Source: Synced Review
In one minute, 40 million tokens gone up in smoke.
'Lobster father' Peter Steinberger probably never imagined he would trend on social media this way.
In February this year, he was personally recruited by Sam Altman to join OpenAI, responsible for developing 'next-generation personal agents'.
In just a few months, he has already become one of the most voracious API consumers inside OpenAI.
Just moments ago, he posted a breathtaking screenshot on X: The OpenAI API rate limit – a quota of 40 million tokens per minute – was completely devoured by him in one go, draining the balance directly to zero.
What exactly is the concept of 40 million tokens? It roughly equates to the amount of text found on 300,000 pages of A4 paper.
A quota that an average developer might take a whole month to use up, Peter burned through in a single minute.
Under the screenshot, he left just one comment: 'brb calling @sama.' (Be right back, calling Sam Altman).
Altman replied instantly: It will be fixed immediately!
He didn't forget to add a teasing remark: That number is too low.
The tweet immediately exploded.
Netizens were shocked, wondering what Peter was actually up to: 'What are you building? A second internet?'
Peter responded that it was ClawSweeper – a bot responsible for maintaining the OpenClaw code repository.
It's worth mentioning that ClawSweeper runs on 50 parallel Codex instances powered by GPT-5.5.
GPT-5.5 Drops a Bombshell
The Lobster Father Burns Through 40M Tokens in a Minute
The 'culprit' behind Peter's ability to burn through 40 million tokens in one minute is none other than OpenAI's latest ace weapon – Codex powered by GPT-5.5.
GPT-5.5 is defined by OpenAI as 'the smartest, most intuitive model'.
You can throw a chaotic, multi-threaded task directly at it, and it will plan by itself, invoke tools, check results, handle ambiguity, and then drive the task all the way to completion.
This capability is particularly deadly in the coding domain.
When Codex was equipped with GPT-5.5, it directly set the developer community ablaze.
Sam Altman and Greg Brockman rarely promote products simultaneously on social platforms, but the developer reactions were even more exaggerated than the official ones.
Someone directly stated that this might be the most significant moment in AI this decade.
Even Peter himself could not escape the 'law of true fragrance' of Codex 5.5.
In his tweets, he admitted that the 'large model temperament' displayed by Codex 5.5 surprised him, but it also brought a tricky problem: the entire OpenClaw toolchain needed an urgent refactor.
Even more noteworthy is Altman's own statement.
When a developer said that GPT-5.5 had become his 'daily necessity', Altman replied: 'Then get ready for a life transformation.'
From GPT-5.2 to 5.3, 5.4, and now 5.5, OpenAI maintains a terrifying pace of roughly one update every six weeks.
At this speed, a 'life-level transformation' might not be a distant promise, but a matter of the next quarter.
On May 5th, OpenAI will also hold a celebration party for GPT-5.5.
The most dramatic scene: Altman publicly stated on social media that Musk, who is currently suing OpenAI, 'can come if he wants to', and added that 'the world needs more love'.
Codex Transforms into a Super-App, Fully Encircing Claude Code
If Codex 5.5 is OpenAI's battering ram, then its 'full family bucket strategy' is the encirclement net.
User Zibo Gao wrote excitedly after trying the Codex Mac App: 'winning SO HARD'.
He conveniently listed a feature wishlist: a native editor, an iOS app, full browser support, and deep integration with OpenClaw.
The response from Tibo, a core member of the Codex team, was concise and powerful: 'All of this and more is coming.'
OpenAI is transforming Codex from a command-line tool into a full-platform product matrix covering Mac, iOS, browser, and IDE extensions.
Meanwhile, OpenClaw 5.2 also underwent a major upgrade on May 2nd.
It now integrates xAI's latest Grok 4.3 model by default, entirely rewrites the plugin architecture—curing the persistent npm dependency conflicts that plagued developers—and performed deep optimization for multiple platforms like Discord, Slack, and Telegram.
The target of this combo is very clear: seizing the ecological high ground in AI programming tools and squeezing Anthropic's Claude into a corner.
After reading this, you probably think OpenAI has indeed won massively.
Beneath the $852 Billion Valuation
Lies a Bottomless Money Pit
And then, cold reality hits like a bucket of ice water.
