Someone calculated that the pre-training calculation amount of DeepSeek V4-Pro is about 1e25 FLOPs. OpenAI has 100,000 GB200s in hand. Even if it only uses 15% utilization, it can complete training of the same scale in 37 hours. In a day and a half, Salvation will reproduce a DeepSeek V4 for you. This is OpenAI in April 2026, and the computing power reserve has been exaggerated to this point. However, OpenAI’s problems cannot be solved by computing power alone.
The release of GPT-5.5 yesterday allowed OpenAI to briefly regain its former glory.
This morning, Ultraman, who was in a good mood, sent an
On the surface, it's all good news.
But insiders know that behind this calm tweet, OpenAI has just experienced the bloodiest "big purge" since its establishment.
Turn your boss’s ideas into reality
Just now, a long article in the New York Times exposed a lot of shocking news.
One detail is particularly ironic. There is a semi-public chat group within OpenAI called "Turn Sam's Tweets into Reality."
There is only one thing that employees do in the group, which is to turn the pie that the boss drew casually on social media into code that can actually be delivered.
Many times, like the rest of the world, they only know what the company is going to do after reading the tweets.
This group has been around for several years.
Altman himself likens this approach to "betting on a series of startups internally at the same time."
ChatGPT’s explosive success lulled him into a false sense of security. The cost quickly emerged.
Computing power is frequently transferred between different teams, product priorities change every few weeks, and employees sometimes really don’t know what the company wants to do.
Google and Anthropic are approaching strongly, eating up OpenAI's leading position one by one.
Finally, Ultraman made up his mind to take action.
Sora is dead, the science club is disbanded, and three people are gone
On April 26th, which is tomorrow, Sora will officially shut down.
This AI video application, which once topped the App Store download list, had 1 million users at its peak and less than 500,000 when it was shut down.
He burns $1 million in computing power every day, but his total lifetime income is only $2.1 million.
In other words, its one-day running cost is half of the money earned during its entire life cycle.
Foreign media The Register gave OpenAI the title of "product killer."
A week ago, three senior executives announced their resignations on the same day.
Sora head Bill Peebles, head of science department Kevin Weil, and corporate CTO Srinivas Narayanan all left within one day.
Kevin Weil's science department only lived six months.
In the 24 hours before being disbanded, the team rushed to release the last model GPT-Rosalind. It's like a suicide note.
Of the 11 co-founders, only Altman and Brockman are left.
In addition to Sora and the Science Department, the NSFW chatbot project, independent social network project, and AI shopping function were also axed. OpenAI internally calls these "side tasks."
Now, all the branches are dead.
One potato and 10,000 NVIDIA employees
The remaining cards in Ultraman's hand are GPT-5.5 and Codex.
GPT-5.5, internally codenamed “Spud” (potato), was launched on April 23.
Terminal-Bench 2.0 scored 82.7%, long context reasoning doubled from 36.6% to 74%, the hallucination rate dropped by 60% compared with the previous generation, and multiple core indicators surpassed Anthropic’s latest Opus 4.7.
Greg Brockman’s words, “This is a new category of intelligence.”
Running points is one thing, but efficiency is the killer move.
The data given by early testers reaches the same level of intelligence as GPT-5.4, and GPT-5.5 consumes significantly less Tokens. Thinking Heavy mode can produce better answers in 2 minutes than GPT-5.4 can produce in 10 minutes. The output quality of the Pro version in 8 minutes is better than the 30 minutes of the previous generation.
However, API pricing has also doubled. Input and output are $5 and $30 per million tokens respectively, twice that of GPT-5.4. OpenAI said that token efficiency has improved, but the actual cost has only increased by about 20%.
As for whether developers will buy this, it depends on whether the agent capabilities of GPT-5.5 can really smoothen the enterprise workflow.
Lao Huang sent an internal letter to the entire company
Nvidia is already betting.
Lao Huang sent an internal letter to the entire company, asking more than 10,000 employees to use Codex. Engineers, legal affairs, marketing, finance, HR, all covered.
The deployment method is very hard-core. Each employee is assigned a cloud virtual machine. The agent has its own computer like the employee. If there is a problem, it can freeze and capture the stack.
The internal feedback used two words, "blasting" and "life-changing".
The debugging cycle has been compressed from days to hours, and experimental iterations have been compressed from weeks to overnight. Scaled down to hours, experimental iterations compressed from weeks to overnight.
|
|
Altman forwarded this letter on
Codex’s weekly active users increased from 3 million to 4 million in two weeks. Seven of the world's top system integrators have signed contracts, Cisco, Ramp, Notion, and Rakuten are queuing up for access, and the number of enterprise users has increased six times compared with the beginning of the year.
Altman’s plan is to merge ChatGPT, Codex and Atlas browsers into a desktop “super application”. No longer splitting up efforts to make a dozen products, stuffing everything into one entrance.
Fidji Simo positions this direction as "AI operating system".
This is OpenAI’s strongest card currently. Whether we can make a comeback or not depends on this.
The moment he was overtaken
The cards have been played, but it still doesn't look good when you open the ledger.
