Greg Brockman: AI wrote 80% of code

Software engineering just got a new baseline. Greg Brockman said at Sequoia’s AI Ascent 2026 that OpenAI saw AI coding tools jump from writing about 20% of code to 80% in a single month — December. Around the same time, Snap CEO Evan Spiegel said more than two-thirds of Snap’s new code is now AI-generated, and Google said last week that 75% of its new code is AI-generated and then approved by engineers. (thenextweb.com) ### What exactly did Brockman say? The striking part is the speed. Brockman wasn’t describing some slow multi-year rollout. He said OpenAI crossed from AI being a side tool to being “the main thing” engineers do, with the jump happening over December. That matters because it frames this less as autocomplete getting better and more as agents taking over the first draft of software work. (tech.yahoo.com) ### Is this just an OpenAI story? Not really. Spiegel said on a recent podcast that more than two-thirds of new code at Snap is now written by AI, and he singled out Anthropic’s Claude as a major driver inside the company. Google made the same point from a different angle — Sundar Pich(tech.yahoo.com) fall. Three big companies are describing basically the same operating model. (msn.com) ### What changed so fast? The tools stopped being just code completion. OpenAI’s recent Codex updates are built around longer-running agent workflows — editing files, running commands, checking tests, reviewing pull requests(msn.com)oves. The limiting factor is less “can it type code” and more “can a human define the task and verify the result.” (openai.com) ### So what is the engineer doing now? More orchestration. More review. More boundary-setting. Google’s phrasing is useful here — AI-generated and approved by engineers. That approval step is the whole game. If an agent writes most of the code, the human job shifts toward specifying what should happen, checking whether the(openai.com)troduce security holes, flaky tests, or weird edge-case behavior. (blog.google) ### Does “80% of code” mean 80% of the work? No — and this is the catch. Writing lines is not the same thing as understanding product requirements, choosing architecture, debugging ambiguous failures, or deciding what should be built at all. These company numbers tell y(blog.google)ignals, not as a clean measure of total labor replaced. The real shift is that typing is getting cheap. Taste and verification are not. (thenextweb.com) ### Why are executives talking about this now? Because the economics are changing in plain sight. Spiegel’s point was that if software becomes easier to produce, companies can spend relatively less on building and more on distribution — getting products in front of users. That is a very diff(thenextweb.com)s faster, attention becomes scarcer than implementation. (msn.com) ### What does this mean for everyone else? “I use AI” is no longer the flex. That’s table stakes. The differentiator is whether a team can turn AI output into reliable software without drowning in junk. The teams that win ar(msn.com) “AI writes 80% of code,” but it’s probably the part that lasts.