TheChek News Flow

Alexandr Wang, the company’s AI chief, said the new model will debut soon, along with a large language model dubbed Avocado

Meta Platforms META 2.30%increase; green up pointing triangle is developing a new image and video-focused AI model code-named Mango alongside the company’s next text-based large language model.

Meta’s META 2.30%increase; green up pointing triangle chief AI officer Alexandr Wang talked about the artificial intelligence models in an internal company Q&A on Thursday with Chris Cox, Meta’s chief product officer, according to people who heard the remarks. The models are expected to be released in the first half of 2026.

Wang also said one of the focuses for the new text model, code-named Avocado, is making it better at coding, and that the company is in the early stages of exploring developing so-called world models, AI that learns about its environment by taking in visual information.

ChatGPT has hit a new milestone of $3 billion in worldwide consumer spending on mobile as of this week, according to estimates from app intelligence provider Appfigures. This figure represents the total spending on iOS and Android devices since the app’s launch in May 2023, when it first arrived, then only on iOS.

What’s notable is that the bulk of that spending took place this year. Worldwide, consumers spent an estimated $2.48 billion in the ChatGPT mobile app in 2025, representing a 408% year-over-year increase from the $487 million spent in 2024. In 2023, the app’s first year of availability, it earned $42.9 million, before growing 1,036% to reach the 2024 figure.

These numbers represent a sharp rise in consumer adoption, compared with other popular apps. For instance, it took ChatGPT 31 months to reach $3 billion in consumer spending, but the top earner, TikTok, took 58 months to do so, Appfigures told TechCrunch.

ChatGPT also reached the milestone faster than top streaming apps like Disney+ and HBO Max, which hit the $3 billion figure in 42 months and 46 months, respectively.

xAI’s Grok, however, is seeing a similar revenue trajectory to ChatGPT, compared with other AI rivals.

Grok was released in late 2023 to X Premium Plus subscribers before becoming more broadly available last year. But if you compare the pace of consumer spending across AI apps, Grok came the closest to matching ChatGPT’s cumulative revenue at the same point, once it began monetizing. (You can see this on the chart below, which plots the AI apps’ spending aligned to when each app began monetizing.)

While the $3 billion in spending represents a significant uptake by consumers, it’s not the only way to measure AI app adoption or potential long-term revenue.

ChatGPT’s mobile customers are buying paid subscriptions, like the $20 per month ChatGPT Plus or the $200 per month ChatGPT Pro for advanced users. But AI apps can generate revenue in other ways, including through developer offerings and, perhaps soon for ChatGPT, ads. In addition, ChatGPT on Wednesday launched its own app store of sorts, which the company suggests will be monetized in some way in the future, its blog post noted.

Google, meanwhile, is exploring the potential to transition its healthy search ads business to AI-powered search, placing ads in AI Mode, AI Overviews, AI shopping, and an increasingly AI-powered Discover page, among other things.

Anthropic is aiming at the business market and is reportedly on track for $70 billion in revenue by 2028.

Birkenstock's fiscal '26 financial targets are limited somewhat by foreign exchange headwinds and a temporary production capacity issue.

Birkenstock Holding Plc shareholders were disappointed with the brand’s fiscal 2026 targets, which were lower than 2025 results due to a combination of tariff headwinds, foreign exchange, and the brand’s own production capabilities.

Birkenstock CEO Oliver Reichert said the company posted “very strong fiscal 2025 results,” but investors — for better or worse — tend to focus just on future guidance.

For the fiscal year ending Sept. 30, 2026, the company said it expects revenue at between 2.30 billion euros to 2.35 billion euros, with adjusted earnings per share at between 1.90 euros to 2.05 euros.

