The New York Times has severed its ties with a freelance writer after discovering that he used A.I. to generate at least part of a recent book review—which bore a noticeable resemblance to an earlier review published in The Guardian. The news is fresh fuel for the ongoing A.I.-journalism debate, kicked off by the revelation that a Fortune writer has been using A.I. to churn out an alarming number of stories, and a Wired piece about how a cohort of independent tech reporters uses A.I. for everything from editing and idea generation to full-on article drafting. This problem isn’t going away…

And now for the main event…

In the months since OpenAI issued its “code red” warning—ostensibly intended to galvanize its troops to beat back the increasingly threatening posture of Google—the startup has secured a fresh $122 billion in funding, cracked open the advertising floodgates, and suddenly found itself playing defense against a younger, smaller adversary: Anthropic. Though they are few, OpenAI’s competitors seem to be closing in on all sides.

But the latest panic inside OpenAI largely comes down to two words: Claude Code. Late last year, Anthropic shipped a new model release that seemed to kick off a revolution in coding, bringing a degree of (mostly basic) web development to non-coders. By the time Anthropic announced the closure of its staggering $30 billion Series G round in February, the company had grown its run-rate revenue to $14 billion, a 14x year-over-year expansion that it credited to its breakthrough product.

Anthropic’s quantum leap reminded the market of the disproportionate value that enterprise technology solutions, with their professional audiences and recurring revenue opportunities, have over consumer products. And, in many ways, the release of Claude Code seemed to crystallize the strategic divergence between the two behemoths: Anthropic was all in on B2B, a potential Microsoft of its day; OpenAI, on the other hand, appeared to have scattered its focus across a messy (and lawsuit-inducing) consumer product landscape. Its “instant checkout” feature, which launched last fall and enabled ChatGPT users to buy stuff without leaving the app, never really worked and was killed in March. Plans for a special “adult mode” for ChatGPT were also recently nixed, according to the Financial Times. The much-hyped device that OpenAI has been developing with Jony Ive was pushed to at least February 2027, according to court documents. Then, of course, came the decision to sunset Sora, its generative A.I. video platform, which also unceremoniously ended its partnership with Disney.

Anyway, OpenAI got the market signal and responded diligently. Earlier this month, The Wall Street Journal reported that the company had decided to pursue a shift in strategy, simplifying its user experience and slimming down on “side quests” in order to reorient the business around what has been working for Anthropic: enterprise customers and developers. “Companies go through phases of exploration and phases of refocus; both are critical,” OpenAI’s C.E.O. of applications, Fidji Simo, posted on X. “But when new bets start to work, like we’re seeing now with Codex, it’s very important to double down on them and avoid distractions.”

In this fast-changing landscape, however, OpenAI is now further behind than many realize. In December, companies purchasing A.I. tech for the first time were picking OpenAI over Anthropic at a 60–40 split, according to data from Ramp. Now Anthropic is capturing 73 percent of that market. Though this shift began before Anthropic’s public defiance of Pete Hegseth’s Department of Defense—juxtaposed with Sam Altman’s opportunistic compliance—the conflict fueled a positive P.R. super-cycle that seems to have accelerated the momentum in Anthropic’s favor.

Of course, early adopters are going to keep switching models until there’s a material performance gap or the cost of switching becomes too onerous—something developers are certainly aware of. “Everybody is now sort of realizing, We need to worm our way into every bit of people’s workflow and build integrations everywhere so it’s a giant pain in the ass to switch,” Will Wilson, the C.E.O. of Antithesis, a software testing company, told me. “Because I think the true answer is that model quality is not going to be enough to keep people loyal, so long as they’re racing neck and neck every day.”

Everyone on Peter Grafe’s team at BlueAlpha, an A.I.-native marketing firm, is on a paid Claude subscription after switching from ChatGPT. Despite his willingness to follow the best model capability—“I believe every tool excels in a specific area,” he said—he surmised that Anthropic is starting to build a moat. “Especially now with setting up Cowork and having different agents run—it’s becoming way stickier.”

Likewise, Joe Tyler, an A.I. researcher at the code review firm Sonar, told me that the new features Anthropic’s been adding—including plug-ins for finance, cybersecurity, and general enterprises, better integrations with Excel and PowerPoint, and code review features—“could be interpreted as moat-building. For long-term stickiness, it remains to be seen how sustainable the costs prove to be.” He expects that, over time, enterprises and individuals will become a bit more interested in open models and open agents—a path that A.I. coding startup Cursor has taken—reducing their reliance on both Anthropic and OpenAI.

This was all echoed by PwC’s A.I. factory leader, Jacob Wilson, who told me that there “is not a durable moat around A.I. coding tools in the way we once thought about platforms like search.” Within PwC, and across “many” of its clients, he described A.I. coding tools “as one of the most tangible early wins in enterprise A.I.”

That said, there are almost certainly going to be market oscillations. To wit: Serious engineers are increasingly convinced that OpenAI’s Codex has edged out Claude Code, at least for now. “In terms of closing the coding gap, I think with Codex 5.4 they arguably already have,” Will Wilson said. “The thing about these systems is that they are scarily interchangeable—scarily for the providers. So if there’s an enduring performance gap in either generation, you’re just going to see everybody switch.”

The next key question facing the industry, however, is model reliance. Even as Anthropic and OpenAI race to cement their lead in the automation of software engineering, smaller players are working toward automating the construction of A.I. models themselves—a push intended specifically to reduce, or even eliminate, dependence on the frontier labs. Breaking that oligopoly is an early focus for Mira Murati, who left OpenAI to launch Thinking Machines Lab, a startup that raised record amounts of funding without a product in sight. Now they have one: an A.P.I. called Tinker that provides engineers access to the infrastructure required to customize and fine-tune openly accessible models.

“It’s interesting that even the ex-C.T.O. of OpenAI defects from OpenAI and builds a company to help enterprises build custom open models,” Manos Koukoumidis, a former software engineer who led development teams at Microsoft, Meta, and Google, told me. “All these big labs know that there’s no silver bullet. If you want to get better results without one day being pushed to the corner by these providers, you need to build your own specialized models.”

Koukoumidis, who launched Oumi, an open A.I. platform, last year, wants to take things even further. This morning, he launched a new addition to Oumi that will enable businesses to use A.I. to… build A.I. The new platform, according to Oumi, will allow enterprises to develop and deploy small, custom-trained A.I. models, all through prompting, effectively automating away the A.I. engineer. His gamble is that open A.I. will eventually beat OpenAI. After all, specialist models can offer lower costs, better control, and data privacy—all without being subject to, say, the rate limits imposed by the biggest frontier labs.

The frontier developers are betting on the idea that it’ll stay so costly to remain at the frontier that open models can’t keep up. But, as Will Wilson pointed out, such a scenario gives all the advantages to Google, which probably isn’t great for OpenAI or Anthropic in the long term. And if model performance plateaus and the cost of inference plummets, then, well, “none of these people make any money.”

That’s all for today. I’ll see you on Thursday.

Ian

Need help? Review our FAQ page or contact us for assistance. For brand partnerships, email ads@puck.news.

You received this email because you signed up to receive emails from Puck, or as part of your Puck account associated with {{customer.email}}. To stop receiving this newsletter and/or manage all your email preferences, click here.

Puck is published by Heat Media LLC. 107 Greenwich St., New York, NY 10006