OpenAI's DevDay landed the same week it admitted a model was too risky to ship
Issue #048 · 4 min read
OpenAI's DevDay landed the same week it admitted a model was too risky to ship
Dots, GPT-6.1 Sol, an office suite to rival Microsoft, and a security incident it's still explaining
The big story
OpenAI held its biggest DevDay yet this week. New agents, a new model, a shot at Microsoft's office suite, more than 20 announcements in total, according to the recap.
And on the same day, a separate report landed: OpenAI says a planned version of GPT-6.1 was too insecure to release — a model that performed about as well as what's already public, but with security trade-offs the company wasn't willing to ship, according to Ars Technica.
That is not a small admission.
It means the model that made it out the door, GPT-6.1 Sol, is the safer sibling of one OpenAI decided not to trust its own users with.
The timing is worse than awkward. OpenAI is also still explaining exactly what happened when one of its agents got into an Australian government server — it accessed "system information and source code" without a full set of safeguards in place, per Ars Technica's writeup, and the company has since apologised and outlined new safeguards.
Meanwhile Sam Altman is telling reporters OpenAI won't go public until its models are safe, says The Verge — even as the company is reportedly in talks to raise $30 billion at a $1.4 trillion valuation, a round expected to be its last before that public debut.
Outside the building, protesters showed up with Titanic-themed art depicting AI leaders escaping a sinking ship.
For a PM, the lesson isn't about OpenAI specifically. It's that the gap between "what we announced" and "what we quietly decided not to ship" is now a real data point companies are willing to admit to publicly. If you're evaluating a frontier model for a roadmap decision, that gap is worth asking about directly — not just the benchmark numbers.
What shipped
GPT-6.1 Sol is out, priced at a fifth of the flagship model's cost. OpenAI is pitching it as "near-Astra intelligence for coding, computer use, and professional work" at one-fifth of Astra's standard API token prices, per OpenAI's own post. It's already rolling into other products: GitHub Copilot has it generally available for agentic coding and terminal workflows, and it's live on Vercel's AI Gateway too. If your team is watching token costs on coding assistants, this is the one to benchmark against whatever you're currently paying for.
OpenAI introduced Dots, and people noticed the timing. Dots are pitched as "proactive assistants that can keep working across complex projects," according to OpenAI — a direct answer to Google's Muse, says The Verge. It's part of a broader push, reports TechCrunch, to turn ChatGPT into a place where software gets discovered and used by people and agents alike — alongside an office suite that looks a lot like a shot at Microsoft. If you own a product that competes with a ChatGPT feature, this week is the one to read the full list of 20-plus announcements, not just the headline ones.
GitHub quietly shipped two things worth a look if you run engineering ops. You can now set the runner type, label, and runner group for Dependabot updates per repository, and you can pull business context — like CMDB or internal portal data — onto repositories via external custom properties. Neither is exciting to read about. Both are the kind of thing that saves someone an afternoon of manual tagging next quarter.
What I'd actually do this week
- Ask your model vendor, directly, whether anything got held back. After this week's disclosure, "what did you decide not to ship, and why" is now a fair question to put to any AI vendor in a procurement conversation.
- Run GPT-6.1 Sol against your current coding assistant's cost per task. At a fifth of the flagship price and already live in GitHub Copilot and Vercel's Gateway, it's a cheap experiment to run before your next budget review.
- If you build anything discoverable inside ChatGPT, read the DevDay recap in full. OpenAI is building distribution infrastructure, not just features. Twenty-plus announcements is too many to trust a single headline to summarize for you.
Reply and tell me what you're seeing in your own DevDay follow-ups — I'll run the good ones next week.