ChatGPT 一夜抄完半个 AI 圈，现场演示频频翻车ChatGPT Clones Half of the AI Scene Overnight, Frequent Failures in Live Demos

Before OpenAI's developer conference had even opened, users had already received a "reduction notice": the $200/month Pro plan would be reopened to new users, but the available quota was cut in half.

The price stayed the same, but the benefits decreased. What can OpenAI offer to convince people to keep paying? This is a challenge that this year's developer conference urgently needs to answer.

Just moments ago, OpenAI CEO Sam Altman announced a slew of updates in one go, including Dot, ChatGPT Space, GPT‑6.1 Sol, UltraFast, Pro 500, Decisions API, Agents API, cloud-based Codex, and the OpenAI Marketplace.

While the keynote appeared packed with new products, the main thread was quite focused. OpenAI is connecting models, agents, collaboration spaces, and commercial distribution, gradually transforming ChatGPT from a tool for generating answers into a work platform that can understand context, invoke software, and handle long-term tasks.

However, this launch is less about defining new products and more of a comprehensive benchmark against competitors: OpenAI intends to rebuild, in a platform-based manner, the forms that others have already validated.

Dot Debut: Rethinking the Division of Labor Between Humans and AI

Without a doubt, "Dot" stole the spotlight as the biggest star of the keynote.

Altman shone with excitement on stage, stating frankly that this was a model he had been waiting for for years. Similar to the currently popular Muse and Grok Bot, Dot is also a type of highly capable, 24/7 online Agent.

Users can name it, assign it long-term responsibilities, and authorize the relevant applications, devices, and work environments to it. Dot will continue to push tasks forward in the background, proactively contacting the user only when encountering critical changes, needing judgment, or requiring permission.

Specifically, users can use Dot to continuously monitor bug reports, test fix ideas and draft PRs, drive next year's budget planning, or migrate the entire codebase to a new API that is about to be deprecated. Dot will sort out dependencies, modify code, run tests, and hand the results over to the team for review.

The foundation for accomplishing these tasks is that Dot has its own identity, a cloud computer, and a browser, and can write and test code when needed. It can also connect to user-authorized laptops to directly enter real work environments.

OpenAI's plugin ecosystem provides it with the ability to connect to over 4,000 applications. Users can communicate with it via ChatGPT, SMS, Slack, Teams, and voice calls, with cross-channel context remaining synchronized.

Subsequently, Holly, head of OpenAI's model team, demonstrated Dot's actual working method using a fictional music app called Blossom.

As the team prepared to launch the model, Dot had already discovered in advance that the release review meeting had been rescheduled, updated the calendar, summarized feedback from test users the previous night, and found the homepage concept submitted by the design team at the last minute.

Upon receiving the instruction to "go ahead," it could invoke Codex on the computer to turn the design into a runnable application and prepare to submit the changes as a PR. However, during the on-site demo, Dot stumbled several times; not only was its response slow, but when Dotty called Codex to build the application, the Codex thread used to display task progress failed to load properly for a moment.

In addition, Dot will join Slack group chats and receive tasks just like a team member. In the demo, engineers only needed to send user feedback and session IDs to the channel, and Dot could investigate the issue and submit a fix.

Oh, by the way, netizens also found a hidden easter egg: logging into "Dot.com" now directly redirects to the promotional interface of Gork Bot.

Unsurprisingly, Meta's Chief AI Officer, Alex Wang, also sent his "best wishes."

Of course, the more proactive Dot is, the harder it is to avoid permission and security issues. OpenAI has configured an isolated cloud computer for each Dot. Users can decide which applications it can access and set permissions for different operations, such as allow, deny, or pre-approval.

Operations involving account changes or information sharing will go through automatic review, while sensitive tasks like changing passwords always require the user to complete them personally.

When Dot conducts "proactive research" in the background, it can only use restricted-permission tools to read connected applications; it cannot directly send messages, modify application content, or control user devices.

For enterprises, OpenAI has also released dedicated Dots.

Enterprises can configure agents with independent identities and clear responsibilities for roles such as finance, marketing, legal, procurement, and customer service, with employees continuously reviewing and providing feedback. OpenAI is collaborating with Microsoft to integrate dedicated Dots into Agent 365, allowing enterprises to manage these agents using their existing identity, permission, and security systems.

Dot will gradually be opened to Pro and Business Premium users in eligible markets, while Enterprise users can try the beta version after administrators enable it.

The first Dot is included in the corresponding plans. Conversations with Dot do not consume ChatGPT usage quotas, but actual tasks initiated via Codex or ChatGPT will still be billed as usual. In the future, users will be able to add more Dots or increase their speed and work capacity.

Personal Dots are responsible for understanding users and continuously pushing tasks forward, while ChatGPT Space is designed to allow enterprise team members and their respective agents to collaborate in the same place.

Space uses Pages as its basic unit, specifically designed for collaboration between humans and agents. It can hold text, files, research materials, images, tables, dynamic charts, and runnable prototypes.

Users can @colleagues just as in a collaborative document, and can also @a specific Dot in comments to have it track dependencies, find data, or update content. Pages can also receive standing instructions, such as checking specific Slack channels daily and updating new findings in the current document.

At the launch event, Holly had Dot convert user feedback into a filterable bar chart that auto-updates hourly; subsequently, she had it find registration data mentioned by a specific engineer from Slack DMs and add it to the model release FAQ.

OpenAI also announced that in the future, presentation capabilities for reading and modification by appropriate agents will be added within Space, allowing teams to create slides together with multiple Dots.