The brand’s revenue target is lower than Wall Street’s projection of 2.39 billion euros. And its revenue growth rate in constant currency is about 13 to 15 percent. That represents a deceleration of the 18 percent growth in fiscal 2025, according to William Blair analyst Sharon Zackfia. She expects sales over the next few years to be bolstered by ramping production, DTC expansion, and white-space categories and geographies that could see growth, as well as a favorable product mix on premiumization and an expanded closed-toe assortment.

Shareholders responded by sending shares of Birkenstock down 8.6 percent, or $4.00, to $42.40 in mid-afternoon trading on the Big Board Thursday.

“These are unusual times,” Reichert told investors during a conference call after posting the fourth-quarter report. He said the company generated significant cash flow and invested over 150 million euros into its production capacity against the backdrop of “global tariff and international trade, a war in the Ukraine and energy crisis and a significant decline in the U.S. dollar.” The CEO also noted that Birkenstock’s healthy brand momentum continued in the fourth quarter, and that the brand is winning in both B2B and B2C, gaining shelf space and taking share.

He also noted a strong back-to-school season with retail sales “at our top 10 partners increasing over 20 percent year-over-year,” adding that the momentum is expected to continue through the holiday season, with over 90 percent of the growth in B2B from within existing doors. “Full price realization, the ultimate indicator for brand health and demand remains over 90 percent,” Reichert said.

While he noted that strong wholesale growth is driven by the younger demographic, that trend also “requires us to produce more pairs in a situation where we are already capacity constrained.” The CEO explained that the strongest demand is for Birkenstock’s premium styles, which require even more production units. “The combination of more wholesale and more premium execution is creating additional pressure on our vertically integrated supply chain,” he said. “We need to manage growth in our production responsibly. This is why we are steering towards a mid-teens pace of growth for fiscal ’26.”

There was nothing in Reichert’s remarks that suggested a slowdown in consumer demand. If anything, the constraints to growth were more limited to the brand’s production capacity. But the brand does have to deal with tough foreign exchange headwinds, as well as incremental U.S. tariffs, and that accounts for some of the pullback in guidance.

As for tariffs, the company said in May when it reported second-quarter results that it had taken actions to mitigate the tariff impacts and that there were still multiple levers it could pull, helped by the fact that the German brand is vertically integrated and had at its disposal efficiencies in production, vendor negotiations and the ability to allocate products between different regions.

Company executives said Thursday Birkenstock was able to offset most of the 2025 tariff impact through targeted price increases, including one implemented in July.

The one difference between 2026 and 2025 is that this past year benefited from shipments prior to the increase in tariffs. That means 2026 will see some impact from tariffs in the cost of goods sold, which is expected to result in a decline in margins.

Price reviews are done each season, with adjustments on a store-by-store basis. Over the longer term, growth in the brand’s Asia-Pacific business is expected to reduce its exposure to the U.S. dollar and to U.S. tariff impact. And while price increases — which are dollar neutral but not margin neutral — will offset much of the incremental tariff impact, Birkenstock executives gave the impression that higher prices would not rise to the level where they would capture the same margin rate as in 2025.

“We believe the guidance level sets expectations for the year, and the top-line outlook reflects continued strength in underlying demand for the brand against an uncertain macro outlook,” Telsey Advisory Group’s chief investment office Dana Telsey said. “As consumers remain more intentional with their purchases, we see Birkenstock as a brand that can continue to win in the marketplace through its competitive advantage of a high-end lifestyle positioning. As such, we maintain our Outperform rating.”

Back in August when the firm posted third-quarter results, Reichert said the shift to in-person shopping amplifies the brand and favors its B2B channel over DTC. He reiterated that during Thursday’s fourth-quarter call.

That’s one reason why the brand also plans to open about 40 new company-owned retail stores globally in fiscal 2026, about 10 more than the doors it’s opened this year. It ended the year with 97 doors, with Reichert noting that the new stores “are performing ahead of our expectations in terms of productivity and return of CapEx,” adding that the brand is “on track” to reach its 150-store target ahead of schedule. Meanwhile, the brand just opened its seventh company-owned retail store in the U.S. in Boston, in the city’s Back Bay neighborhood.

