OpenAI takes aim at Anthropic's coding mojo
Execs took shots at rivals during the Codex press briefing. Also: A new ACCESS ep, notes on the SpaceX/xAI merger, and news from my interview with Google's head of search.
I took a break from the usual newsletter last week to post three of my Sources Live interviews from Davos. Tomorrow, I’ll share my full conversation with Reflection AI co-founder Ioannis Antonoglou. I still have a few more convos from Davos to share in the coming days as well. Otherwise, I’m back to regular programming.
Three things:
Are we about to see more of an xAI exodus? As part of the merger with SpaceX, “shares in xAI will be converted to 0.1433 shares of SpaceX stock,” and xAI employees will be able to sell their shares back to the combined company. An odd thing to allow right before an IPO…
Something I missed over the weekend: SpaceX told the FCC it wants to launch as many as one million satellites to function as “orbital data centers” and make Earth a “Kardashev II-level civilization.”
OpenAI employees have a term for when Codex operates independently during an internal meeting: “babysitting.”
“But if you really want to do sophisticated work…”
Sam Altman doesn’t join most OpenAI press briefings these days, so I took notice when he joined the call I attended for today’s Codex Mac app release.
Anthropic’s Claude Code is having its “ChatGPT moment,” as one reporter noted on the call. OpenAI, meanwhile, wanted to remind everyone that it’s still very much in the AI coding race. Altman called the Codex Mac app “the most loved internal product we’ve ever had.” He said more than a million developers used Codex in the past month, with usage up 20x since August.
During the briefing, OpenAI leaders highlighted Peter Steinberger’s viral success in using Codex to develop OpenClaw, the autonomous agent that has quickly become one of the most talked-about AI tools in the industry. Despite using Codex, Steinberger originally named the project Clawdbot — a play on Anthropic’s Claude. (Anthropic’s legal team promptly requested a name change. After a brief stint as “Moltbot,” the project landed on OpenClaw.)
When a reporter asked directly about Claude Code’s momentum, engineering lead Thibault Sottiaux emphasized Codex’s ability to work on multiple, long-horizon tasks autonomously. “One of the things that our models are extremely good at is that they really sit at the frontier of intelligence and do reliable work for long periods of time,” he said. “Whereas other models we are not building might require a lot more handholding, and you can only have one thread going at a time.”
“There are many tools out there that can do a good job of vibe coding front ends,” product lead Alexander Embiricos added. “But if you really want to do sophisticated work on something complex, 5.2 is the strongest model by far.”
Then Altman chimed in: “We have a model that I think people agree is the most intelligent, smartest model. If you go on Twitter and you search the competitor name plus Codex at the same time, you’ll often see anecdotes of people saying, ‘I’ll use this other tool, but then I have to send Codex into the code base to find bugs and fix bugs or make the thing work that wasn’t working beforehand.’”
One goal of the call was to emphasize that OpenAI itself is now running on Codex. Sottiaux said four engineers shipped the Sora Android app using Codex in eighteen days. The majority of the code for OpenAI’s Atlas browser was written by Codex. “There is no screen within the Codex engineering team that doesn’t have Codex running,” Sottiaux said.
While rumors swirl about Anthropic releasing a new Sonnet model as soon as this week, Altman made sure to also tease that OpenAI is focused on coding gains for what sounds like its own imminent model release: “You should expect us, with upcoming models, to continue to drive coding performance while also improving the warmth and clear thinking, clear expression, and writing that people like.”
Where do we go when the bots take over?
On the most recent episode of ACCESS, Ellis Hamburger and I chat about Moltbot mania, the scary aspect of AI agents, and the latest essay from Anthropic CEO Dario Amodei.
Then, we’re joined by Sublime founder Sari Azout to discuss what she learned from the crypto bubble, taste as a differentiator in the age of AI, founder life in Miami, the limits of “second brain” tools, and what it actually means to design for curiosity on the internet.
Like and subscribe on YouTube, Spotify, or Apple Podcasts.
Elsewhere
Google’s head of search dishes on its new publisher program for AI Overviews: I managed to get some details about the new pilot with a handful of news outlets from Google Search boss Liz Ried during an interview this morning at the IAB Leadership Summit in Palm Springs. Rather than the start of a broad rev‑share regime, she framed the new deals as narrowly focused pilots around “specific rights,” such as richer display of content than fair use permits, real‑time data feeds, and metadata or signals that go beyond what Google can obtain through ordinary crawling.
OpenAI and Nvidia are good, OKAY? This report from Reuters, saying that OpenAI has become dissatisfied with Nvidia’s inference speeds, has ruffled feathers from San Francisco to Taiwan. Even Oracle felt the need to weigh in.
Don’t get one-shotted by the Moltbook hype: Aside from the many fake screenshots that have been getting shared on X, there’s strong evidence that the social network for AIs is, in fact, being puppeted by a handful of bots that have been seemingly tasked with creating the illusion of a growing hive mind.
OpenAI is building out its ads team under Vijaye Raji and another former Facebooker, Asad Awan. One of the job listings for the new team seeks senior engineers to “build OpenAI’s foundational ads ranking and recommendation systems.”
Waymo is worth half what xAI was worth in the SpaceX merger. Groq this: A money-losing, frontier-ish AI lab with a penchant for perviness is apparently worth $250 billion in the SpaceX transaction, but the most scaled and reliable self-driving company on earth is worth $126 billion in its fundraising round today.
ICYMI: My interview with ElevanLabs CEO Mati Staniszewski
Sources is a newsletter by Alex Heath about the AI race, featuring scoops, unique analysis, and exclusive interviews. Every week, Sources is read by thousands of decision makers in tech, finance, policy, and media. Click here to learn more.