I Was at OpenAI DevDay 2026. Here's What OpenAI Is Turning ChatGPT Into
What did OpenAI actually show at DevDay 2026 — and what of it is worth the money, and what is still a raw experiment?
The new $500 ChatGPT Pro gives 25x Plus, while the $200 plan now gives 10x instead of 20x: half the limits for the same money, and for Ilnar Shafigullin that is DevDay's main disappointment. At a closed session OpenAI's architects calculated that an hour of Astra's autonomous work used to cost $32, and with 6.0 it costs $16: a month of one such session comes to $23,000, ten parallel ones to $230,000. Ultra Fast speeds work up eight times at six times the price and works only in Codex and Work with 6 Astra. Dots is one window to all of OpenAI's infrastructure: it reads chats and sorts email, but it is not connected to WhatsApp or iMessage — and in essence it is OpenClaw moved inside ChatGPT. The Sites presentation, with a Google Sheets table and a choice of widget colour, reminded Alexander Volchek of a 2010 CMS, and Siri AI lost to Codex on a simple contacts task: five contacts found against 324.
What to watch for
Key takeaways
Two hours before the event Alexander was offered a free trial of the $100 ChatGPT Pro, and the next morning he was credited $2,500. Such a seemingly small slip, he says, shows how the corporation works.
OpenAI released a $500 plan, and Alexander upgraded right at the conference. The new plan gives 25x Plus, while the $200 Pro gave 20x.
Ilnar counts limits from Plus: the $200 plan gave 20x, now it gives 10x, and the $500 one 25x. For him this is DevDay's main disappointment: limits were one of OpenAI's main advantages over Anthropic.
At a closed session an OpenAI developer said he has sessions that have been running for more than three months, and Astra 6 can work without a break for about 300 hours. Long autonomous work began roughly with Fable.
An hour of Astra's autonomous work cost $32, and with 6.0 it costs $16. A month of one session is $23,000, ten are $230,000, so, Alexander believes, strong systems will be expensive, and people will be left an interface for life.
By Ilnar's estimate, Ultra Fast costs six times as much as a normal request and will eat limits just as fast, so the $500 plan's 25x will vanish instantly. He doesn't even use the regular Ultra.
Alexander bought the plan for cases when he needs an answer right during a meeting: six minutes of analysis instead of an hour. The downside is that the mode works only in Codex and Work and only with 6 Astra, and ChatGPT doesn't have it.
A week passed between 6.0 Sol and 6.1 Sol. Ilnar thinks the model was rolled out to have something to talk about after the cut limits: there won't be enough limits for Astra, and 6.1 Sol is similar to it on benchmarks and much cheaper.
Dots reads chats in Codex, Work and ChatGPT, works with Slack and Microsoft Teams but not with WhatsApp or iMessage, is available on Pro and is created from Codex. For Alexander the main plus is one window, though the dot's access is still limited.
Alexander had his dot sort his email and his children's school emails and then notify his wife too. If dots could talk to each other, he would communicate with his wife only inside ChatGPT, not on WhatsApp or iMessage.
Ilnar sees Dots as OpenClaw appearing inside ChatGPT: it is easier to set up than OpenClaw or Hermes, but you are locked into the ecosystem. Most likely it is a project of OpenClaw's founder, whom OpenAI poached.
OpenAI's leadership says the company is a platform and won't build all software solutions. Alexander is sure of the opposite: sites in ChatGPT kill CMSs and hosting, and Dots will one day kill messengers.
8 million sites, a Google Sheets table inside OpenAI, a site from one prompt to view a report, a choice of widget colour and login only through ChatGPT — for Alexander that is the level of a 2010 CMS.
Siri AI found five, then eight old contacts and suggested searching by hand. Codex found 324 in a couple of minutes, saved a copy and deleted them. 'Siri, goodbye,' says Alexander.
Skills that used to live for generations, then for one generation, now go out of date in days. The main skill that has still not lost its value is learning to learn, says Ilnar.
