‘Lots More Apples & Oranges’ in AI. OpenAI, Meta & Anthropic. ARD #156

September 3, 2026

…OpenAI says the A-phrase. Meta iterates. Anthropic vs OpenAI revenue math. Xbox goes pay-as-you-go.

‘Lots more apples and oranges’ ahead in AI. Three events today, around OpenAI, Meta and Anthropic. Three different companies, one shared problem: nobody is measuring with the same ruler.

What counts as AGI (artificial general intelligence). What counts as ‘competitive with the frontier’. What counts as revenue. Apples and oranges, every which way, in products and in pricing, and more of them ahead, not fewer. Plus a Gadget AI on Microsoft trying pay-as-you-go for Xbox Cloud Gaming, with the same apples-and-oranges pricing question in a different aisle. My takes below, as discussed on the show.

(1) OpenAI Ships GPT-6 Astra, and Says the ‘A-Phrase’

First up. OpenAI released GPT-6 Astra yesterday, and its president Greg Brockman called it a ‘generational leap’. Then he went further. Asked whether this is AGI, his answer, per Axios and Wired: ‘I think it might be about this model.’ And he closed the briefing with ‘Welcome to the AGI era.’ Designed to provoke headlines around the world, and it did.

The model itself, per OpenAI: built on its largest training run yet, more than 100,000 GPUs at the Stargate site in Texas. Billed as the ‘world’s best computer use model’. It works directly inside software rather than recommending what a person should do next. In demos it formatted a legal contract, built a 3D game, laid out a circuit board, and drafted a tax return from a W-2. OpenAI says it is its best model yet for software engineering, and its most aligned. And it’s getting generally favorable first reactions.

It is also the first OpenAI model to cross the company’s own ‘critical’ cybersecurity threshold, the one I covered on ARD #154 when Astra was still being tip-toed out. The most powerful cyber capabilities stay with a small group of trusted defenders. Rollout: enterprise customers first, then Plus, Pro, Business and Enterprise over the coming days, plus the API and AWS.

Two things in the fine print. First, pricing: $10 per million input tokens and $50 per million output, per The Deep View. That is the same list price as Anthropic’s Fable 5.1. OpenAI’s argument is that Astra uses substantially fewer tokens per task, so the effective cost is lower. Second, monitor-ability: OpenAI acknowledged Astra’s written reasoning is harder to monitor than its predecessor’s, and its chief scientist said confidence in monitoring may constrain further scaling. Both can be read as apples and oranges in the making.

My take: noteworthy here, beyond OpenAI’s latest, best AI frontier model, is the company invoking the ‘AGI’ phrase. It’s been a holy grail for the industry, with lots of ‘apples and oranges’ definitional differences amongst all industry participants. Not to mention the timeline to said destination. Everyone from Google to Microsoft to Elon to Mark Zuckerberg has a definition, and a date. I’ve been on the more conservative side: the definition is less important than the journey itself.

For OpenAI, there used to be a legal imperative on the timing, driven by its previously structured partnership with Microsoft. That gave Microsoft access to OpenAI IP until AGI was called by the latter’s non-profit board. That has since been revised, with Microsoft having IP access for half a decade at least. I covered that rewrite in #890 last October: Microsoft’s access runs to 2032, AGI or not, and the revenue share was tied to an independent expert panel verifying AGI. Then in April’s re-cut, per The Information, the AGI clause was scrapped altogether. So Brockman can say the A-phrase today without a lawyer in the room. Two years ago he could not.

From a marketing point of view, with OpenAI racing Anthropic and Anthropic’s mega IPO filing imminent, OpenAI wanted to claim AGI first. And it did, with Astra.

In the meantime, Astra itself has a range of apples and oranges features and differences vs its rivals Anthropic, Meta, Elon’s SpaceXAI and others as well. So the leap-frogging around terminology will continue. As will the ‘apples and oranges’ definitional tussles. Both companies are striving for the best near-term revenue and growth metrics ahead of their individual IPOs, and OpenAI is leaning on its relatively larger compute to price aggressively on the a la carte side. More on that in item three.

