OpenAI Announces New Pricing and AI Agent Features at Developer Conference

Author | Alan

At 3:00 AM Beijing time on September 30, OpenAI held its self-described "largest ever" developer conference. The official summary page listed a total of 25 announcements, ranging from new models and APIs to an application marketplace for enterprises to purchase software.

The quietest moment on stage came when Sam Altman mentioned the adjustment to ChatGPT Pro memberships. No one in the audience clapped.

The adjustment itself is straightforward: the Pro 200 price remains at $200 per month, but the usage allowance for new subscriptions is reduced; additionally, a new Pro 500 tier is introduced at $500 per month, offering 25 times the usage of Plus, plus access to the Astra Ultrafast speed tier—available exclusively in this personal plan.

That silence spoke louder than any of the 25 announcements—it revealed today’s core theme: what this conference truly changed was how ChatGPT charges: it no longer charges like a messaging app, but now charges like an “employee,” with clear pricing for how much work is done and how quickly.

Monthly fee and usage are both 25 times that of Plus | Source: OpenAI

As previously hinted, it’s no surprise that OpenAI devoted the majority of this announcement to dots, and the official Chinese webpage also retained the English name.

According to official claims, each 'dot' is powered by GPT-6 Astra, equipped with its own cloud computer and browser, capable of connecting to over 4,000 applications and running 24/7.

The usage is also similar to delegating tasks. You give it a goal and define which actions it can handle on its own, and it will proactively get to work. It can operate within ChatGPT, Slack, and Teams—OpenAI product lead Tibo (Thibault Sottiaux) tweeted live that you can even call it directly while it’s working.

OpenAI describes it verbatim as "an agent capable of continuously taking responsibility," and envisions a future where multiple agents form teams to collaborate on your behalf.

The phrase "take on responsibilities" is precisely chosen. Previous AI assistants responded only when you asked a question; Dot, however, keeps working even when you're not around. This reads more like a job posting than a software manual.

Dots offers Pro starting at $100 per month only; Image source: OpenAI

The threshold for DOT is not low. Among individual users, only the Pro plan is eligible, and Tibo specifically noted that it includes the Pro 100 plan at $100 per month. For businesses, it’s the Business Premium plan.

The official statement says the first DOT is included free with your plan and chatting with it doesn't count toward your usage. However, using it for tasks in "ChatGPT Work" or Codex will still deduct from your quota.

The question is, what’s installed on this cloud computer. Tibo is confident in the cloud—he tweeted on-site that once the new Codex cloud environment is set up, you’ll never want to go back to developing on a laptop.

During a live-streamed event organized by Geek Park in Silicon Valley, a participant working in mathematics noted that his code relies on a commercially licensed mathematical optimization solver. While the local Codex could call, run, and test this tool, migrating to the cloud does not guarantee that these tools and licenses will follow. Could AI simply rewrite it? His assessment is that it is not yet at the level of mature tools.

He typically uses AI for specific research, not for booking tickets or courts. During the event, some attendees also noted that each company’s personal assistant is beginning to look increasingly similar. The day before, Manus released its 2.0 version, which also emphasized personal AI assistants and cloud computers. For these professional users, what dot lacks isn’t thousands more app integrations, but their own customized local environment.

Apart from DOT, the membership tiers were the most heavily modified aspect of this update and also drew the most criticism. Tibo tweeted about two hours before the event, clearly stating the new multipliers: Plus is 1x, Pro 100 is 5x, and Pro 200 is 10x. The previous Pro 200 tier was 20x.

To add a bit of surprise, Tibo even prepared a physical 'reset' button on-site | Image source: OpenAI

When you convert the monthly fees into multiples, it becomes clear: Pro 100 is 5 times the fee of Plus, Pro 200 is 10 times, and Pro 500 is 25 times—exactly matching their respective usage multiples. The old Pro 200 offered 20 times the usage for 10 times the cost, a kind of “bulk discount” that is no longer available; now, each tier provides exactly 1 unit of Plus usage for $20.

The official help page phrased this more gently: Pro 200 has reopened for new subscriptions at the same monthly price of $200, but new subscriptions now come with lower usage limits; eligible existing subscribers will retain their original limits until October 29, after which their limits will also be reduced. Tibo calculated beforehand that the new Pro 200, when converted to API costs, is roughly half the cost of the previous version.

The most concentrated dissatisfaction at the event also centers here. Some attendees felt that the available quota under the original plan had been reduced, effectively amounting to a price increase, while the anticipated new model failed to materialize; others said they were most eager for a model with stronger mathematical and algorithmic capabilities, but the newly released version instead focused more on consumer-facing features.

The official reason given is that the models have become more efficient.

Yet at the same conference, OpenAI unveiled its only new model, GPT-6.1 Sol.

It is an upgraded version of Sol, and the official statement says it has intelligence close to Astra’s, with API pricing at just one-fifth of Astra’s. Tibo also mentioned that GPT-6 Sol and Luna have just been reduced to half their original price.

The models are becoming cheaper, but the usage allowance provided to members has been halved.

Putting these two things together isn’t contradictory—it’s just a change in the unit of measurement. Following OpenAI’s logic, as models become more efficient, each unit of capacity can accomplish more work, so less capacity is needed. What you’re paying for is no longer the number of conversations or tokens, but the amount of work accomplished.

