OpenAI 年度发布会最全总结：25 个新品很夯，但额度减半拉完了OpenAI Annual Event Roundup: 25 Impressive New Products, But Quotas Cut in Half

In the meritocratic world of AI, user reputation hinges solely on model capabilities and cost-performance.

Thus, with the recent launch of Opus 5.5 and Sonnet 5.5—models whose capabilities have shattered everything—Anthropic is in full bloom, with all things competing for growth. Meanwhile, under pressure, OpenAI has recently reverted to its cryptic mode, with various vague tweets essentially implying: "Watch DevDay to see how I strike back!"

In the early hours of today, OpenAI's annual developer event, DevDay 2026, finally arrived. Unlike previous years, which merely featured lukewarm updates to APIs and developer tools, this year's DevDay truly delivered high volume and high value, bringing over 20 new capabilities and model updates.

New models, new human-AI collaboration methods, new model speeds and subscription quotas... In summary, OpenAI, with its 1.2 billion weekly active users, is attempting to build its ecosystem into a complete "online office," where you can stay from morning until night.

AI Employees On Board, Moving Into the AI Office

Standing in the center of the stage at this "menu recital" launch event was the new model: dots.

In short, it is a 24/7 online agentic assistant, powered by GPT-6 Astra. Like a real colleague, it comes with a cloud-based computer, work environment, and browser (though you can also have it operate your own work environment), continuing to work even when you are not watching.

For example, suppose you are preparing for a model launch event. While you are off work, a competitor suddenly releases a strong rival product, directly impacting your model's selling points. At that moment, you can give your dot a voice call, having it work overtime to reorganize the model's positioning and modify the launch materials.

When you come to work the next day, you will receive its organized new launch plan and materials.

The author knows that good models have similar traits. Dots have the adorable big eyes of Grok Bot and are compatible with customizations similar to Muse, allowing you to modify the dot's appearance and name. Whether a triangle rice ball, a painter's hat cloud, or a big-eyed frog is up to you—let dots look more like your exclusive little sidekick.

Relying on its long-accumulated plugin ecosystem, dots can also operate over 4,000 applications, including Slack, Teams, Notion, email, and SMS, thereby becoming more "all-rounder."

Even when you are not looking for it, dots will find work for itself using "proactive research," such as traversing the entire workflow and proactively reminding you of tasks you might have overlooked.

Of course, OpenAI clearly understands the danger of putting an AI in a company that can control the platform, knows passwords, and likes to find work on its own when you are absent. Therefore, they added Custom Rules, permission controls, and automated operation audits to ensure dots do not disseminate sensitive information or change passwords by mistake.

To meet enterprise needs, OpenAI has also launched Specialist dots. These are company-standardized dots, each assigned specific identities, permissions, and credentials. Like highly specialized virtual colleagues, they take on specific responsibilities within the organization, such as legal and accounting work.

Dots are initially available to ChatGPT Pro, Business Premium, and Enterprise, with Pro and Business Premium plans including one free dot. Moreover, chatting with a dot does not consume your regular ChatGPT quota—

However, if you want to add a new dot or have it call Codex or ChatGPT Work to do the job, you will still have to pay up.

With these "employees" like dots, there is no reason not to have them work in the same environment as you, and thus ChatGPT Space was born.

ChatGPT Space includes several collaborative capabilities. For example, "Dynamic Pages" feel similar to Feishu Docs, allowing you to edit pages together with colleagues, ChatGPT, and dots. Just @ a dot, and it will start modifying the document or updating real-time information.

Similarly, in "Collaborative Slides," the entire team can create slides with dots. Everyone can comment online on areas they want to change, and dots will handle each request. Additionally, as long as the person who created the project has a ChatGPT membership, others without a membership can still comment.

Business and Enterprise users can also create teams, share tasks, and assign jobs like "organize project progress every Monday" to ChatGPT. Tasks can be triggered on a schedule or automatically started by conditions such as "receiving an email" or "a message arrives on Slack."

In another dimension, Codex, a favorite among developers, has been further moved to the cloud.

In his speech, Sam Altman first teased the phenomenon of not daring to close the laptop lid when using Codex, then announced: this phenomenon will become a thing of the past.

Now, you can control the cloud-based Codex Cloud from any device. This cloud development environment is retainable and reusable, eliminating the need to restart projects every time. It is even possible for an entire team to share a single specification and one environment.

The process of integrating various platforms continues. Sam Altman announced that in the future, users will be able to log in to third-party applications like Notion directly with their ChatGPT accounts, using the quota from their existing ChatGPT plans to utilize these models, thereby reducing the perceived cost of using different models.

GPT-6.1 Sol: One-fifth the price of Astra, is it enough?

After detailing the above capabilities, OpenAI unexpectedly released a new model: GPT-6.1 Sol.

Gemini's release of three models in the previous six weeks was already fast enough, but OpenAI has broken this record again—only 7 days have passed since the release of GPT-6 Sol.

The update speed of frontier models has now entered an era measured in days.

The wave of models has also brought some suffering to OpenAI: because the launch event was held while new models like GPT-6.1 Sol and bots were being rolled out, massive server pressure caused the live stream to stutter like a PPT, and on-site demonstrations frequently failed due to service crashes.

Getting back on track, the uncached API price for GPT-6.1 Sol remains $2 per million input tokens and $10 per million output tokens, consistent with the previous generation GPT-6 Sol (i.e., one-fifth the price of Astra), but the cached input price has halved compared to the previous generation, reaching $0.1 per million tokens.

Considering that Agent work frequently reads from the cache, the 50% drop in cached input prices may have a significant cost-reduction effect in practical multi-step Agent workflows.

In terms of capability, the official position is: achieving intelligence levels close to Astra at one-fifth the price.

