I spent three days unable to work because I ran out of Codex tokens.
That sounds absurd written down. It felt worse. My organization runs on both Claude Teams and GPT Teams, which not everybody needs to do, but we do. On the GPT business side you can upgrade a single seat to premium for $100 a month, and that buys you a 5x multiple on tokens. The only route to the 20x multiple was a $200 personal pro plan. Power users have been quietly stacking three or four of those and rotating between them. OpenAI closed the ability to open new $200 accounts. So when I hit the wall, I hit the wall.
I burned through my resets. I am cheap enough that when a reset landed mid-afternoon and I knew I couldn’t use it all before the 6pm reset, I sat on it. I still have one banked.
This is the part of the AI cost conversation nobody puts in a deck: the person telling you costs are out of control is often just annoyed that he can’t buy more.
Then Luna 6 landed
Anthropic names models after literature. Haiku, Sonnet, Opus, Fable. OpenAI has Luna at the lower end, then Soul, then Astra. Luna is supposed to be the cheap seat.
Luna 6 just dropped with computer controls, and it is forty times less expensive per token than Fable. Forty. That is what is currently driving my Codex sessions, which is an experiment more than a settled decision. But the early receipt is good.
We have all been wringing our hands about cost for months. Meanwhile the answer to the subsidy question keeps arriving on its own: the models get more efficient, and the new cheap ones are as capable as the old expensive ones. If Luna performs anywhere near Astra at a fraction of the price, then when the subsidy goes away we should land at comparable performance for something close to comparable cost.
I am not as worried about the cost apocalypse as a lot of people are. I could be wrong. But I would rather be wrong about this than sit out another year of building.
What forty times cheaper actually bought me
Here is the concrete version, and it happened while we were recording.
We built a new offering: 2027 AI budget advisory, because there is a giant gap in how organizations are planning for this spend. Launching anything means a landing page, funnels, ads, all of it. Claude Code did a great job on the landing page, the marketing language, a draft blog, and it published. Then I hit a pile of things Claude was simply not going to do, mostly anything requiring computer control. Configure the campaigns in Google AdWords. Touch Search Console.
So I stacked those up and handed them to Codex. It logged into Google Tag Manager, Google Analytics, Google Search Console, Google AdWords, Slack, and HubSpot, with my permissions and restrictions, and did every manual configuration by hand that I would have done by hand.
Matt keeps calling the current state of these agents a good intern. I pushed back on that one. I have had skilled marketing people on my teams. That work is a week. Easily a week. It did it in four hours while I was at dinner. That is not intern work.
The part I can’t answer
I am still an Anthropic fan. I believe in the technology and the teams, and Claude Design is the best design tool I know of. But Anthropic suffered from the inference demand that hit when they started winning at the beginning of 2026. Tokens spent serving customers are tokens not spent training the next model. OpenAI has been investing more capacity in new model development.
So the question, and it is a question, is whether Anthropic can catch up when they haven’t been doing the training spend. If we get to Astra 7 before something comparable arrives from Anthropic, that is a real problem for them.
I don’t have an answer. Everything shifted from OpenAI to Claude once already. People are fickle and there is no real barrier to switching, so it can shift back.
What I can tell you is what I would do with my own money. If I could have my $200 plans, I would be running three Codex plans right now and using Anthropic as a double check.
If you are trying to figure out how to budget for any of this in 2027, that is the thing we just built. Give me a call.
