OpenAI named its next model in the strangest way possible
On August 1, OpenAI published ten advances in mathematics and theoretical computer science. The news was sitting in the third paragraph: the work was done by an internal version of a model called Astra, which OpenAI describes as its next major model.
These weren't graduate level math problems. These are proofs and problems not solved by anyone… ever (and yes, OpenAI released formal Lean certificates so mathematicians can check the proofs themselves). One decorated mathematician said on Twitter “this is likely the most significant day in the history of mathematics.”
This doesn’t change any work you’re doing now. It’s really just a signal in the noise that models are moving faster, and knowing what to use and when to use it matters most. It’s likely this will be GPT-6, and by all accounts will be released in the next two weeks (almost certainly before the next issue of The Huddle).
CHAT, WORK, or CODEX? (AND WHY THE ANSWER LANDS ON YOUR BILL)
Since ChatGPT was rebuilt on July 9, this is the question I get more than any other: "Which one am I supposed to be in?"
It's a fair question. OpenAI put three doors next to each other, labeled them Chat, Work, and Codex, and left a lot of teams guessing which one to open.
What each one is actually for
CHAT is the one you know. Ask a question, get an answer, keep talking. Use it for search, brainstorming, a quick rewrite, generating a flyer for the company picnic, or thinking out loud.
WORK is where you hand over a job instead of asking a question. It can break the task into steps, use connected apps and files, and come back with something finished: a spreadsheet, deck, document, a report or even a full-fledged website. You review the output the way you'd review a direct report's work instead of steering it message by message.
CODEX is still aimed most directly at software and technical work. If your team is not building or maintaining software, Chat and Work are the two doors to understand first.
I owe you a correction here. In Issue 10, I said Work should replace ChatGPT in your life. That was too strong. They are modes for different jobs, not versions of the same thing. Using Work for a question you could have asked in Chat is an expensive way to learn the difference.
Why the confusion costs credits
For credit-metered usage on Business and Enterprise (which includes all Trailblaze clients), OpenAI uses two rate structures, and the distinction is easy to miss.
In Chat, the everyday thinking model uses 10 credits per message. Its Pro version uses 50. Deep Research uses 50 credits per task, and agent mode uses 30 per message. Instant is listed as unlimited usage.
One setting is especially easy to misread: Medium, High, and Extra High all use the same 10 credits per message. If anyone on your team has been choosing a shallower setting to protect the allowance, they have been buying a worse answer at the same rate.
Work uses token-based pricing instead. Credit use depends on how much you feed the task, how long it runs, and how much it writes back. A single ambitious ChatGPT Work job can use more credits than a week of ordinary questions, so matching effort to the job still matters there. Teams are still deciding how to distribute tokens, how to track them, and how to funnel the intelligence to the team members chasing work that matters.
THE OTHER CLOCK
What we’ve covered so far has to do with the clock on the AI industry’s scoreboard. But it’s important to talk about the clock on your personal scoreboard too…
I've watched this happen in more than one client meeting. Interview 10 different people at a company about where the company is on their AI journey and adoption curve, and you get ten different answers.
- Some people on the team are afraid to admit they struggle with the basics
- Some people on the team are overwhelmed at the sheer potential of AI and wish they had 10 hours a week just to take a breath and think about how they could work smarter
- Some people on the team think they’re training their robot overlord replacements without knowing it
- Some people on the team are ready for more AI firepower but living with restrictions and policies that were written 18 months ago when the technology was 5% as capable as it is today
- Some people on the team have harnessed the potential of AI but want to keep it a secret to advance their own careers
-
Meanwhile the coaching staff (leadership team) has no way of knowing who is in what camp and what real progress is. I’ll say it for the people in the back - you are not alone!
For leaders: You owe it to your people to understand all of these feelings and approaches are almost definitely present in your company. And a new AI policy or more token budget isn’t going to solve the problem. If a policy can’t solve it, what can? Reply to the email, we want to help.
For team members: If you are ahead, keep taking reps inside the rules your company already has. Use your own workflows, non-sensitive material, and the work you control. It may not be long until your company looks to you to help lead the culture change required to do this right.
If your company is ahead of you, its timeline does not have to become yours overnight. Pick the one task you do every week and get good at handing that one over. One rep at a time.