For the fourth quarter ended Sept. 30, the brand reported a 78.9 percent gain in profits to 93.9 million euros, or 0.51 euros a diluted share, from 52.5 million euros or 0.28 euros. Revenue rose 15.5 percent to 526.3 million euros from 455.8 million euros.

Wall Street was expecting adjusted diluted earnings per share of 0.25 euros on revenue of $407.09 million.

For the year, profits rose 81.8 percent to 348.3 million euros, or 1.87 euros a diluted share, from 191.6 million euros, or 1.02 euros. Revenue rose 16.2 percent to nearly 2.1 billion euros from 1.80 billion euros.

As demand for the Cybertruck can’t reach more than about 10% of Tesla’s planned production capacity, Elon Musk used his privately owned company to try to boost demand.

We now learn that SpaceX has bought tens of millions of dollars’ worth of Cybertrucks – potentially over a hundred million.

Elon Musk said that he expects Tesla to sell as many as 500,000 Cybertrucks per year.

Tesla actually planned to produce up to 250,000 Cybertrucks annually at the Gigafactory Texas. It never came even close to that.

The automaker is extremely opaque about its sales data, bundling Model S, Model X, and Cybertruck sales together.

However, based on registration data and historical split of Model S/X sales, we can estimate that Tesla is having issues selling even 20,000 Cybertrucks per year – less than 10% of its planned capacity.

By definition, the Cybertruck is a commercial flop.

Tesla boasted over 1 million reservations for the vehicle ahead of production, but it is estimated to have converted only about 60,000 of those reservations into orders since production began more than 2 years ago.

There are many reasons for this, but it is primarily because the Cybertruck costs much more than initially announced at the unveiling in 2019 and has less range and fewer cool features than the prototype.

SpaceX to the rescue

Earlier this year, we reported that Tesla started delivering truckloads of Cybertrucks to SpaceX and xAI, Elon Musk’s privately owned companies.

Now, a source familiar with the matter told Electrek that SpaceX bought over 1,000 Cybertrucks from Tesla and that it could ramp up to about 2,000 over time.

Hundreds of Cybertrucks can currently be seen parked in SpaceX’s lots in Southern Texas:

With a base price of $80,000, it would represent between $80 million and $160 million in sales.

It would be a significant help to Tesla’s performance in the fourth quarter, as the automaker is suffering from EV incentives ending in the US at the end of the third quarter, which remains Tesla’s most important market.

Electrek’s Take

SpaceX has been helping out Elon’s other companies quite a bit lately. It has reportedly committed to invest $2 billion into xAI, which is burning cash at an insane rate. Now, they are buying tens of millions to over a hundred million worth of Cybertruck, which are sitting in Tesla’s inventory, making its 4th quarter look even worse than it already is.

Can’t blame him here. This is legal. Although SpaceX investors might have concerns about how smart a purchase this is and what the utilization rate of those trucks looks like.

FTC: We use income earning auto affiliate links. More.

The AI adoption race in the workplace is on

Sector companies are seeking to show that time saved on one-off tasks can translate into real-world business value

A financial analyst turns to a chatbot to study the possible risks and returns from an investment programme. A job that would have taken 50 minutes is completed in 10.

That is one of the real-world cases that AI company Anthropic claimed to find recently when it asked its Claude chatbot to analyse how it was being used at work. But showing how one-off tasks like this translate into real-world business value for employers is not so simple.

This is set to become one of the main fronts in the battle between leading AI companies in 2026. OpenAI chief executive Sam Altman said recently that his company was shifting its focus to enterprise customers as it looks to boost its revenues — a market where Anthropic is currently ahead.

Yet even if people are starting to take to generative AI at work, most companies aren’t yet able to measure whether the technology makes individual workers more effective in their jobs, let alone trace any productivity gains at the level of the company as a whole.

