An AI twin that works while you sleep, and a Pro plan cut in half. Inside OpenAI DevDay 2026, plus a build-off against Claude Opus 5.5

On September 29, 2026, OpenAI made more than 20 announcements in San Francisco (OpenAI). Its new flagship model costs half what the flagship cost a month ago. Yet on the same day, OpenAI said new subscribers to the $200 Pro plan would get half as much usage.

The AI got 50% cheaper, and $200 now buys half as much. Wait, how does that add up?

We'll get there. First, short answers to the four questions you're probably asking.

01

DevDay 2026's five big launches in three minutes

There were more than 20 announcements. These are the five that matter.

• dots: an AI "second self" that runs on the top model, GPT-6 Astra, and keeps working 24 hours a day. Each dot gets its own cloud computer and connects to more than 4,000 apps. You can talk to it from ChatGPT, Slack or Teams, and it learns your preferences the more you use it. Available first on Pro and Business Premium, with the first dot at no extra charge (OpenAI)
• GPT-6.1 Sol: the new workhorse model, with performance close to Astra at a fifth of Astra's API price (OpenAI). It is also half the price of GPT-5.6 Sol, the flagship just a month earlier (OpenAI API)
• Ultrafast: an express lane that returns answers up to 8x faster. On the API it costs 6x the normal rate (OpenAI)
• Pro 500: a new $500-a-month top tier, and the only personal plan that includes Ultrafast (OpenAI Help Center)
• Pro 200: new sign-ups, paused since September 10, reopen. But usage drops from 20x Plus to 10x Plus (Thibault Sottiaux on X)

A packed venue, and the star demo froze on stage

The venue was Fort Mason, an event space on the San Francisco waterfront. The line stretched out before the doors opened, and protesters were reportedly gathered outside the entrance.

The roughly 54-minute keynote opened with dots, then moved on to GPT-6.1 Sol and Ultrafast. During the dots demo, a dot stopped answering mid-conversation on stage, and the room went silent. The show ended with someone on stage pressing a button that reset usage limits for every user worldwide, to applause (Simon Willison's live blog). The full keynote is on OpenAI's official channel.

02

Social media cheered dots and raged at Pro 200

A day after the announcements, X (formerly Twitter) was dominated by two topics: amazement at dots, the AI second self, and anger that the $200 ChatGPT Pro plan would give you half as much.

dots drew amazement, and a good laugh

Reactions to dots fell into three threads: it acts before you ask, you can call it, and a rival pulled off a prank.

Developers with early access kept reporting moments where dots moved first.

• It read their email and flagged a double booking and a suspicious login (Firtman's post)
• While its owner slept, it dreamed up a tool nobody asked for and built it (Linton's post)
• It found a way to cut a recurring bill by about $500 a year, then handled the back-and-forth with customer support (a post by OpenAI researcher Noam Brown, so not a neutral party; Brown's post)

dots let you talk by voice, like a phone call. Tap the call button in a conversation with your dot, and you're connected. You can check in by voice while it works: "How far along are you?" "What should we do about this?" And when you hang up, the work keeps going (ChatGPT Learn). In Japan, a video of someone actually calling their dot made the rounds (paji's post, in Japanese).

The reverse, a dot calling you, doesn't exist yet. OpenAI's help page lists it as "planned for after launch," and for now progress reports and questions arrive by chat. In that same video, the dot itself says it can't place calls right now.

The most-liked dots post wasn't from OpenAI. It was a joke that "dot.com" belongs to Elon Musk's SpaceXAI and redirects to the download page for rival chatbot Grok. It drew 36,000 likes, beating OpenAI's own announcement (35,000) (TechCrunch).

Dan Shipper of the US AI publication Every, who tested dots early, noted some glitches and suggested waiting a week or two before trying it (Shipper's post).

Pro 200 triggered "back to Claude." The math told a different story

Now the second topic. OpenAI cut the usage new subscribers get on the $200 ChatGPT Pro plan in half (Part 04 has the details).

The news came the day before DevDay (US time) in a post by Thibault Sottiaux, who leads Codex at OpenAI: the new Pro 200 is worth half the old one, measured in API dollars (Sottiaux's post). It drew 22,000 likes and 7,180 replies. Compare the dots announcement, with 35,000 likes and 1,906 replies. This one sparked argument, not applause.

The replies were full of people announcing they were switching to Claude.

People who sat down and did the math, though, pointed out that not everything gets cut in half (one such post). The logic is simple.

• The allowance was halved, measured in API dollars
• But the new GPT-6.1 Sol costs half what last month's flagship did
• Half the dollars on a model that costs half as much buys the same amount of work

The allowance that really shrinks is the one for the top model, Astra. The angry users are the ones who had been running Astra flat out. When OpenAI says you can "get more done than before," it is assuming you use Sol. The two sides are looking at different allowances, so they talk past each other.

03

How big is the gap? Astra, Sol, Fable and Opus, side by side

Right now, OpenAI and Anthropic (the maker of Claude) each have two models at the cutting edge, the so-called frontier models. These are the four we compared.