In late April, an in-depth investigation by WSJ reporter Berber Jin ripped off the band-aid.
OpenAI failed to meet its own revenue targets.
Multiple monthly sales targets in early 2026 were missed. The ambitious goal of hitting 1 billion weekly active ChatGPT users by the end of 2025 also fell short (the actual number published at year-end was 800 million).
More fatal is that OpenAI has already committed to spending money it doesn't have.
Combining contracts with Stargate, Oracle, AWS, and Microsoft Azure, the total exceeds $1.4 trillion. This figure is larger than the GDP of Spain and Australia!
CFO Sarah Friar's job is to figure out where the money will come from.
But the problem is, the books simply don't balance.
And the matter goes beyond just 'disputes'.
According to The Information, since August 2025, Friar no longer reports to Altman, but instead reports to Fidji Simo, the head of applications. For a large company's CFO not to report to the CEO is practically unheard of in Silicon Valley.
More suspiciously, Friar was subsequently excluded from investor meetings. Moreover, Simo herself also announced a brief leave of absence due to health reasons.
The day the news broke, SoftBank's Tokyo stock plummeted 10%, Oracle dropped over 7% in pre-market, and CoreWeave and AMD were dragged down as well.
Morningstar predicts that OpenAI's IPO window will have to be postponed from Q4 2026 to mid-to-late 2027.
Anthropic Seizes the Opportunity Amid Chaos
More lethally, the competitor has already completed an overtake.
According to Counterpoint Research Q1 2026 data, the global share of LLM revenue is 31.4% for Anthropic and 29% for OpenAI. OpenAI was surpassed for the first time in this core metric.
In terms of ARR, Anthropic is at $30 billion, while OpenAI is at $24-25 billion. Fifteen months ago, Anthropic was only at $1 billion, meaning it has grown 30-fold.
In the code generation market, Anthropic holds 42%-54%, while OpenAI has only 21%.
Ramp data shows that among enterprises purchasing new AI services in March, 65% chose Anthropic, and only 32% chose OpenAI.
Not only that, but Anthropic’s average monthly revenue per user is $16.2, while OpenAI’s is only $2.2—an eightfold difference.
OpenAI's 900 million weekly active users, a terrifying figure for 2026, sounds impressive, but over 95% of them are free users, and each one is burning cash.
In a leaked shareholder letter, OpenAI listed Anthropic as its biggest competitive threat. But by the time the letter was sent, the overtake was already a fait accompli.
Musk Adds Insult to Injury in Court
Furthermore, Altman can't even focus fully on handling his own business crisis.
In the same week, Musk's lawsuit against OpenAI opened in the federal court in Oakland, California, with a damages claim potentially reaching as high as $180 billion.
On the day of the court hearing, Altman was supposed to attend a collaborative event with AWS in San Francisco, but he had to appear remotely via a recorded video.
My schedule has truly been beyond my control lately.
Legal experts generally believe that Musk's chances of winning the case are low. But that doesn't matter.
Over the past two years, Musk has been tirelessly portraying Altman as 'Scam Altman', a con man from the AI age who rose through bluster and deception.
The cognitive damage caused by this PR war is far more profound than the legal judgment itself.
The More Explosive the Product, the Faster the Money Burns
So the current situation is this: Codex 5.5 ignites the developer circle, GPT-5.5 benchmark scores obliterate everything, and a super-app empire is taking shape.
But on the other side, revenue targets weren't met, the CFO is sidelined, Anthropic has surpassed them in revenue share, and Musk is waiting in court to deliver the final blow.
The crack between OpenAI's product frenzy and business reality is widening. And the better the product, the wider the crack gets.
At the current burn rate, OpenAI is expected to lose $14 billion in 2026, cash consumption could skyrocket to $57 billion in 2027, and profitability may not arrive until 2029 at the earliest.
Over at Anthropic, training costs are only a quarter of OpenAI's, and they might reach positive cash flow by 2027.
The more explosive the product, the faster the money burns. The more users, the bigger the losses.
The myth of Sam Altman is far from a closed case.
References:
https://www.wsj.com/tech/ai/the-lore-of-sam-altman-is-being-tested-like-never-before-968227ea