Today’s Codex will feel “primitive” by fall — and its own team’s roadmap backs it up

Thibault Sottiaux doesn't run some rando dev shop — he leads core products at OpenAI, so when he posted on X late Monday that Codex is "a good harness" but will feel "primitive in 2-3 months," people paid attention. His follow-up line was the real tell: "the next generation of models need more than your laptop." No roadmap, no dates, just a warning shot from inside the building.

The timing lines up with something OpenAI already announced. Back in June, the company said it plans to buy Ona, formerly known as Gitpod, a firm that builds secure cloud dev environments already used by 2 million developers. OpenAI framed the deal as the "next phase of Codex" — one where an agent keeps grinding away in a customer's cloud long after the laptop that kicked off the task gets closed and stuffed in a bag. Right now, Codex leans on cloud compute but still often needs that laptop nearby to reach a project's files and tools. Kill the connection, and the agent can lose its footing mid-task.

The ambition here isn't small. OpenAI showed in a February test that Codex could run solo for about 25 hours, burn through 13 million tokens, and spit out roughly 30,000 lines of code building a design tool from nothing. Impressive, until you clock what Alibaba just did: its Qwen3.8-Max agent worked autonomously for 16 straight days, landing 265 commits without a single human nudge. That's the bar apparently being set for late 2026, and it's the kind of endurance no laptop session survives.

Getting there means solving a much messier problem than model quality. An agent with standing access to a company's network, credentials, and CI/CD pipeline is a genuinely scary attack surface, and OpenAI knows it — hence the promise that Ona's setup stays customer-controlled, with OpenAI supplying only the model and orchestration layer. Anthropic is chasing a similar itch with its Mendral acqui-hire, aimed at automating flaky-test cleanup and dependency reviews. Everyone building these persistent agent environments is quietly rediscovering that autonomous coders need identities, audit logs, and access rules — basically the same governance headaches human engineers already have, just handed to software that never clocks out.

So Sottiaux's prediction isn't really hype about a smarter model dropping soon. It's an admission that Codex's current shape — bound to a session, tethered to a machine — is already the bottleneck, and OpenAI is racing to build the infrastructure that removes it before someone else does.