Why OpenAI’s DevDay Looked Like Catch-Up

While the rest of us were distracted by details within Anthropic’s IPO filing and OpenAI’s plans to raise $30 billion in a new round that could value it around $1.4 trillion, app developers on Tuesday flooded to San Francisco’s Fort Mason for OpenAI’s annual developer conference, where it made a flurry of new product and feature announcements.

All in all, OpenAI’s DevDay felt a bit reactionary. Take “Dots,” an always-on or persistent agent product—which we broke news about earlier this month—that promises to automate complex processes like turning customer feedback into improvements for an app or filling out invoices for employees.

In case you’re counting, this is OpenAI’s fifth attempt to launch an agent product, counting past tries like Operator, Deep Research, ChatGPT Agent and ChatGPT Work. The company said that this time was different, though, because of how good frontier AI like Astra has become (and they might be right!).

Dots looks similar to recent breakout agent products like Instinct, Meta Platforms’ Muse and SpaceX’s Grok Bot, which allow people to talk to agents through text and calls. To be fair, though, OpenAI’s version felt distinctly productivity- and work-focused compared to some of these other examples, especially with features like having Dots work together to create documents and presentations. (Separately, can anyone tell me how to test out all these different agent products? I’m personally overwhelmed.)

With all the news about rogue agents recently, though, users might be hoping that Dots are better at behaving themselves than other OpenAI agents!

Then there was the Decisions application programming interface, which drew lots of excitement from the developer crowd. It works by asking OpenAI’s fast and cheap Luna model to respond to a question using a finite set of pre-defined answers. For instance, a developer might ask the Decisions API to route customer support tickets between three options: billing, fraud or technical issues. That allows the model to respond much faster than it otherwise would, the company said.

The Decisions API announcement looks like a response to Jev, a new model that’s been gaining popularity this month for its ability to respond to questions super fast and at a fraction of the cost of frontier models like OpenAI’s. Corporate executives told us last week at AI Agenda Live that they’re using the model for applications like classification.

Otherwise, OpenAI made a notable move to expand the power of ChatGPT as a kind of platform. The company said customers would be able to sign into other applications they use, such as Cognition, Notion and Vercel, with their ChatGPT accounts. Those accounts typically come with a specific usage limit, and users will be able to use the tokens allowed under their ChatGPT plan towards those other apps, just by using those other apps the way they normally do. (This initiative has been long in the works: We first reported on this effort a year ago.)

OpenAI also sought to catch up to Anthropic in launching an app store of sorts. Known as OpenAI Marketplace, the feature lets OpenAI enterprise customers that have committed to spending a certain amount on OpenAI products pay down that commitment by using AI-powered apps from Adobe, Salesforce, Harvey and Palo Alto Networks. The aim is to lock more customers into OpenAI’s ecosystem (those other apps are powered in part by OpenAI models) and encourage them to make bigger spending commitments. (Anthropic has long had such a program, Claude Marketplace.)

Outside the venue, protestors criticized OpenAI’s work with the U.S. military and immigration enforcement, chanting, “Hey girl, what’s tea? It’s time to quit GPT.” For its part, OpenAI said ChatGPT’s weekly active users rose to 1.2 billion, compared to around 1 billion in July.

Here’s what else is going on…

GLM Pro, GLM Con

Anthropic and OpenAI on Tuesday took very different approaches to handling Chinese developer Z.ai’s latest GLM open-weight model.

OpenAI said enterprises could use GLM or (Moonshot’s Kimi K3) through its coding assistant Codex, and the usage would count toward whatever spending commitment the enterprises had made to OpenAI. (Inference provider Baseten, part of the aforementioned OpenAI Marketplace, is hosting the models.) In other words, OpenAI may be trying to turn Codex into a quasi-open platform for customers to pick and choose the models they want.

On the other side is Anthropic, which posted a warning about GLM, saying Anthropic’s tests showed the Chinese model lacked safeguards to prevent malicious people or governments from using it to hack websites and other software systems.