On DeepSWE v1.1, which evaluates complex software engineering tasks, GPT-6.1 Sol surpassed Astra by one percentage point, while the cost to complete tasks was only $0.65, far lower than Astra's $4.43.

On OSWorld 2.0 computer operation tasks, 6.1 Sol scored 7% higher than 6 Sol, only 2.1% behind Astra, with the cost per task being roughly one-seventh of Astra's.

When it comes to complex professional PDF tasks like GDP.pdf, its score was higher than Claude Opus 5.5 with fallback mechanisms, and the cost per task was less than half of the latter. As for the multi-step business process test AutomationBench, at medium inference intensity, 6.1 Sol scored 2.2% higher than Opus 5.5, with costs approximately one-third.

In other words, although the release times were only a week apart, the performance gap between 6.1 Sol and 6 Sol might be comparable to the gap between Claude Sonnet 5.5 and Sonnet 5, also raising doubts about whether the previous 6 Sol was essentially Terra.

As the officials have repeatedly emphasized, the greatest competitive edge of GPT-6.1 Sol lies in its price: in the same test, the average cost per task for GPT-6.1 Sol was about $5.47, Opus 5.5 was $23.21, and Astra was $23.80.

Ultimately, OpenAI simply wants to tell people one thing: Opus is good, but don't get too attached; the cheap and abundant Sol is better!

Currently, GPT-6.1 Sol has been rolled out to Plus, Pro, Business, Enterprise, and Edu users, and the API is simultaneously open, but it is limited to ChatGPT Work and Codex. It cannot be used in regular chat for now.

The speed of flying, at the cost of shrinking quotas

With "higher" and "stronger" models, how can there be no "faster"?

Thus, the Decisions API was officially announced. It is basically a replica of the previously popular Jev: developers preset several options, input materials, and the Decisions API can use a customized Luna model to quickly classify, route requests, or decide the next step of the Agent.

In the case published by the officials, the Decisions API can output results within 150 milliseconds, whereas standard GPT-6 Luna called via API takes 1.6 seconds; the former is about 10 times faster. In large amounts of repetitive work, this time difference will accumulate into a huge efficiency gap.

In another dimension, GPT-6 Astra has also received the UltraFast mode, achieving 6x output speed in the API, with Codex capable of reaching 8x speed, hitting 300 tokens per second. GPT-6.1 Sol will also support this speed tier later.

Speed comes with a price. Astra's existing Fast mode roughly exchanges a 2x price for 2.5x speed. According to reports by The Decoder, the API price for UltraFast mode is about 6 times the standard price, which corresponds to $60 per million input tokens and $300 per million output tokens for the Astra model.

As for regular ChatGPT subscribers who want to use the UltraFast mode, they can only subscribe to the newly launched Pro 500 membership tier.

As the name suggests, Pro 500 costs a whopping $500 per month. That is, Pro is now divided into three tiers: $100, $200, and $500. Each tier includes one free dot, but only the highest $500 tier supports the UltraFast mode.

OpenAI also took this opportunity to reopen the Pro 200 membership subscription, which had been suspended for nearly three weeks. However, it came with bad news: the new Pro membership quotas have been halved compared to the original.

Specifically, the new Pro 100 quota is only 5 times that of the Plus membership, while the Pro 200 quota has also dropped from 20 times to 10 times. The double quota for existing users will also expire on October 29. The only compensation is the removal of the 5-hour usage limit.

Before the new policy was implemented, Tibo explained on X that model prices will continue to drop in the future, and even with reduced quotas, the amount of work completed will still be more than before. However, this statement feels like a telecom carrier saying "5G is faster, so I will cut your data plan in half," and it is clearly difficult to convince people.

A more likely reason is that OpenAI is tilting towards enterprise services, and the previous practice of subsidizing individual user subscriptions with large quantities and preferential prices is naturally unsustainable.

If you watch the entire DevDay launch event from beginning to end, you will find that beneath the seemingly chaotic stack of new products, there is actually a main thread—

ChatGPT is becoming more and more like a SaaS service.

If the way of using AI in the past was constantly starting "new conversations," what OpenAI is clearly trying to do now is replace it with a "persistent working relationship."

This change targets the enterprise market.

Last month, well-known tech analyst Ben Thompson, in a podcast discussing OpenAI's commercialization process, specifically cited the example of Dropbox taking two years to rebuild its client for enterprise users: consumer models can bring huge brand awareness and user scale, but truly stable, high-value business models often still have to return to enterprise customers, organizational deployment, and workflows.

In this dimension, Anthropic has already been a pioneer.

According to a Reuters report, last year, over 300,000 enterprise and commercial clients of Anthropic contributed approximately 80% of revenue, and Ramp's enterprise spending data also showed that Anthropic at one point captured 73% of global new enterprise AI customer spending.

It was this "overtaking on the curve" in the enterprise market that allowed Anthropic to achieve quarterly profitability first, with its latest valuation now approaching $2 trillion.

By comparison, ChatGPT already has 1.2 billion weekly active users, with consumer mindshare that is almost unmatched. However, how to convert this awareness into revenue and prove to the market that its business model has self-sustaining capability is a challenge OpenAI must address before its IPO.

Additionally, considering the current situation where each model vendor is trading blows, each leading for a few weeks, the author can safely assume: no single model can lead competitors in capability by more than half a year.

This also means that building a moat solely by relying on model capability advantages is no longer a viable business model.

In the future, AI vendors must bet big on hard-to-migrate things such as projects, Agents, permission management systems, and software workflows to maintain user stickiness.

From this perspective, it is not hard to understand OpenAI's decision over the past year to cut C-end "toys" like the Sora and Atlas browsers—

At DevDay 2026, what the author witnessed was actually OpenAI's "coming of age" ceremony in its transformation into an enterprise service company.