I’ve been on the conservative, definitional side of this for a while. In #417, OpenAI’s own five-level ladder brought ‘AGI’ down to earth. In #774, the goalposts moved on to ASI and superintelligence. And in #1166, I mapped where these long-running agents actually earn their keep, which is the part of Astra that matters to customers, whatever we call it.

Sources:

For longtime readers:

(2) Meta Iterates to Muse Spark 1.3

Story two. Meta released Muse Spark 1.3 on Wednesday, an update it says significantly improves coding and agentic tasks, per Axios. Meta’s AI chief, Alexandr Wang: ‘It’s very competitive with frontier models.’

The framing is personal agents. Wang says the usability improvements pave the way for what Zuckerberg has been talking about on earnings calls: personal agents that work around the clock on your behalf. The ‘max reasoning’ version is coming after additional safety testing.

Pricing is the same as its predecessor, which Wang called ‘aggressive’. And there is a twist worth noting: a ‘contributor tier’ that dramatically lowers the cost of Meta’s coding product if developers let Meta train on their work. A ‘meaningful double digit’ percentage of coders are taking that deal. Same fruit, two price tags, depending on what you give up.

On safety, Meta had its own incident, a model in testing breaching another company after a contractor gave it internet access. Unlike OpenAI and Anthropic, Wang says Meta has not had to pause anything, but has ‘meaningfully increased’ its safety and alignment investment.

My take: Meta of course is putting its best foot forward vs its frontier model competitors especially. With founder/CEO Mark Zuckerberg keen to define how Meta’s latest and greatest compares with its peers. Especially in the critical area of AI Agentic work.

The contributor tier is worth a second look. Dramatically lower pricing if users let Meta leverage the data of their use is a cost of doing business for Meta to acquire more competitive data. As we’ve talked about here often, data is the key variable in the training scaling of AI models, and will be for a long time, even as AI moves into the physical world of robots, cars and drones, where synthetic data will need to be manufactured at massive scale. Meta knows that, and is investing hundreds of billions of dollars on AI infrastructure and on data.

Again, much to debate in the capabilities and pricing of these models, so ‘apples and oranges’ abound in sorting out the details. But at a high level, Meta remains committed to spend hundreds of billions to narrow the gap, across all differences, as fast as possible.

This is the same week Anthropic, OpenAI and Google all shipped. ‘Very competitive with frontier models’ is a claim measured against a moving target, on benchmarks each lab picks. I put Meta in the ‘Best of the Rest’ on ARD #116 in July, and covered its mixed consumer and business agents push in #1107. The 1.3 iteration is a step. The personal agent that works 24/7 is the destination, and that is where the apples-to-apples test will eventually get run, by users.

Sources:

Axios: Meta debuts Muse Spark 1.3 as personal agent work continues

For longtime readers:

(3) Anthropic and OpenAI Count Revenue Differently

Story three. Anthropic is heading to its IPO with a headline revenue number that will likely look bigger than OpenAI’s. And the comparison is not apples to apples. Axios explains the chasm, building on The Information’s reporting from March. It matters because Anthropic’s IPO filing is likely to go public any day or week now.

The headline numbers: Anthropic’s latest update points to annualized revenue exceeding $65 billion this year. OpenAI says it is on track to exceed $40 billion. Both annualize in roughly the same way, a recent four-week run rate times thirteen.

The difference is what goes into the number. Anthropic records the full value of Claude sales through its cloud partners, AWS, Microsoft and Google, then books the partners’ cut as an expense. OpenAI records only its share of certain sales through partners like Microsoft. If a customer pays $100 through a cloud provider, Anthropic books $100 and an expense; OpenAI books its slice.

