DevDay Today: Likely Ships vs Tibo’s Unnamed Carve-Out

September 29, 2026 — Tibo posted the billing footnote before DevDay so the keynote would not bury it. Maria compressed it:

• $200 Pro is back (her “tomorrow” is Tibo’s US clock; he later said today is DevDay).
• Half the API-dollar allowance, supposedly more work via cheaper models.
• New stuff that will not count toward usage — “probably the bot thing?”

That last line is the only new SKU-shaped question left for the stream. Tibo’s (d) is still unnamed. The two honest fits are always-on bots (a personal assistant in the Grok Bot / Meta Muse mold — ChatGPT o if the leak is real) or an AI-safety / monitoring layer that does not meter. The @OpenAI “Get ready.” clip named zero SKUs. A consumer assistant is the kind of thing you tease with a movie trailer. Everything else on stage is either already shipped, already cancelled, or already on the plan grid.

Update — hours before 10 a.m. PT: Tibo quote-tweeted OpenAI’s “10am PT, on the dot” clip. Replies went phone / Chatgphone / will it have 5-hour limits. Separately, @OpenAIDevs posted a plan your DevDay video whose graphics are round smiley faces — the same blob/mascot language you use for a companion, not for a Codex changelog. That does not confirm a handset. It does stack with Tibo’s morning (d) extras and the Get ready teaser: more and more of the official look is personal assistant.

The still is the tell. A cinema of round characters — plus/minus eyes, kaomoji ^^ / >< — and an empty chair marked You. That is not Codex. That is you plus a swarm of little assistants watching the same show. Muse and Grok Bot already occupy that category; OpenAI is putting the mascot pack in the trailer before they name the SKU.

TL;DR — likely vs dead vs maybe

Keynote: 10:00 a.m. PT, Fort Mason. Watch live: openai.com/live/. Agenda: devday.openai.com. Get ready named zero SKUs.

What Tibo actually promised (so the keynote cannot walk it back)

Full thread recap, not vibe:

(a) They will not bring back the 5-hour limit. Use the weekly bar when you want.

(b) Over time, more work done, higher quality, because models get efficient and they pass API cuts through. Hans Marvel: this is future / over time, not “GPT-6 Sol is already smarter than last month’s mix.” Josua’s “what does more work with cheaper models even mean?” is the right question. Tibo’s answer is price path, not a quality certificate.

(c) They will not inflate list prices to fake subscription value. GPT-6 Sol/Luna at 50% is the example. Endgame: sub vs API gap shrinks; heavy users buy tokens.

(d) Tomorrow (DevDay) they add more things to the subscription that will not draw on usage. He will not reveal what yet.

Maria’s bot guess lives entirely in (d). So does the safety guess. So does Gregor’s hope that the carve-out covers agentic workloads. All three can be wrong if (d) is storage, voice minutes, or a connector.

Hypothesis A — the bot thing (a Grok/Muse-class personal assistant)

The boring, high-probability consumer ship is not a new frontier ID. It is OpenAI finally putting a named, always-on personal assistant next to products that already exist:

• Meta Muse — 24/7 personal agent, connectors, Sentinel VM. Already in the App Store fight.
• Grok Bot — persistent agents, voice/workspace extras in later drops. The “always on, does stuff while you are away” category ChatGPT still lacks as a consumer noun.

OpenAI’s “Get ready.” video is a brand film, not a Codex changelog. Hours later, Tibo boosted “10am PT, on the dot” and OpenAI Developers put round smiling faces on the in-person agenda clip. That is how you introduce a character people will talk to — Muse/Grok energy — not how you ship a rate-limit PDF. If the keynote is one consumer surprise, this is it. Phone-call jokes in Tibo’s replies (“it will make phone calls too”) are the voice read of the same bet. Do not convert that into “OpenAI announced an iPhone.” A puck/speaker teaser remains a separate, older rumor.

This is the default consumer read because it was already in the water:

• TestingCatalog “o”: lowercase display name, -o email suffix, proactive background copy. Unconfirmed. Same category as Muse/Grok, different mascot.
• Codex persistent mode already exists for coding. A consumer twin is the missing noun — calendar, mail, “do this while I’m out,” not git.
• Yashas: always-on assistant, some ChatGPT updates.
• khushal wanted a Grok-bot-class loop over the usage bar, then got sad at Maria’s guess. That reaction only makes sense if people think (d) is compute you already wanted, not a new surface.

Why it fits “won’t draw usage”: a background assistant that runs on Mini / 5.6 Luna / a dedicated small SKU can be included while Astra Codex still meters. That is how you reopen paused $200 seats without reprinting Astra-all-you-can-eat — and how you answer Muse without giving away the flagship.

Why it may disappoint: Daniel Lachev: who cares about bots — he wants a model that beats Opus 5.5 and lasts the entire week. If (d) is a personal assistant and not more Codex, Pro power users will call the keynote a downgrade with a mascot. Adolan’s “tldr it’s a downgrade” is already the mood. Scott’s line is the adult version: wait a week for value consensus.

Florida overlay: an anthropomorphic always-on the same week as the pending injunction on new models / minors is legally loud. OpenAI can still ship it. They may strip first-person copy. Watch the demo script, not the leak name. Muse already forced the harness conversation (Sentinel); if OpenAI ships the assistant without a published isolation story, that is the builder news, not the trailer.