Looking for the effects of a new technology like AI in the broader economy is even harder. IT’s impact on overall labour productivity has been famously difficult to identify from the official data. The impact of digital technologies did not show up in US economic data for years, until productivity growth started to rise steadily after the late-1990s. By the beginning of this decade, though, growth had fallen back to its earlier level, at around 1.5 per cent a year.

The encouraging news for the AI companies and their investors is that many people are starting to find uses for generative AI in their working lives. Summarising a long report, drafting a marketing presentation and analysing financial data are the kinds of things workers have tried for the first time this year. If and when any of these use cases take hold widely, the effects in terms of AI model usage could be significant.

So far, generative AI has one “killer app” at work, in the form of the coding assistants used by software developers. Its effects have been explosive. In May this year, 11 per cent of all the tokens generated by large language models were related to coding, according to a study of AI model usage by OpenRouter. By November, that proportion had soared to around 50 per cent.

Workers themselves certainly believe AI is starting to make them more effective. Earlier this month, OpenAI said workers it surveyed found AI saved them 40-60 minutes a day. That is up from the 2.2 hours a week that workers believed they were saving in a similar study conducted by the St Louis Fed a year ago.

Self-reporting like this is highly subjective, which makes Anthropic’s study of real-world tasks potentially more revealing. Based on 100,000 work-related conversations, Claude estimated it was slicing 65 minutes off the 85 minutes an average task would have taken.

Yet demonstrating virtuosity on individual tasks doesn’t translate directly into business advantage for customers, as Anthropic is the first to admit. The figures don’t show, for instance, how much extra work goes into checking the output from chatbots, or how overall quality affects the results.

Also, a single job may lead to more than one chat session. The ease and speed of getting a result out of a chatbot might lead workers to produce many more reports or emails, leading to a cascade of unproductive “workslop”. Nor can Claude tell what workers do with any time the technology has managed to save them.

Another drawback is that the task-based analyses that lie at the heart of most studies of technology’s impact on productivity fail to capture the reality of working life. For most people, work doesn’t fall into discreet, self-contained segments. Looking at single tasks in isolation, as Anthropic admits, doesn’t capture the tacit knowledge and personal relationships that affect how work is done, or the connections between different tasks.

That appeared to explain the counter-intuitive results of one study this year, which found that a group of experienced developers took 19 per cent longer to complete a task when they used an AI coding tool.

The full benefits of generative AI will only become apparent when companies have redesigned entire work processes to make best use of the technology, and when they have overcome the cultural barriers that always stand in the way of this kind of change. But with workers starting to take to experiment with AI, the race is on.

Law firms Hogan Lovells and Cadwalader set to merge in record $3.6bn deal

Combination that would deliver 3,000-lawyer behemoth is latest transatlantic merger

Hogan Lovells and Cadwalader, Wickersham & Taft are set to merge in the largest-ever law firm tie-up, creating a behemoth with combined revenues of more than $3.6bn.

The new firm, which will be known as Hogan Lovells Cadwalader, will become the world’s fifth-largest law firm by revenue and house more than 3,000 lawyers globally.

Hogan Lovells — which itself was created from the transatlantic merger of US firm Hogan & Hartson with the UK’s Lovells in 2010 — is by far the dominant party, with more than six times the number of lawyers and offices than New York-headquartered Cadwalader. Hogan Lovells chief executive officer, Miguel Zaldivar, will be CEO of the combined firm.

Partners will vote on the combination in the spring with the merger expected to go live as soon as June 2026, according to Zaldivar.

“This is history in the making,” Zaldivar told the Financial Times on Thursday. “We’re doing this in part to address our client needs . . . the days of smaller firms teaming up and multiple relationships [for each client with different firms], we see them as passing.”

“When you have a strong balance sheet — $3bn plus — when you have no debt, when you have a very well capitalised firm, when your brand has consistently improved . . . then the icing on the cake is getting New York right.”