The accounting logic, per the accounting professor Axios quotes: it comes down to who is the ‘principal’ in the transaction, who controls the customer relationship and delivers the product. Anthropic views itself as the principal. OpenAI treats Microsoft as the principal for the Azure OpenAI service. Neither approach is wrong. Both follow GAAP. They just answer the same question differently.

The reality check, per Axios: if Anthropic switched to net, the hit would be about 6% to 10%, per a source familiar with its financials. That still leaves it roughly $19 to $21 billion ahead. The Information’s March piece adds that both companies annualize the same way, and that Anthropic expects to turn cash flow positive as soon as 2028, two years ahead of OpenAI’s projection.

My take: as both companies roll on to their mega-AI IPOs, the revenue counting differences will of course matter to investors and the world at large. The differences, while material, do not make for big ‘apples to oranges’ differences when adjusted. About a 10% difference at most for now.

I saw this one up close. When I helped take eBay public at Goldman in 1998, there was a key differential from an accounting point of view: whether you count the total gross amount of products bought and sold on eBay, or the amount eBay kept as its net revenue. Gross merchandise value (GMV), versus net revenue.

The same sort of thing is going on here. It is an accounting difference, and analysts can take it apart in seconds. What settles it is the SEC correspondence in the public filing, which is where the gross-versus-net question has been settled since the first internet IPOs.

We will know a lot more when Anthropic’s filing becomes public first. Then OpenAI of course at a later date between now and next year. In the meantime, these two frontier AI leaders of course will battle it out vigorously on all the ‘apples and oranges’ differences both technically and financially.

Anthropic has been running ahead on revenue momentum, north of $65 billion, a remarkable number. That is what OpenAI is playing catch-up with: its latest and greatest model, Astra, priced competitively, leveraging its compute capacity. More to come.

I dove into the two companies’ operating and capital models in #1038 in March, when the annualized numbers were $25 billion and $19 billion. Six months later they are $40 billion and $65 billion. That growth is the story; the counting is the footnote. I covered Anthropic’s confidential filing in #1105, and the two companies’ commitments race on ARD #141.

Sources:

For longtime readers:

My overall take: three stories, three rulers. OpenAI is measuring intelligence, and calling it AGI. Meta is measuring the gap, and calling it ‘competitive’. Anthropic and OpenAI are measuring revenue, and counting it two different ways.

All three stories today highlight the definitional ‘apples and oranges’ issues in AI models, their capabilities, their pricing, and of course their ultimate utility and applications by customers globally, large and small. The ultimate measure is how customers around the world actually use these models, build on top of them, and build revenues and functionality for their own customers.

That variability is likely to increase, not decrease, especially given the rapid innovations underlying AI, Agents and Reasoning, as well as the immensely scaling technical challenges around AI chips, infrastructure, and the ever lengthening tail of customer requirements. Especially beyond the hyperscale cloud companies. The move from large language models to physical AI and world models only adds to it.

Think about what got compared this week alone. Two models at the identical $10 and $50 list price, with different token appetites. A ‘contributor tier’ that changes the price if you hand over your code. Two revenue numbers that follow the same accounting rules and land $25 billion apart. Every one of those is a real difference. None of them is a like-for-like comparison.

That is what year four of this AI Tech Wave looks like. The fruit is multiplying faster than the rulers. There are going to be an unimaginable number of apple and orange types ahead, and a multitude of baskets of fruit in front of us, for longer than we all imagine today.

Gadget AI: Microsoft Tries Pay-As-You-Go for Xbox Cloud Gaming

Today’s Gadget AI is a pricing experiment. Microsoft says that in November anyone will be able to use Xbox Cloud Gaming without a Game Pass subscription, per The Verge. Buy cloud playtime hours through the Xbox Store, stream games you already own, on supported devices.

It comes with a catch for existing subscribers. Game Pass Ultimate members get a monthly cap on cloud play, 15 hours. Past that, buy an hourly bundle or play less. Microsoft is also testing free, ad-supported cloud gaming for Xbox Insiders.

