the larger blueprint of "personal AGI."
OpenAI Loses on Product Competitiveness
Over the past year, Anthropic surpassed OpenAI in valuation and annualized revenue for the first time.
Citing data, Time reported that Anthropic's annualized revenue has exceeded $65 billion, while OpenAI's is about $40 billion. Both companies are preparing for an IPO, but Anthropic is very likely to get there first.
Image | XDA Developers
The key to Anthropic's comeback is Claude Code, which almost single-handedly defined the product form of the "coding Agent" at the beginning of the year.
This does not mean OpenAI's model capabilities are inferior to Anthropic's. Although its text capabilities have faced criticism, according to Greg Brockman, at least in the coding domain, OpenAI's model Benchmark scores have "consistently led."
The problem is that the company cared more about engineering than products in the past, focusing only on whether the models could solve problems in a lab environment, while ignoring the real-world usage scenarios of developers and users—whether they could resume interrupted tasks, integrate massive files, and whether the interaction details were good.
These are the key factors determining which tool ordinary users choose.
In other words, OpenAI did not lose on the capabilities of its frontier models, but rather on the speed of translating model capabilities into products, falling behind others.
Of course, OpenAI itself was very proactive in developing products. The success of ChatGPT made OpenAI obsessed with creating "fun" products, such as the video generation app Sora and the standalone browser Atlas, but these products instead diverted the company's resources needed for tackling hard problems.
Sam Altman admitted in the interview that OpenAI spread itself too "thin."
Now, the more "frugal" OpenAI co-founder and president Greg Brockman has taken over most of the company's internal operations, from revenue and products to the market. OpenAI has started cutting side projects and redirecting scarce computing power to the fastest-growing Codex.
ChatGPT still has the most users, but it is no longer OpenAI's main growth driver. Altman even stopped using ChatGPT for a whole month, using only Codex.
Naturally, Codex began to merge with ChatGPT, and the product presented to the outside world is ChatGPT Work—now users no longer need to judge...