Equity partners at Hogan Lovells earn an average of $3.07mn. Cadwalader’s profit per equity partner — a key metric to compare in any law firm merger — is $3.7mn.

Hogan Lovells will account for a greater proportion of the equity partners and a larger amount of the income of the combined firm, according to a person familiar with the deal’s terms.

The deal comes after Cadwalader, one of Wall Street’s oldest law firms and a pre-eminent name in finance, has lost tens of partners this year, including from senior management. 

The firm has also faced scrutiny over its decision to do a deal with US President Donald Trump in April, committing at least $100mn in pro bono legal services to the administration to avoid being targeted in a White House crackdown on Big Law earlier this year.

Pat Quinn, Cadwalader’s co-managing partner, told the FT that the firm was coming to the end of its second-ever strongest year by revenue and had hired more than 95 lawyers, including 10 partners, in the past 12 months. The firm predicts it will finish the year with revenues of $625mn.

The firms started talking about a possible deal in November 2024 before pausing conversations in March, Zaldivar said. Discussions restarted in autumn this year.

Zaldivar said the firms did not envisage major lay-offs from the combination, despite duplicate offices in London, New York and Washington DC.

The tie-up marks the latest in a run of large mergers in the legal industry as firms look to get a stronger foothold on both sides of the Atlantic and scale up in the face of pressure to make heavy investments in technology and artificial intelligence. 

UK law firm Ashurst and the US’s Perkins Coie agreed a merger last month, which will create one of the top 20 law firms globally by revenue, while America’s Winston & Strawn announced a tie-up with Britain’s Taylor Wessing this week for a 1,400-lawyer firm that will be known as Winston Taylor.

UK “magic circle” firm Allen & Overy and New York’s Shearman & Sterling also merged in 2024 to create A&O Shearman. The combined firm posted revenues of $3.7bn for the last financial year.

In this story that Sri, Amir and I published this morning, we revealed that over the past year OpenAI staffers had noticed that improvements to the company’s models weren’t necessarily lifting ChatGPT usage. Their quandary is that focusing too much on increasing the chatbot’s appeal could draw away resources from the company’s stated goal of developing artificial general intelligence, or AI that meets or surpasses human abilities.

Nevertheless, there are growing signs that the company is shifting its research strategy to be more aligned with its products. For instance, OpenAI hopes to release updates to its models for ChatGPT more frequently than in the past, as often as once a month, a current employee told me. Already, we’ve seen this trend with GPT-5.1 released in November and GPT-5.2 released this month.

This seems like a smart way to avoid building up too much hype around any particular model release—that can end up being a letdown (as we saw with GPT-5). But some current and former employees told me that they’re worried that OpenAI could become too focused on incremental updates to its models at the expense of the longer-term vision of what its AI should look like.

Other issues could arise if researchers focus too much on users’ initial reactions to changes in ChatGPT.

For instance, if a user is shown two versions of a ChatGPT response where one version is longer, they might choose the longer answer because people tend to assume longer responses are more accurate, a former employee said. In the long run, though, overly wordy responses can become annoying to users, the former employee said.

“Product and research are deeply interconnected, not oppositional,” an OpenAI spokesperson said in a statement. “Research breakthroughs shape our products, and product feedback shapes research. This is a single, unified strategy for building and safely deploying increasingly capable models, not a division between competing sides.”

For what it’s worth, a secretive new model from OpenAI, codenamed Garlic (whose existence we first reported here), could give the company an opportunity to realign its research and product teams. That’s because Garlic was created through improvements in pretraining, the first step of model training in which a model is shown lots of data from the internet and other sources so it can learn to make connections between them.

Pretraining is the part of the model development process that OpenAI has lately struggled to show improvements in. But getting better results in pretraining matters, because they flow through to the model’s performance in a variety of areas, from coding to creative writing to a model’s personality.

