Sora shuts down, core team leaves, Anthropic surpasses: OpenAI's darkest hour

Title: “OpenAI, Perilous Times Ahead”

Author: Aeneas, Hao Kun

Source:

Reprint: Mars Finance

Introduction: Ultraman personally killed Sora and the Science Department, bloodily clearing a large wave of “side quests,” finally successfully releasing GPT-5.5. However, facing Anthropic’s revenue overtaking and internal upheaval, this high-stakes IPO gamble—was it a return to glory or a final farewell?

Someone did the math.

The pre-training compute for DeepSeek V4-Pro is approximately 1e25 FLOPs.

OpenAI has 100k GB200 units, even with only 15% utilization, it can complete training of the same scale in 37 hours.

One and a half days, and you can replicate a DeepSeek V4.

This is OpenAI in April 2026, with computational resources already exaggerated to this extent.

However, OpenAI’s problems cannot be solved by compute power alone.

The release of GPT-5.5 briefly restored OpenAI’s former glory.

In a cheerful post on X, Ultraman said, “It’s been a good week, proud of the team, wishing everyone happy development!”

All the obvious news is good.

But insiders know that behind this calm and peaceful tweet, OpenAI has just experienced its bloodiest “great purge” since its founding.

Turning the boss’s ideas into reality

A long article in The New York Times revealed many explosive details.

One detail is especially ironic. Inside OpenAI, there is a semi-public chat group called “Turning Sam’s Tweets into Reality.”

Employees only do one thing in this group: turn the boss’s casual promises on social media into deliverable code.

Many times, they, like the public, only learn what the company is about to do when they see the tweets.

This group has existed for several years.

Ultraman compares this approach to “betting on a series of startups internally at the same time.”

The explosive success of ChatGPT gave him a false sense of security. The cost quickly surfaced.

Compute is frequently shuffled between teams, product priorities change every few weeks, employees sometimes really don’t understand what the company wants to do.

Google and Anthropic are closing the gap, gradually eating into OpenAI’s leading position.

Finally, Ultraman decided to take action.

Sora is dead, the Science Department is gone, three people left

On April 26, Sora was officially shut down.

This once top-ranked AI video app on the App Store, peaked at 1 million users, but by shutdown, fewer than 500k.

Burning $1 million in compute daily, but total lifetime revenue was only $2.1 million.

In other words, its daily operating cost was half of what it earned over its entire lifecycle.

Foreign media The Register dubbed OpenAI a “product killer.”

Some time ago, three executives announced their departure on the same day.

Bill Peebles, head of Sora; Kevin Weil, head of the Science Department; Srinivas Narayanan, CTO of the enterprise—left all within one day.

Kevin Weil’s Science Department lasted only six months.

In the 24 hours before disbandment, the team rushed to release the final model, GPT-Rosalind. Like a farewell letter.

Now only Ultraman and Brockman remain of the eleven co-founders.

Besides Sora and the Science Department, other projects cut include the NSFW chatbot project, the independent social network project, and the AI shopping feature. Internally, OpenAI calls these “side quests.”

Now, all side quests are dead.

A potato and ten thousand Nvidia employees

Ultraman’s remaining cards are GPT-5.5 and Codex.

GPT-5.5, internally codenamed “Spud,” launched on April 23.

Terminal-Bench 2.0 scored 82.7%, with long-context reasoning jumping from 36.6% to 74%, hallucination rate down 60% from the previous generation, and multiple core metrics surpassing Anthropic’s latest Opus 4.7.

Greg Brockman said, “This is a new category of intelligence.”

Benchmark scores are one thing, efficiency is the real killer.

Early testers reported that GPT-5.5, at comparable intelligence to GPT-5.4, used significantly fewer tokens. In Thinking Heavy mode, answers generated in 2 minutes were better than GPT-5.4’s 10-minute outputs. The Pro version’s 8-minute output quality exceeded the previous generation’s 30-minute results.

However, API pricing has doubled. Input and output are now $5 and $30 per million tokens, twice GPT-5.4. OpenAI claims token efficiency has improved, and actual costs only increased by about 20%.

Whether developers buy into this depends on whether GPT-5.5’s agent capabilities can truly streamline enterprise workflows.

Simo sends an internal letter to the entire company

Nvidia is already betting.

Simo sent an internal email to over 10k employees, enabling everyone to use Codex. Engineers, legal, marketing, finance, HR—all covered.

The deployment method is hardcore: each employee gets a cloud virtual machine, with the agent and the employee having their own computer, so issues can be frozen and stacks can be debugged.

Internal feedback used two words: “explosive” and “life-changing.”

Debugging cycles are compressed from days to hours, experimental iterations from weekly to overnight. From days to hours, and from weekly to overnight.

Ultraman reposted this on X, saying, “Deploying Codex across Nvidia’s entire company has been very effective. Want to try it in your company?”

Within two weeks, Codex’s weekly active users increased from 3 million to 4 million. Seven top global system integrators signed on, including Cisco, Ramp, Notion, Rakuten, with enterprise user scale growing sixfold since the start of the year.