The combination of Dot and Space also reveals OpenAI's judgment on next-generation office software. Chatting is just the entry point for issuing intent; persistent context, an operable software environment, and collaboration spaces with human approval involvement are the real infrastructure for agents to enter daily work.

GPT‑6.1 Sol: Lowering the Cost of Long-Term Work

To keep agents working for long periods, the model must possess advantages in capability, speed, and cost. For this purpose, OpenAI released GPT‑6.1 Sol, positioning it as the daily main model for developers and agents.

According to detailed pricing, it charges $2 per million input tokens and $10 per million output tokens, both one-fifth of the corresponding GPT‑6 Astra prices. Cached input costs $0.10 per million tokens, which is suitable for tasks requiring repeated reading of project materials and reusing context.

The price drop did not come with significant capability regressions.

OpenAI stated that the new model, in practical tests for complex software engineering, professional document processing, multi-step workflows, and computer operations, approaches Astra's performance at a lower cost, while the capability ceiling for the most difficult scientific research tasks is still maintained by Astra.

In the DeepSWE 1.1 software engineering benchmark, GPT‑6.1 Sol achieved a level close to Astra at roughly one-fifth of the cost, and exceeded GPT‑6 Sol's best score by 6.4 percentage points.

In GDP.pdf, which measures complex PDF understanding capabilities, it outperformed Opus 5.5 with fallback mechanisms at all reasoning levels, with a single-task cost of less than half of the latter.

In AutomationBench, which targets multi-step business processes, GPT‑6.1 Sol led Opus 5.5 by more than two percentage points at medium reasoning intensity, with a cost of about one-third of the latter.

In the computer operation benchmark OSWorld 2.0, at the highest reasoning intensity, the gap with Astra narrowed to within 2.1 percentage points, and the single-task cost was about one-seventh of Astra's.

Starting today, Plus, Pro, Business, Enterprise, and Edu users can use GPT‑6.1 Sol in ChatGPT Work and Codex, but temporarily cannot select this model in regular chat. Developers can invoke it via the API using gpt-6.1-sol.

Around Dot and GPT‑6.1 Sol, OpenAI also released a set of infrastructure services for agents.

UltraFast mode boosts inference speed to eight times that of the standard mode, reaching up to 300 tokens per second. It is currently available for Astra and will soon be compatible with GPT‑6.1 Sol.

OpenAI also launched the Pro 500 dollar plan, offering approximately 25 times the usage quota of Plus and UltraFast privileges in ChatGPT and Codex, and reopened the Pro 200 dollar plan. At the end of the keynote, OpenAI added a one-time quota reset for global users.

For scenarios requiring instant decisions, OpenAI also released the Decisions API, benchmarked against Jev. It allows Luna to quickly select results from preset options, applicable to request routing, image classification, and determining the next action of agents, while retaining visual understanding, multilingual, and safety capabilities.

Codex has also been fully migrated to the cloud.

Developers can initiate tasks from their phones and continue viewing them in the browser or on desktop; even if the computer is closed, the agent will still work in the cloud. The new version of Codex CLI adds bidirectional voice capabilities, and Codex Security Cloud can continuously scan for vulnerabilities, automatically deduplicate, and prepare verified fix ideas.

The Agents API opens up capabilities used internally by OpenAI for Codex and Dot to developers, including the execution framework, hosting, memory, multi-agent control, and computer use.

OpenAI also previewed Private Intelligence, aiming to allow enterprises to invoke frontier model capabilities without retaining content by leveraging zero data retention, private secure processing, and private inference.

Beyond models, OpenAI wants to control the entry point for applications

After building the model, OpenAI also wants to master the distribution stage.

In the future, users will be able to log in to third-party applications with their ChatGPT accounts and consume tokens from their subscription plans, thereby allowing developers to reduce the cost for users to initially experience the applications.

Plugin Extensions allow developers to put editors, dashboards, and even complete workspaces into ChatGPT and Codex; Figma and Adobe have already demonstrated corresponding cases.

The OpenAI Marketplace launched on the same day, with its first batch including over 30 partners. Enterprise customers can use existing OpenAI contract credits to purchase third-party applications, while developers can access a distribution channel with approximately 1.2 billion weekly active users.

Plugins will also be proactively suggested during conversations, so ChatGPT simultaneously plays the role of entry point, runtime environment, and application store.

After the new model introduction, Altman returned to the topic of quotas that people are most familiar with. He said that OpenAI has done many "banked resets," to the point where Tibo jokingly suggested renaming the company the "Reset Company."

But what I care about more is that there actually seems to be a physical button for resetting quotas in reality.

It is worth noting that at the end of the keynote, Sam Altman referred to the ideal future of AI as a "new Renaissance," rather than a new Industrial Revolution.

He hopes that AI will give individuals more ability to create, learn, and explore. From a business perspective, OpenAI's goal is equally clear: it is pushing ChatGPT from a super app to a platform hosting users, agents, and third-party software.

But grand metaphors ultimately need to be tested by reality.

The underlying logic of internet services is a relatively stable set of servers and rules. Webpages crash, platforms go down, and software has bugs, but most of the time, users know what they have handed to whom and roughly how it will be processed.

The change brought by AI agents is that they may misinterpret an instruction, be biased by external content, or make a decision that deviates from the user's intent in a series of seemingly reasonable steps.

More troublingly, as Dot, Codex, and Private Intelligence take over more tasks, a sharp problem lies before everyone: are we really ready to hand over so much information, permission, and judgment to a black box without reservation?

The answer is clearly no, but in the face of the temptation of seemingly ultimate efficiency and convenience, we seem to have no time to hesitate.