So theoretically once Garlic is incorporated into ChatGPT, it should improve the chatbot’s performance on a variety of areas users might ask questions about. It should also improve the chatbot’s personality, while also helping the model’s general intelligence, an area that researchers would care more about.

(Importantly, OpenAI’s recently-released GPT-5.2 isn’t based on the full-blown Garlic model, but rather, an early checkpoint of the model, according to a person with knowledge of the model, so we’ll likely see more improvement there still.)

One last note: somewhat counterintuitively, some of the information we’ve learned about ChatGPT usage has made me more optimistic that it still has a lot of room to grow. We learned during the reporting process for today’s article that many ChatGPT users simply don’t understand the full range of things they can do with the chatbot. That’s partly because the chatbot is so text-based that users might not realize ChatGPT can do things like analyze a picture of your houseplant to understand why it’s dying or look at a picture of your hand in mahjong and tell you whether you’re making an illegal move or not (both of which I’ve done in the last week).

If OpenAI can make ChatGPT’s interface more intuitive and visual, as leaders have promised, more users could discover these under-the-radar applications over time.

How OpenAI’s Organizational Problems Hurt ChatGPT

Over the past year, some OpenAI staffers noticed a concerning change in the way people who used ChatGPT were reacting to improvements in the chatbot.

In prior years, every time OpenAI made a big upgrade to the artificial intelligence that powers the chatbot, usage surged as people easily found ways to get useful responses out of it, one employee said.

But even as ChatGPT attracted more users this year, improvements to the underlying AI model’s intelligence—and the in-depth research or calculations it could suddenly handle—didn’t seem to matter to most people using the chatbot, several employees said.

The trend left employees scratching their heads. The company’s research team had spent months working on reasoning models that spent more time computing answers to complex questions about math, science and other topics than ChatGPT’s previous models. OpenAI bragged over the summer about how its AI had achieved gold-medal–level performance at the 2025 International Mathematical Olympiad, and in the fall it aced the 2025 International Collegiate Programming Contest.

Most of the questions users asked ChatGPT, though, didn’t take advantage of those types of improvements, according to some OpenAI employees.

OpenAI’s focus on “science, math benchmarks, frontier math, coding competitions…doesn’t seem to match who the typical ChatGPT user would be,” said Peter Gostev, AI capability lead at LMArena, a startup that evaluates AI performance and publishes widely cited results. Much of the time, ChatGPT users are “probably asking about pretty simple things, like movie ratings, where you wouldn’t need a model to think for half an hour.”

Data OpenAI published in September about ChatGPT queries seem to back up Gostev’s view.

That issue and other disconnects between OpenAI’s underlying technology and its products created an opening this year for OpenAI competitors such as Google to make gains against ChatGPT. As a result, OpenAI CEO Sam Altman earlier this month issued a “code red” warning to refocus the company on improving its chatbot so it can appeal to more people.

The change in the way ChatGPT users have reacted to new models powering the chatbot shows how the goals of OpenAI’s core AI research division, which develops its technology, don’t always serve the needs of ChatGPT, which drives most of the company’s revenue. It also shows how OpenAI’s reliance on the chatbot for revenue could become a weakness as Google and other competitors find various ways to add AI features to a wide range of consumer products, including search, mobile devices and productivity apps.

The idea that ChatGPT might not crater Google Search is a shift from a year or two ago, when executives at both OpenAI and Google believed ChatGPT could effectively replace the popular search engine. Google Search has since added AI-powered answers at the top of search results, and the company reported in October that the feature was driving “meaningful” search query growth and revenue as “users continue to learn that Google can answer more of their questions.”

OpenAI’s renewed ChatGPT push comes as the company is at risk of falling short of its ambitious user growth goals. At the start of the year, when it had 350 million weekly active users, OpenAI said it aimed to hit 1 billion this year. With less than 900 million weekly active users as of earlier this month, that goal will be tough to meet.