Ultraman’s plan is to merge ChatGPT, Codex, and the Atlas browser into a single desktop “super app.” No longer spreading resources across a dozen products, everything is consolidated into one entry point.

Fidji Simo describes this direction as “AI Operating System.”

This is currently OpenAI’s most powerful card. Whether they can turn the tide depends on this move.

The moment they are overtaken

The cards are played, but the ledger still looks grim.

Anthropic’s annualized revenue has reached $30 billion. OpenAI is at $24 billion.

A year ago, the numbers were reversed. OpenAI at $6 billion, Anthropic at $1 billion. The gap seemed insurmountable.

Then, Anthropic used 15 months to grow annualized revenue from $1 billion to $30 billion—30 times.

Epoch AI originally predicted this crossover would happen in August 2026, but it arrived four months early.

On platforms like Forge Global, Anthropic’s valuation once exceeded $1 trillion.

Although Anthropic lacks a flagship app with 900 million weekly active users, video generation, shopping features, or social networks, it has API and enterprise contracts.

Over 1,000 enterprise clients pay over $1 million annually; two months ago, that number was 500. Eight of the top 10 Fortune companies are Claude clients.

Ramp’s enterprise expenditure data is even more brutal. Anthropic’s invoice share in the enterprise AI chat market has exceeded 60%, up from 10% a year ago. OpenAI’s dominance has shrunk to about 35%.

OpenAI’s CFO Sarah Friar can’t help but exclaim, “I’ve never seen such rapid growth.”

She is talking about OpenAI itself, but this phrase is even more fitting for Anthropic.

More troubling is the rate of burning cash. OpenAI expects a loss of $14 billion in 2026, with a burn rate at 57% of revenue. Anthropic’s burn rate has dropped to 33%, with positive cash flow projected in 2027.

OpenAI spent four times the training cost but earned less money.

CEO of a $852 billion company, holding 0% equity

On the equity table, the CEO’s line reads “None/Pending.”

Ultraman owns 0% of OpenAI. His annual salary is about $66k.

A founder and CEO of a $852 billion valuation company, with no equity in his own firm.

His $2 billion net worth comes from elsewhere: Stripe, Reddit, fusion company Helion, space company Stoke Space—all unrelated to OpenAI.

More subtly, according to WSJ, Ultraman once pushed OpenAI to invest $500 million in Helion, in which he personally holds shares. This investment would multiply Helion’s valuation over six times, and his own stake would skyrocket. Some inside OpenAI are uneasy about this.

On the other hand, Dario Amodei, a typical founder-shareholder, saw his wealth soar to $7 billion in early 2026.

As the company grows, so does he. Interests are naturally aligned.

The new operator’s 72 days

Back to OpenAI’s transformation.

The real behind-the-scenes driver is Fidji Simo.

She was brought in from Instacart as “Application CEO,” but effectively took over OpenAI’s daily operations.

Her first move was to sound the alarm.

She characterized Anthropic’s rise as a direct blow, then, together with Ultraman and CFO Sarah Friar, reviewed all projects, judging their viability one by one.

The ones kept are Codex and the ChatGPT super app. Everything else was cut.

Then Simo fell ill.

On April 3, she announced a few weeks’ leave due to worsening POTS (a neurological immune disorder).

On the same day, CMO Kate Rouch resigned for cancer recovery, COO Brad Lightcap was reassigned to “special projects.”

Brockman temporarily took over product management. CFO Sarah Friar, Chief Strategy Officer Jason Kwon, and Chief Revenue Officer Denise Dresser split Simo’s responsibilities.

One person’s work was handled by three people.

The more delicate issue is the schedule.

Ultraman is pushing for an IPO, possibly by the end of this year. But multiple insiders, including Sarah Friar, say this timeline is “too aggressive.”

Ultraman himself admitted in a podcast, “I am at zero excitement about becoming a public company CEO.”

A company losing $14 billion annually, halving its product lines, with its operator on leave, and the CEO uninterested in going public—are they really ready for an IPO?

Bald but stronger?

In March, OpenAI gave everyone a week off, officially to “prevent overwork.”

Veterans interpret it differently.

100k enterprise paying users, 900 million weekly active, $24 billion annualized revenue—these figures are legendary for any company.

But OpenAI’s problem has never been insufficient size.

It’s that it spent two years chasing every possible direction, only to find that the former VP of research, who left with a sense of idealism, has already surpassed itself in revenue.

The most ironic layer is that OpenAI was founded with the mission “to develop general artificial intelligence for the benefit of all humanity.” It was born with a saintly tone of resisting monopolies.

But now, to catch up financially with a rival that defected from within, it has killed the Science Department, shut down the video platform, and eliminated those “warm” and “fun” side quests that once made AI seem more human.

It has gone bald, but is it stronger?

GPT-5.5 is indeed very powerful. Codex is truly changing how enterprises work.

But Ultraman’s latest statement, “We want to be the platform for every company, every scientist, and every ordinary person,” contrasts sharply with what he’s actually doing this week, revealing a widening gap.

The science department was cut. The creative platform was shut down. After two years of trying every direction, the answer turned out to be enterprise markets and coding tools.

And Anthropic, in these two years, turned that answer into $30 billion.