Anthropic's annualized revenue has reached $30 billion. OpenAI is 24 billion.
A year ago, that number was reversed. OpenAI 6 billion, Anthropic 1 billion. The gap seemed insurmountable.
Then Anthropic took 15 months to increase annualized revenue from 1 billion to 30 billion. 30 times.
Epoch AI originally predicted that this intersection would occur in August 2026, but it was four months ahead of schedule.
On unlisted equity trading platforms such as Forge Global, Anthropic's valuation once exceeded US$1 trillion.
Although Anthropic does not have a national-level application with 900 million weekly active users, it has no video generation, no shopping function, and no social network. But it has APIs and enterprise contracts.
More than 1,000 corporate customers pay more than $1 million a year, up from 500 two months ago. 8 of the Fortune 10 companies are Claude's clients.
Ramp's business spending data is even more grim. Anthropic’s invoice share in the enterprise AI chat market has exceeded 60%, compared with only 10% a year ago. OpenAI dropped from dominance to about 35%.
OpenAI's own CFO Sarah Friar couldn't help but sigh, "I have never seen such a scale of growth."
She was talking about OpenAI itself. But this sentence is more accurate for Anthropic.
Even worse is the rate of cash burn. OpenAI is expected to lose US$14 billion in 2026, with a cash burn rate of 57% of revenue. Anthropic's cash burn rate has dropped to 33%, and cash flow is expected to turn positive in 2027.
OpenAI spent 4 times the training cost and made less money.
CEO of 852 billion company, holding 0% shares
On the leaked equity statement, there are two words written on the CEO row, "None/Pending".
Altman holds 0% equity in OpenAI. The annual salary is approximately US$66,000.
The founder and CEO of a company valued at $852 billion doesn’t own a penny of shares in his company.
His $2 billion fortune comes from elsewhere. Stripe, Reddit, nuclear fusion company Helion, and aerospace company Stoke Space have nothing to do with OpenAI.
More subtly, according to WSJ reports, Altman once pushed OpenAI to invest US$500 million in Helion, which he personally holds shares. The investment will more than sixfold Helion's valuation, and his own holdings will also skyrocket. There are people inside OpenAI who are upset about this.
Opposite Dario Amodei has a standard founder shareholding structure, and his net worth has soared to US$7 billion in early 2026.
When the company goes up, he goes up. Interests are naturally bound.
72 days for a new trader
Let’s talk about the transformation of OpenAI.
The real man behind this is Fidji Simo.
She was hired from Instacart to serve as "application CEO" and actually took over the day-to-day operations of OpenAI.
The first thing you do after arriving at work is sound the red alert.
She characterized the rise of Anthropic as a wake-up call, and then worked with Altman and CFO Sarah Friar to review all the company's projects and judge whether to retain them one by one.
What remains are the Codex and ChatGPT super apps. Cut off everything else.
Then Simo collapsed.
On April 3, her condition worsened due to POTS (a neuroimmune disease) and she announced that she would be taking a few weeks off.
On the same day, CMO Kate Rouch resigned due to cancer recovery, and COO Brad Lightcap was transferred to "special projects."
Brockman temporarily takes over production. CFO Sarah Friar, Chief Strategy Officer Jason Kwon, and Chief Revenue Officer Denise Dresser will share the work left by Simo.
One person's work is done by three people.
Even trickier is the timeline.
Ultraman is pushing for an IPO, as early as the end of this year. But several people familiar with the matter revealed that some executives, including Sarah Friar, believed the timetable was "too aggressive."
Altman himself said the truth in the podcast, "My excitement level about being the CEO of a public company is 0%."
Is a company with an annual loss of 14 billion, a company that has just cut off half of its product line, a trader on vacation, and a CEO who has no interest in going public, really ready for an IPO?
You are bald, but have you become stronger?
Last month, OpenAI gave all employees a week off. The official reason is "to prevent overexertion."
Older employees have another interpretation.
9 million enterprise paying users, 900 million weekly active users, and 24 billion annual revenue. These numbers are mythical for any company.
But OpenAI’s problem has never been that the numbers aren’t big enough.
It spent two years chasing all possible directions. When I looked back, I found that the former vice president of research who felt that he was not idealistic enough and left with a group of people to start a business had already surpassed himself in terms of revenue.
The most ironic thing about this is that OpenAI’s original intention was to “develop general artificial intelligence for the benefit of all mankind.” It was born with the color of saints fighting against commercial monopoly.
But now, in order to catch up with an opponent who defected from its own company in terms of financial statements, it has killed the science department, the video platform, and the side tasks that once made people feel that AI has "warmth" and "fun".
Has it become bald and stronger?
GPT-5.5 is really strong. Codex is really changing the way businesses work.
But the words Altman wrote in his latest post, "We want to be a platform for every company, every scientist, and every ordinary person," are put together with what he actually did this week, and the cracks in the middle are getting wider and wider.
The science department was cut. The creative platform is closed. Altman spent two years trying all directions, and finally found that the answer was the enterprise market and coding tools.
Anthropic used the past two years to achieve the answer of 30 billion.