But OpenAI has done an impressive job of convincing more individual workers and businesses to subscribe to ChatGPT for extra features and unlimited usage. The company is now generating more than $19 billion in annualized revenue, largely from such subscriptions, up from $6 billion in annualized revenue in January. That suggests it will hit its goal, set in August, of reaching $20 billion in annualized revenue by the end of this year. The company is also on track to beat its 2025 revenue goal of $13 billion, up from around $4 billion last year, and is looking to raise funds at a $750 billion valuation, 50% higher than its valuation two months ago during a share sale.

These are big and enviable numbers by most standards. But to generate $200 billion in revenue in 2030, as OpenAI has projected, it will likely need to convince weekly users to use the chatbot daily. That way it will have more opportunities to sell ads that it has planned to run in the chatbot or to take a cut from the sale of products users find through the chatbot, employees and investors say. (ChatGPT has roughly 70% of assistant usage globally while also ranking as Apple’s most downloaded free app of the year and capturing 10% of search share in less than three years, a spokesperson said in a statement.)

That will require ironing out its organizational and product challenges. This year, OpenAI’s research group, which has more than 1,000 people and is mostly cordoned off from the rest of the company, has largely focused on developing reasoning models rather than improving AI specifically for the company’s chatbot, according to several employees.

But reasoning models don’t do much for ChatGPT, whose users generally want answers fast, according to a person who worked on them. Typically, reasoning models can take anywhere from several seconds to several minutes to answer a question—which can feel like an eternity to people accustomed to Google’s snappy search results. OpenAI says reasoning models are best suited for completing complex, multistep tasks, reviewing large amounts of code or finding specific information across a trove of company documents.

ChatGPT faces an even bigger problem than the issue with reasoning models: Consumers don’t appear to understand the full range of topics the chatbot can answer questions about, even when it’s using nonreasoning models that give faster answers. That limits how much time they spend using it, several employees said.

In particular, ChatGPT’s text-based design has made it harder for people to discover its other features, such as its ability to analyze images of mechanical or computer errors and provide advice on how to fix them. The current look and feel is somewhat akin to MS-DOS, the text-only Microsoft operating system for PCs used in the 1980s, said ChatGPT product chief Nick Turley. The PC revolution didn’t take off until Microsoft launched Windows, a more visually appealing and intuitive operating system, he added.

Other OpenAI leaders have echoed this sentiment, saying that ChatGPT’s interface will have to change to gain mass appeal. OpenAI’s head of applications, Fidji Simo, said in a blog post this week that ChatGPT is moving “from being primarily text-based and conversational, toward a fully generative UI that brings in the right components based on what [users] want to do.” On Tuesday, the company launched a new image-generation model for ChatGPT users and said it would incorporate more imagery in ChatGPT responses.

But Simo, who joined OpenAI from Instacart several months ago to oversee ChatGPT and other applications, seems aware of the limitations on her side of the company. She recently wrote on her blog that at its core, OpenAI remains a research-focused company and “products aren’t the goal themselves.”

Some rivals to OpenAI, whose main product primarily caters to consumers, don’t seem to be experiencing the same disconnect between their research and product efforts. Anthropic, for example, mostly focuses its research efforts on an application programming interface aimed at businesses. So far, the smarter Anthropic’s models get—especially in generating computer code—the more API sales it makes to other businesses and application developers. At OpenAI, improvements to its models also boost API sales, but they are a small fraction of its current and projected revenue.

For much of the year, Altman seemed to be running OpenAI as if it had already conquered the chatbot market, according to some employees. He embarked on countless other product efforts, including the Sora video app, music-generating AI, a web browser, specialized AI agents, a consumer hardware device, robots and more.

There’s a growing sense among OpenAI executives that ChatGPT could be vulnerable, which has rattled many of them, according to a person who recently spoke with leaders at the company. Altman, as part of his code red declaration, said he wants to move some employees back to working on ChatGPT.