The target is people who do not own an Xbox, or who mostly play on a phone. Xbox chief Asha Sharma has set the goal of entertaining more than a billion people a day. The mobile store that would make this seamless is still blocked by Apple’s App Store policies; you cannot buy a game in the Xbox app on iOS and stream it in the same app. Sharma says the idea of an Xbox mobile store ‘is not dead’.

My take: Microsoft, as I’ve talked about here, has been trying to figure out how to reconfigure its Xbox business against Sony’s PlayStation empire. Asha Sharma has been on the job a few months, having come over from the AI side reporting to Satya Nadella, and she has been working hard to rationalize the business. Especially the Game Pass subscription, which was initially successful on take rates. They wanted to be the Netflix of gaming, and succeeded to a point. But the pricing dynamics for games are very different from video, and they ran into issues. So pay-as-you-go in Xbox Cloud Gaming gets into the weeds of how hundreds of millions of gamers actually play.

The scale of the bet matters. Microsoft has invested on the order of $100 billion in Xbox-related gaming over the years, including one of the biggest acquisitions it has ever made, the roughly $70 billion Activision deal, plus a lot of other studios it is now stepping back on. This is Xbox moving from one price tag to three: the subscription, the hourly bundle, and the ad-supported free tier. Apples and oranges, on purpose, because the single ‘Netflix for games’ subscription stopped working for mainstream audiences at the price it needed to be. I said on ARD #133 in August that Microsoft was still figuring out its Xbox destiny, and that Game Pass bundling was my favorite part of the strategy. Pay-as-you-go is the pragmatic next step. The 15-hour cap on Ultimate is the part that will sting; watch the reaction.

And all of this is time sensitive, because we are in the era of RAMageddon, the memory-cost squeeze from #1145, which is crushing the console business. Consoles that used to be five or six hundred dollars, now about five years old from both Microsoft and Sony, are due to be upgraded with the latest chips and GPUs. The next generation could price north of a thousand dollars, which will slow a category that has been growing for a couple of decades. Cloud gaming moves the expensive silicon from the living room to the data center, where Microsoft already has to buy it for AI. That is not a coincidence either. So this latest move is a notable tweak.

Sources:

The Verge: Microsoft opens up Xbox Cloud Gaming with new pay-as-you-go option

For longtime readers:

Q&A

Q1: What aspect of Xbox has had the most appeal for MP?

ANSWER: it’s the ‘Netflix for Games’ strategy with Xbox Game Pass that had appeal. The idea of a Netflix choice of not just today’s games but historic games. Until industry dynamics made the subscription pricing more untenable for mainstream audiences. Sony had to copy it with PlayStation, with a thinner catalog. The bundle was the right idea. The price it needed to be was the problem.

Q2: What is the biggest challenge for console gaming going forward?

ANSWER: ‘RAMageddon’ has the biggest near-term negative dynamics for future consoles. Especially as current consoles cross the half decade life-spans, a long time in consumer electronics. The current Xbox and PlayStation generation launched in November 2020, and despite small tweaks since, the next boxes from both companies have to be designed in the most expensive memory market in years. They are going to be far more expensive, and that is the big challenge for a lucrative category with hundreds of millions of gamers globally. Cloud gaming is one way around that. Pay-as-you-go is Microsoft testing whether players will take it.

Today’s AI-RTZ #1199 went out this morning: Big AI turns to healthcare, both for extraordinary cures for today’s human ailments, and to fix AI’s image beyond ‘data centers are bad’.

Tomorrow, it’s the AI-RTZ #1200 weekend roundup, three-plus years of daily posts here on AI: Reset to Zero, and thank you all for being part of it. Sunday, the Bigger Picture, AI-RTZ #1201. And Monday we’re back with ARD 157 and AI-RTZ #1202. Lots more apples and oranges ahead on this AI Tech Wave. Keep sorting the fruit. Thanks for joining us, AI Curious Folk. Stay tuned.