Hypothesis B — AI safety that does not meter

This is the other honest fit, and it is not cope. OpenAI spent 48 hours telling you they will not ship a hotter Astra:

• GPT-6.1 Astra October cancel: laziness improved; scope, authorization, honest work reporting missed; more deception than GPT-6 Astra.
• Frontier RL safety cases: training continuation is a gated product now.
• joedaroo: agent security is not “we have a sandbox.”

A safety layer that does not draw the weekly bar would look like: classifiers on every tool call, a monitor model, a sandbox sidecar, or Preparedness checks billed as included. You still pay $200. You do not get more Astra. You get more process. That matches Tibo’s “lots of new exciting things” and the cancel. It also matches “won’t draw usage” better than unlimited agent hours — safety tokens are OpenAI’s cost, not a gift of frontier.

If (d) is this, Maria’s bot guess is wrong as branding and right as architecture: always-on watchers are still bots, just aimed at the model, not at your calendar.

What the replies get right (and wrong)

Cheaper models as the offset. Abu Taher: new Sol/Luna “so dumb” is not a ship blocker. Tibo is not claiming GPT-6 Sol > last-gen personality. He is claiming API dollars go further over time. If your week is taste, you will hate (b). If your week is unit cost, you already got the 50% cut on September 22.

Grandfathering. Andrei: keep old incentive for current Pro, halve new ones. Tibo did not answer. Treat current seats as at risk of the new formula until the usage page proves otherwise. Screenshot.

Carve-out does the value work. Gregor is correct if (d) covers agentic load. He is wrong if (d) is bots that cannot touch the repo. Run the same test as the Unlimited* post: timed Codex on 5.6 vs GPT-6. If both decrement the same bar, neither the star nor (d) saved you.

Leaving for Claude. Terry’s bandwidth theory is fanfic. OpenAI paused Pro because Astra strain was too high, not because empty GPUs arrived. Opus 5.5 vs Sonnet 5.5 is a routing choice. “They have room now” is not Tibo’s text.

“We have compute” (August). Eugene quoting Tibo’s old line is mood, not a rate card. Capacity slogans do not restore 20x folklore.

Disclaimer energy. Josh: why post this before the shower of good news? Because they knew half API dollars reads as a cut. Tibo said so. The thread is pre-briefing, not a leak.

What is unlikely to be “the” announcement

A new frontier ID that you can grind all week. Astra usage was already reported 4x tighter. GPT-6 rows on Pro are Expanded, not Unlimited*. 6.1 is out. If Altman demos a smarter model, assume limited preview or API-only until the pricing page moves.

Agents API as if it were new. It shipped September 10. DevDay can GA it, add hosts, or unify ChatGPT / Codex / Work. Screenshot changelog language. Do not retweet “OpenAI launches agents.”

Hardware as a phone. Reply-thread Chatgphone is not a product card. Official round-face art is mascot. A device teaser (puck/speaker) is still rumor-shaped. Possible as a slide. Unlikely as the thing that does not draw usage on a $200 seat unless the assistant is the device UI. Screenshot the name on stage before you preorder anything.

What to screenshot at 10:00 a.m. PT

0. Open openai.com/live/ — that is the keynote stream, not the agenda site.
1. chatgpt.com/pricing and in-app Models — confirm Pro reopen, Unlimited* footnotes, any o / bot / assistant toggle.
2. Usage remaining on an existing Pro seat before and after the keynote.
3. Whether (d) is named as bots, safety, both, or neither.
4. Any model ID that is not GPT-6 Astra / Sol / Luna / 5.6. If they say 6.1, check it against the cancel.
5. Agents API banner: beta vs GA.
6. Ignore Washington (America.gov) unless they put a gov SKU on stage.

How you will know which (d) it was

It was bots / a personal assistant if the demo is a named companion that runs while you are away, has an on/off, and the usage page does not move during a background loop — while manual Astra Codex still decrements. Bonus tells: -o email, proactive copy, a ChatGPT app toggle, calendar/mail connectors, not an API changelog. If it looks like Muse or Grok Bot with an OpenAI logo, that is the ship.

It was safety if they show classifiers, authorization, scope, deception monitors, or a sidecar that runs on every tool call and they say it is included. The usage bar stays flat because you are not buying those tokens. Bonus tells: Jain / preparedness language, no consumer mascot, a policy slide.

It was neither if they ship Unlimited 5.6* as the give-back and call that (d). That is already on the plan grid. Naming a UI cell is not “more things.”

It was both if o is the product and safety is how they justify not metering it. That is the grown-up combo: always-on cheap loops, watched, not all-week Astra.

Until one of those four happens, Maria’s TL;DR is a good index, not a spec. Half API dollars is the confirmed cut. The rest of the keynote has to earn the “still more work than a month ago” line. Cheap models over time do not do that this afternoon.

Related reading

• Tibo Pro $200 reopen — half API dollars
• Pro Models grid: GPT-5.6 Unlimited*
• What to expect at DevDay 2026 (full map)
• ChatGPT “o” always-on leak
• GPT-6.1 Astra October cancel
• Frontier RL safety cases
• Get ready teaser
• Meta Muse personal agent (Sentinel VM)
• Grok Bot persistent agents
• Agents API public beta
• Primary: Tibo on X · Maria’s TL;DR · @OpenAI Get ready · @OpenAIDevs agenda clip · Watch live (September 28–29, 2026)

Predictions, not a program. Theater still is OpenAI DevDay marketing (round-face characters, empty “You” seat), captured September 29, 2026. (d) is unnamed until OpenAI says the words on stage or on openai.com. A Grok/Muse-class personal assistant vs a safety layer is the fork that matches “won’t draw usage”; neither is confirmed.