“Product and research are deeply interconnected, not oppositional. Research breakthroughs shape our products, and product feedback shapes research. This is a single, unified strategy for building and safely deploying increasingly capable models, not a division between competing sides,” an OpenAI spokesperson said.

The divergence between OpenAI’s research and ChatGPT product groups emerged over the past year as the company struggled to improve its large language models through traditional training methods that had previously worked. So it focused its energies on reasoning models to eventually reach artificial general intelligence—AI that meets or surpasses human abilities.

Initially, the researchers thought reasoning models could also improve ChatGPT. But at the start of the year, when OpenAI converted its most advanced reasoning model to a version that users could use in ChatGPT, the model’s performance got worse. It turns out that creating a chat-based version of the model unintentionally dumbed the model down.

OpenAI still found ways to bring reasoning models into ChatGPT, where they currently power its Thinking mode and Deep Research agent—launched in February—that generates reports. (Reasoning models also help power Codex, OpenAI’s coding assistant app.) But relatively few of the chatbot’s nearly 900 million weekly active users regularly access those features, several employees have said.

It’s also far from clear that reasoning models will lead to AGI, according to some researchers at OpenAI.

A change OpenAI recently made to ChatGPT implied its reasoning models were becoming a liability for the chatbot. Earlier this month, the company quietly rolled back a feature in its free and lowest-priced ChatGPT subscription tier that automatically routed people’s questions to reasoning models.

OpenAI faces other hurdles in bringing improved models to ChatGPT: Even traditional, nonreasoning AI models can clash with ChatGPT features, such as efforts to personalize the chatbot so it provides different answers based on what it already knows about the person using it.

For instance, in the weeks leading up to its August release of GPT-5, a flagship LLM model powering OpenAI’s products, researchers found that the model’s performance on certain tasks such as coding questions declined after it was integrated into the chatbot, the current employee said. That’s because when users sent a query to ChatGPT, the chatbot would also use information it had about them, such as their occupation, to personalize the answer, the employee said.

But that personal information at times interfered with the model’s understanding of what the user was asking, leading to incorrect answers, the employee said. The company fixed the issue before it released GPT-5, but employees believe they’ll continue to face interference between new models and ChatGPT features.

There are other signs of schisms between research and product at OpenAI. Earlier this year, the company deprioritized the development of its image-generation model, which had temporarily boosted usage and user growth of ChatGPT in March, according to two employees.

It couldn’t be learned why it made that decision, but when Google in August released its own image-generating AI, Nano Banana, to much acclaim from consumers, leaders at OpenAI rushed to improve its image technology, one of the employees said. That prompted a disagreement between Altman, who felt image-generation capabilities were important to ChatGPT’s growth, and Mark Chen, the company’s research chief, who wanted to prioritize other initiatives, they said.

Earlier this month, Altman said in his “code red” alert that image generation would be a key priority in the push to improve ChatGPT. This week OpenAI released a new image-generating model for the chatbot.

In Google, OpenAI faces a competitor that has long held advantages in getting its AI in front of people, including through search, Chrome and its workplace apps such as Gmail. And in recent weeks, Google’s AI models have improved to be roughly on par with ChatGPT in terms of the questions or tasks they can handle adeptly, including generating images and computer code. That makes Google’s chatbot and other AI products more appealing.

Inside OpenAI, leaders are concerned the average chatbot user won’t see much of a difference between ChatGPT and Google’s Gemini, according to a person who has spoken to them. And unlike social apps like Facebook and Instagram, which had staying power because the products got better as the number of people that used them grew, such network effects generally don’t exist in chatbots.

OpenAI is at another disadvantage: its finances. It is burning through billions of dollars a year in cash as it rents more and more servers to train and run its AI, including ChatGPT. While OpenAI is endeavoring to develop its own data centers and server chips to lower those costs over time, Google today can operate more efficiently because a decade ago it began developing specialized servers for AI.