Sources: OpenAI API, “Pricing”; Anthropic, “Plans and pricing” (in Japanese) (both accessed September 30, 2026). A token is the unit AI uses to read and write text.

Every AI company's own numbers tend to favor conditions where it wins. So we used the overall index from Artificial Analysis, an independent evaluator. It averages 10 evaluations, covering business documents, coding, hard reasoning problems and more, with every model set to maximum reasoning effort.

Opus 5.5 leads with 58, Fable 5.1 and Astra tie at 53, and GPT-6.1 Sol scores 52 (Artificial Analysis). The real gap is cost: $0.72 per task for Sol versus $7.63 for Fable 5.1. A few points apart in intelligence, up to 10x apart in cost. Opus 5.5 has a lower API price than Astra, but it thinks through about 120,000 tokens per task, more than 4x Astra, so it ends up costlier per task.

Even Anthropic, announcing Opus 5.5, wrote that at this level of capability, benchmark differences are becoming an unreliable guide to real-world differences (Anthropic). So what changes when you actually make them build something? We gave all four AIs the same one-line prompt.

Both OpenAI models drew detailed illustrations with a sun, clouds and even hills. Both Claude models drew clean scenes: a white pelican on a red bicycle, blue sky, a strip of ground.

We asked for Japanese neon, wet pavement, rain, an elevated train and moving cars, with a camera that moves on its own (high reasoning effort).

Opus 5.5 built the most detailed city, down to a giant screen running an ad for a new "Night Sky Soda, ¥160" and a train marked as a rapid service to Takao, a real destination on Tokyo's Chuo Line. Fable 5.1 picked the most Tokyo-sounding signs, like a neighborhood bathhouse called "Matsu-no-yu" and "Open until the last train." For the fine detail of a Japanese street, both Claude models were a cut above.

Sol had fewer signs, but the most readable text. The surprise was Astra: it made 65 Japanese signs, but dark backing panels ended up in front of the lettering, so not a single one was visible. What a waste.

All four produced games you can play right away. The difference was what's inside. We ran each self-playing demo for three minutes and counted features in the code.

Enemy waves change every run, so results vary (in another run, Astra beat the boss).

Opus 5.5 even built the best autopilot, testing 85 possible moves every 0.05 seconds to dodge bullets. The two OpenAI games looked polished but were simple underneath, and both named their boss "DREADNOUGHT", a fitting coincidence for two models from the same company. Building them cost about $0.17 with Sol and about $2.04 with Opus 5.5.

04

So, pricier or cheaper?

It depends on which AI you use, and how.

Sources: OpenAI API, “Pricing”; OpenAI Help Center; Thibault Sottiaux on X (all accessed September 30, 2026)

The "half" is really a change in how the allowance is counted. OpenAI redefined the new Pro 200 allowance as half the old amount, valued at API prices. Because it's counted in dollars, the cheaper the model, the more you get.

The losers are people who have kept Astra, the top model released September 3, running nonstop. Demand for Astra was so heavy that OpenAI paused new Pro sign-ups on September 10 (TechCrunch). When sign-ups reopened, the allowance changed with them.

Existing subscribers keep the old allowance through October 29, and got a one-time grant of 62,500 credits (worth about $2,500, valid through the end of the year) (TheNextWeb). The same amount landed in our team's account.

What felt most expensive to many people was paying for speed. Compared with standard speed, Astra on Ultrafast drains your allowance 8x faster (OpenAI Help Center).

In Japan, Yota Ishikawa tried Ultrafast on Pro 500 and reported that a build finished in 4 minutes 29 seconds but used about 7% of his weekly allowance in one go (Ishikawa's post, in Japanese). At that rate, the allowance runs out after about 14 runs a week.

EDITOR'S TAKE ── Editor-in-Chief Ogasawara's view

From selling AI's answers to selling AI's working hours

dots that act before you ask, a new model at half the price, and a Pro 200 cut in half. They landed on the same day and look like separate stories, but I think one thread runs through all three.

dots run on Astra, OpenAI's most capable model, and keep working while their owners sleep. Every person running a dot around the clock keeps Astra's compute busy around the clock.

But Astra was already so stretched that just one week after launch, OpenAI had to stop taking new Pro subscribers. Add always-on dots, and the scramble for Astra only gets fiercer. So on the same day, OpenAI halved how much Astra you get on Pro 200 and created the $500 Pro 500 for people who want more. That's how we read it.

Once AI goes from a tool that answers to a twin that keeps working, the scarce resource is no longer intelligence. It's AI working time. The excitement over dots and the halving of Pro 200 are two sides of the same story.

Sottiaux wrote that he doesn't want to inflate API list prices just to offer discounts, and that over time, what a dollar buys on a subscription and on the API will converge. On DevDay, OpenAI CFO Sarah Friar told CNBC that customers are starting to move from flat per-seat pricing toward paying for what they use (CNBC).

Subscriptions are drifting from all-you-can-eat toward something closer to a prepaid card, where everything is measured in dollars. The same allowance lasts 5x longer on Sol and drains 8x faster on Ultrafast. Which AI you use, and at what speed, now matters more than which plan you're on: the value of the same $200 can swing several times over.