What this episode is about
A ToTheMoon episode right after OpenAI DevDay in San Francisco: Alexander Volchek was at the company's main conference, including its closed sessions, and together with Ilnar Shafigullin, who is recording from China, he works out what of the announcements really changes work with artificial intelligence, what you will have to pay more for and which announcements are not worth your time. It opens with gifts and oddities: a Chromatic for Codex that is already being sold on eBay, an offer to try the $100 ChatGPT Pro, and $2,500 in credits the next morning. The organization shows, Alexander believes, that OpenAI is not an AI-first company.
The main news is the $500 ChatGPT Pro. Alexander bought it right at the conference, while Ilnar calls it DevDay's main disappointment: the $200 plan now gives 10x Plus instead of 20x and the $500 one 25x — half the limits for the same money. Limits were one of OpenAI's main advantages over Anthropic: in Claude they add up across the chat, Claude Design and Claude Code.
At a closed session OpenAI's architects said their sessions run for more than three months, and Astra 6 can work without a break for about 300 hours. An hour of such work cost $32, and with 6.0 it costs $16; a month of one session is $23,000, ten sessions $230,000. Alexander's conclusion: building on strong systems will keep getting more expensive.
Ultra Fast speeds work up eight times but costs six times more and, in Ilnar's estimate, will eat through limits instantly. Alexander sees it as a tool for urgent analysis — six minutes instead of an hour — but the mode is available only in Codex and Work and only with 6 Astra. GPT-6.1 Sol came out a week after 6.0 Sol and is ten times cheaper than 6.0 Astra; Ilnar thinks the model was added to make the presentation look 'beefier' against the cut limits. Codex Cloud became available from the phone, but not in full; Anthropic released Opus 5.5, and Google announced Gemini 4.
Most of the time goes to Dots — a central interface to all of OpenAI's infrastructure. It reads chats in Codex, Work and ChatGPT, is integrated with Slack and Microsoft Teams but not with WhatsApp or iMessage, is available on Pro and is created from Codex. Alexander gave it his email, his children's school emails and messages to his wife, and believes such dots could replace messengers. Ilnar sees Dots as OpenClaw moved inside ChatGPT: the entry threshold is lower, but you are locked into the ecosystem, the interface is turning into a dump, and it may die like Sora, plugins and GPTs.
The marketplace and using the Pro subscription in partner apps like Vercel are still unclear. Alexander is sure OpenAI is not a platform, as its leadership claims, but an operating system: sites in ChatGPT kill CMSs and hosting, and Dots will one day kill messengers. The hosts call the work example — Dots picking up a task in Slack — pathetic.
The ChatGPT Sites presentation disappointed Alexander most of all: 8 million sites, a Google Sheets table inside OpenAI, a site from one prompt to view a report, a choice of widget colour and login only through ChatGPT. Most demos looked like pet projects, every speaker, Sam Altman included, read from the screen, and the only interactive moment came from Tibo with a robot. Ilnar thinks DevDay is good for competitors, OpenAI has no central line, and Siri AI lost to Codex on a contacts task.
At the end the hosts go through real-time voice and vision — great but expensive technologies — Spaces as a Notion analog and the Decisions API, which works only with Luna. Ilnar tells a story from the physics department: skills that used to live for generations now go out of date in days, and the main skill of the future is learning to learn.
DevDay 2026 showed two OpenAIs at once. One builds strong models and Codex, which the hosts work in every day, and is turning ChatGPT into one window for work and life — and eventually, Alexander Volchek believes, into an operating system. The other cuts limits, names models so that no one can make sense of them, and shows pet projects on stage while reading from the screen. For the user the conclusion is the same: strong AI is getting more expensive, you will have to choose your plan and tools yourself without trusting the hype — and learn faster than the systems change.
Episode transcript
The episode is in Russian; below is an English reading guide to the transcript (the full EN transcript is a machine translation). Voice matching applied to 180 segments: 180 identified, 0 mixed, 0 probable, and 0 unresolved.
Read transcript on a separate page
Loading…