“The release of GLM-5.3 is a meaningful step change in the cyber capabilities available to attackers,” Anthropic’s researchers said.

Both OpenAI and Anthropic now sell their own cybersecurity-focused models to select governments and businesses to defend themselves. However, the process of selecting which entities get access to such advanced models has been thoroughly confusing, as it involves getting U.S. government approval, my colleague Leo has reported.

Anthropic, in its post about GLM, seemed to be asking for clarity or the ability to release such technology to more businesses. “GLM-5.3 underscores the urgency of expanding access to advanced frontier models to a broader set of entities to empower cyber defenders,” the post said.

As a reminder, Anthropic also stands nearly alone in Silicon Valley in effectively opposing the release of open models with advanced capabilities, including Chinese ones. Anthropic’s protestations on this subject go back at least several years. Still, it isn’t clear how Anthropic will prevent such usage.—Amir Efrati

Big Number

OpenAI is in early talks with investors to raise a pre-IPO round and is targeting raising $30 billion, according to a person familiar with the matter. The ChatGPT maker could seek a valuation around $1.4 trillion, though it hasn’t signed a term sheet yet with any investor, the person said.

Overheard

President Donald Trump announced on Tuesday that leading AI companies had agreed to non-binding commitments on AI safety provisions, including internal controls on model alignment and external auditing. Separately, President Trump signed an executive order formally instructing his administration to stop using the terms “artificial intelligence” and “AI” in favor of “super intelligence” and “SI.” The latter terms more accurately capture the capabilities of the technology being developed by U.S. labs, according to the order.

OpenAI is nearing $70 billion in annualized revenue, up around 70% from its revenue pace at the beginning of the third quarter, according to a person briefed on the numbers.

Meta Platforms is adapting its AI agent Muse for small businesses, adding integrations with workplace software including Salesforce’s Slack, Intuit QuickBooks and Zoom.

Google is in the process of joining Apache Ossie, a software industry group with members such as Snowflake and Nvidia that is trying to make it easier for AI tools to access data from applications and databases, according to a company spokesperson.

Deals and Debuts

See The Information’s Generative AI Database for an exclusive list of private companies and their investors.

Anthropic has agreements to pay SpaceX up to $84.5 billion to use the rocket and AI company’s Nvidia-based computing capacity through 2029, according to a confidential initial public offering prospectus reported by Reuters on Tuesday.

GMI Cloud, a five-year-old provider of Nvidia chip servers to enterprises, said Wednesday it raised $668 million in equity and debt financing. The new fundraise indicates Nvidia is continuing to help smaller, upstart cloud providers, especially those that plan to boost Nvidia in more ways than one.

General Intuition, which trains AI models on video-game footage to help robots learn, raised $220 million at a $6.2 billion valuation in a funding round led by Valor Equity Partners and Atreides.

Kahua, which makes AI software for managing large construction projects, received a minority growth investment from Bain Capital's Tech Opportunities fund at a valuation above $1 billion.

Efficient Computer, which builds energy-efficient computer chips, raised $97 million in a Series B funding round led by TQ Ventures at a $650 million valuation.

ai3Bio, which uses an AI discovery platform to find treatments for autoimmune disease, raised $48 million in a Series A funding round led by UPMC Enterprises and Ziff Capital Partners.

Ascerta, which helps businesses track the ROI they’re getting on their AI, raised $18 million in Series A funding led by Dell Technologies Capital.

Bolt.new acquired Dokai, a startup that built AI agents for enterprise sales teams, to strengthen its own AI coding agents.

Amazon introduced new AI-powered tools to help marketers buy ad inventory across the vast variety of media it sells, as it steps up its effort to strengthen its position in the ad tech world.

Thank you for reading the AI Agenda Newsletter! I’d love your feedback, ideas and tips: [email protected].

If you think someone else might enjoy this newsletter, please pass it forward or they can sign up here.