OpenAI's big launch — and bigger departure
Platformer Casey Newton ● Covered by 7 sources
OpenAI shipped GPT-5.6 in three tiers this week, and testers are impressed. But its president is stepping down, adding to a long 2025 exit list.
Based on reporting by Platformer, Casey Newton — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Thursday was a busy day for OpenAI, and not just because of one release. The company folded Codex into its desktop app, rolled out a new ChatGPT Work agent, quietly retired the Atlas browser, and — almost as an afterthought amid all that shuffling — dropped GPT-5.6 in three flavors named Luna, Terra, and Sol. It landed just days after a voice model that analyst M.G. Siegler called proof that computing is finally tilting toward talking instead of typing, the kind of shift that makes OpenAI's hardware bets look less speculative and more inevitable.
The reception for Sol, the top-tier model, has been genuinely warm. OpenAI's own benchmarks show it beating Anthropic's Claude Fable on tests like Agents' Last Exam and the Artificial Analysis Coding Agent Index. Every's Katie Parrott called it her favorite model to work alongside, even while admitting Fable still wins the jobs she wants to fully hand off. Box CEO Aaron Levie praised its leap on data-heavy reasoning tasks, and Ghostty creator Mitchell Hashimoto said he'd reach for Sol most of the time, saving Fable for narrow debugging work with a clear scoring rubric. Figma's Dylan Field pushed back on the whole comparison exercise, insisting the two models solve different problems entirely. Wharton's Ethan Mollick offered maybe the sharpest framing: whatever your preference, Sol and Fable have pulled far enough ahead that anyone doing serious knowledge work is now choosing between exactly two options, not a dozen.
And then, on the same day, came word that Fidji Simo — OpenAI's applications CEO and effectively its second-in-command — is stepping back. She told staff her chronic illness, postural orthostatic tachycardia syndrome, has worsened, and she'll move into an advisory role less than a year after joining. It's a legitimate, sympathetic reason to step away. But it also slots her into a roster of high-profile exits that's grown uncomfortably long this year: chief futurist Joshua Achiam, research VP Jerry Tworek, former product chief Kevin Weil, enterprise sales lead Barret Zoph, model behavior lead Joanne Jang, and researcher Max Schwarzer.
None of that turnover shows up in a benchmark chart, which is exactly the problem with judging AI companies purely on release day. GPT-5.6 looks like real progress, and the people testing it aren't just being polite. But a company can ship a great model and still be hollowing out at the top, and OpenAI is currently doing both at once.
My take — AI-written commentary, not fact-checked reporting
My fiancé works at Anthropic, so take this with the appropriate grain of salt, but I'll say it anyway: a company can't out-benchmark a leadership bleed forever. Six senior departures in a year isn't bad luck, it's a pattern, and eventually the org chart matters as much as the eval scores. OpenAI keeps winning the week it ships something, then losing the week after — that's not stability, that's momentum covering for something else.
Read more about this at: Platformer