This is the logic used to calculate employees' salaries. The boss doesn’t care how many times you type on the keyboard—only how much work you deliver.

Speed is now priced separately. According to official data, Astra Ultrafast is up to 8 times faster than standard speed in Codex, processing 300 tokens per second, and up to 6 times faster in the API, with API pricing exactly 6 times the standard rate. Within ChatGPT subscriptions, it consumes credits at 8 times the rate of standard mode and is available only to Pro 500 and Enterprise plans. To make the AI work faster, you’ll need to pay an express fee—just like sending a package overnight.

$50 to $300 per million output tokens | Image source: OpenAI

This trend has long been evident in Silicon Valley: In May, Microsoft paused vibe coding, concluding that burning tokens had become more expensive than paying employees; in September, Silicon Valley audited employee AI expenses, with the median monthly AI spending reported by Microsoft employees reaching $300.

Back then, companies were learning how to account for AI. Now it’s OpenAI’s turn, pricing AI based on computational costs—and the monthly fee for Pro 500 has already surpassed the median on that chart.

What about the $20 monthly Plus plan and free users?

The heaviest features this time—Dots, Data Space, Dynamic Pages, and Ultrafast—all require a Pro plan. GPT-6.1 Sol will gradually become available to Plus users, but currently it has only been integrated into "ChatGPT Work" and Codex; the official statement is that it has "not yet entered Chat." Features labeled as "available for all plans" are mostly developer-oriented, such as Codex command-line tools, code review, and plugin extensions.

The other half of the announcement focused on transforming ChatGPT into a gateway. OpenAI explicitly stated at the beginning of the summary page that it aims to open ChatGPT as a shared platform for human-agent collaboration, enabling developers to directly launch native experiences to a combined user base of 1.2 billion.

Plugins are the gateway to this entry point. Developers can build apps directly within ChatGPT: occupying a fixed slot in the sidebar, creating an interactive panel for use alongside conversation, or building a viewer for their own file formats. Accompanying these are plugin creation tools, a redesigned submission process, and a new sorting system; plugins can also automatically trigger a sequence of actions when something happens in a connected app.

When Tibo tweeted live, he added an even more crucial point: ChatGPT directly recommends relevant plugins during conversations.

The weight of this statement for developers far exceeds the importance of the plugin functionality itself. Users don’t need to search in a store—plugins are simply presented to them in conversation, but ChatGPT decides which ones to recommend and which to omit. Developers gain access to a billion-two-hundred-million users, yet simultaneously surrender control over ranking.

Another feature, called “Sign in with ChatGPT,” allows users to log in to partner products using their ChatGPT account without needing to register again. The official Chinese page explicitly states that partners can only access your name, email, and profile picture—chat history and memory are not shared. Plus and Pro users can also directly use their subscription credits on select partner products; this is currently available in limited preview. The official English page lists 16 partners, including Devin, Notion, and Vercel.

In the context of Chinese internet culture, this combination is far from unfamiliar. The plugins recommended in the conversation are like mini-programs; logging in with ChatGPT resembles logging in with WeChat; membership credits are interchangeable across partner products, functioning like a cross-app wallet.

This layer of the wallet is connected to the membership tiers discussed earlier. According to the official English page, OpenAI usage consumed through these partner products counts toward your own plan allowance. In other words, the usage calculated earlier at "$20 per unit" can now be used not only in ChatGPT but also in other partners' products.

There is also an “OpenAI Software Marketplace” for businesses: eligible enterprise customers can use their committed spending credits with OpenAI to purchase software from the first 32 partners, including Adobe, Figma, and Salesforce.

Things needed for work have also been moved into ChatGPT. The Docs space has replaced the original Library, allowing teams to continue working alongside ChatGPT and their own dots in the same space; Dynamic Pages are a new type of document, officially described as “anything you can do with ChatGPT can be done on a page”; Slides can be collaboratively edited and exported as PowerPoint or Google Slides.

These features are available only to Pro and Enterprise plans. Enterprise users can @ChatGPT in Slack and Teams, allowing colleagues without ChatGPT licenses to join the same conversation.

Regarding Codex, in addition to the cloud version mentioned earlier, the command-line interface has also been updated to support voice-based task assignment, and new versions of code review and code security scanning have been released.

The API has added the Decisions API powered by Luna, which Tibo says can make decisions in hundreds of milliseconds; the Agents API, designed for computer operations, has entered preview, with Tibo noting that dots uses the same technology. OpenAI, in collaboration with Amazon, has integrated its agents into AWS’s Bedrock.

Tibo also mentioned another number in his live tweet: token traffic on the Responses API has increased 100-fold compared to this time last year. He later acknowledged that the live demo was affected due to all updates being deployed simultaneously.

Tibo's conference summary page | Image source: X

Looking at the 25 releases, OpenAI wants ChatGPT to become a workplace: colleagues with their own computers, shared spaces for editing documents, integrated channels with Slack and Teams, plus a pricing table based on workload and speed.

At that moment, no one in the audience applauded Sam Altman: because the developers seated below were the first to pay according to this pricing schedule.