OpenAI president Greg Brockman says the share of code at the company written by agentic AI tools quadrupled inside a single month, jumping from 20 percent to 80 percent over December. The disclosure, made on a recent podcast appearance, is the most aggressive internal adoption number any frontier lab has put on the record and lands as the industry argues over how close large models are to replacing professional software engineers.
"Oh, even over the course of December, we went from these agentic coding tools writing 20% of your code to writing 80% of your code," Brockman said, according to host Jaeden Schafer on AI Chat Daily. Schafer flagged the figure as "a 4x jump in one single month" and tied it to a broader claim from Brockman that OpenAI is now "70 to 80% of the way to AGI." Brockman echoed Sam Altman and Google DeepMind chief Demis Hassabis in arguing the field is roughly two breakthroughs from artificial general intelligence.
The numbers were not received quietly. Meta chief AI scientist Yann LeCun, a long-running AGI sceptic, has spent the week pushing back on Brockman's timeline on X. Schafer noted that LeCun "predictably has been dunking on the AGI claim" and has been basically doing this all week, framing the dispute as the latest round of a familiar fight between OpenAI's leadership and rivals who think large language models are nearing a ceiling.
Key facts
- 01OpenAI president Greg Brockman says agentic coding tools went from writing 20% of company code to 80% over the course of December, a 4x jump in one month.
- 02Brockman also claims OpenAI is 70 to 80 percent of the way to AGI, and aligns with Sam Altman and DeepMind's Demis Hassabis that the field is roughly two breakthroughs away.
- 03Andrej Karpathy counters that the 20-to-80 figure measures tokens emitted, not the share of code that becomes merged commits a senior engineer would approve.
- 04Meta chief AI scientist Yann LeCun has spent the week publicly dismissing the AGI timeline claims on X.
A more technical critique came from former OpenAI and Tesla researcher Andrej Karpathy, who Schafer described as "a lot more, I guess you could say, like measured." Karpathy's argument is that the 20-to-80 framing flatters the model: it counts tokens emitted, not code that survives review. The harder question, he has argued, is what fraction of those AI-written tokens become merged commits a senior engineer would actually approve.
“Oh, even over the course of December, we went from these agentic coding tools writing 20% of your code to writing 80% of your code.”— Jaeden Schafer
That distinction matters for how investors and engineering leaders read the metric. A model that drafts most of a pull request but still needs heavy human rewriting is a productivity tool; a model whose output ships unedited is a labour story. Brockman's number, taken at face value, implies the second. Karpathy's reading keeps it firmly in the first camp, with humans still doing the structural work on top of machine-generated scaffolding.
Schafer, who said he has been building software apps with the tools despite not being a developer, sided with the productivity reading. He compared the workflow to sculpting: "I feel like I'm starting out with a giant block of granite, and I'm chiseling it down to be what I want it to be." The AI, in his telling, does not produce a finished statue on command; the user iterates, flags bugs and rewrites until the code works.
His bottom line was that the debate over the precise percentage misses the point for most users. Even when AI code needs editing, monitoring or rework, he argued, "you're still going 100 times faster than if you're doing it from scratch." That framing tracks with what GitHub, Anthropic and Cursor have all reported from their own usage data: raw acceptance rates are noisy, but throughput gains for individual developers are real and large.
For OpenAI, the strategic value of Brockman's 80 percent claim is less about engineering bragging rights than about positioning ahead of the next funding and product cycle. If agentic coding inside the company really has scaled that fast, it gives Altman a concrete data point to wave at enterprise buyers weighing Codex-style products against Anthropic's Claude Code and Google's Gemini coding agents. If Karpathy is right that the figure overstates merged output, the gap between demo and deliverable is the story competitors will press on next.
Working on something we should cover, or seeing a story we missed? Send leads, documents, or feedback to hello@aichatdaily.com. For sensitive tips, see our secure tips page for Signal and PGP options.
Spotted an error? Email hello@aichatdaily.com with the URL and the issue, or read our full corrections policy.
