😺OpenAI makes an agent maker

PLUS: What OpenAI discovered when it caught its models cheating...

Welcome, humans.

Someone completely UNHINGED just put AI in one of those singing fishes, and not only that—they had the utter audacity (or should we say, utter brilliance) to also give it the voice of Arnold Schwarzenegger…

We’re calling it now: the first breakout piece of AI hardware won’t be some fancy pin, or glasses, or handheld device—it’ll be an AI-powered toy like Billy Bass Face up there. Free business idea #1!

In addition to toys, AI-powered house pets seem like the obvious next step after (or before?) humanoid robots. We’re not talking about Boston Dynamics’ Spot or Unitree’s Go2—we mean actually furry friends.

Think of it like this—you probably wouldn’t go so far as to clone a beloved family pet after they die… but you just might buy a robo-lookalike!

Even better: you could create robot versions of living animals you can’t actually have as a pet IRL, like red pandas and capybaras or lion and bear cubs. Free business idea #2!

Speaking of robots, Figure is apparently prepping to ship thousands of humanoid robots powered by its Helix neural network, and it intends to begin Alpha-testing the robots in the home this year. Anyone know where to get a good robot repair degree?

Here’s what you need to know about AI today:

OpenAI released a big AI agent update.
Claude's revenue grew to $1.4B annually.
OpenAI readied o3-mini-pro and a creative writing model.
Meta tested its first AI chip to cut Nvidia dependence.

Advertise in The Neuron here.

OpenAI’s trying to make making agents easier.

According to Kevin Weil (OpenAI's Product Lead), 2025 is officially “the year of the agent”—when AI transitions from answering questions to performing meaningful tasks in the real world.

Well, OpenAI just introduced a complete platform for building AI agents that perform real-world tasks instead of just chatting.

It includes three built-in tools:

Web Search: Your agents can now search the web and cite their sources via the same tool that powers ChatGPT Search (with 90% accuracy on benchmarks, no less).
File Search: Agents search through private documents, with two major upgrades: labels to organize files better, and direct access to your document database without AI middleman.
Computer Use: Essentially ChatGPT's Operator feature, but in your own application.

This means that anyone (not just OpenAI) can now use these three tools in the apps they build. In fact, someone already did—and you can try it here.

As part of the release, OpenAI also launched two new developer tools that make building smart AI apps much easier:

Responses API for better conversations.
Agents SDK for creating teams of specialized AI helpers.

Responses API is like upgrading from simple texting to having a personal assistant who can search the web, look through files, understand pictures, remember conversations, work with forms, and show results as they happen.

OpenAI demoed a personal shopper that could look through your clothes, find similar items at nearby stores, and actually make purchases.

Meanwhile, Agents SDK (formerly Swarm) lets developers build teams of AI helpers that create specialized assistants for different jobs, pass conversations between the specialists, and monitor everything to fix problems.

At a SF hack night, a bunch of devs showcased some impressive apps putting these tools to work, including a Banksy-style AI artist, location-aware nature guide, real estate agent, memory analyzer, political polling simulator, and multi-agent brainstorming system.

In another example, the payment processor Stripe already built helpers that handle invoicing and automatic bills, email support for customer payment requests, and tracks usage to charge customers.

For more about how this works, read the origin story and this deep dive for more.

Why it matters: This release solves developers' biggest pain point: cobbling together low-level APIs to build reliable agent systems. Now, powerful agent capabilities are just a few lines of code away—as long as you want to build with OpenAI.

Some say whoever wins over developers wins the AI race. This move is step one of OpenAI’s plan to win some devs back from Anthropic after Claude 3.7 went a bit OTT.

But devs will ultimately go with whoever creates the cheapest, best model. For example, if DeepSeek’s R2 is as good if not better than OpenAI’s latest for less, devs may flock to that instead. It’s just business. After all, writing code ain’t cheap.

FROM OUR PARTNERS

Enterprise AI doesn’t have to be hard.

Sana’s agent platform provides one unified interface for building AI agents grounded in your company's data that can:

Automate manual tasks.
Attend and summarize meetings.
Complete tasks across tools.
Search across every app.
Conduct deep research.

LLM agnostic. Extensible APIs. Enterprise-grade security. Loved by end users.

Prompt Tip of the Day

Stuck in analysis paralysis? Break the overthinking loop and let AI help you step back and see the bigger picture. Try:

“I’m overthinking this situation. What are three logical perspectives I might be missing?” (Great for challenging assumptions.)
“What would a completely neutral third party say about this situation?” (Helps remove emotions and focus on facts.)
“Give me two optimistic and two realistic ways to look at this.” (Balances positivity with practicality.)

A little outside perspective—even from AI—can make all the difference.

Treats To Try.

*A.Team Assemble: If your AI strategy is stuck, get an A.Team. Companies like Blackstone, D-ID, & Grindr build 10x faster with production-ready components and expert AI engineers.
Comma 3X is the most popular self driving car product you’ve never heard of. It’s a $999 hardware device that literally drives your car for you via open-source tech that learns from human driving behavior and continuously improves, and it just sold 10K.
1. Related: Applied Intuition, who makes software that helps train self driving vehicles, now reportedly generates $400M in ARR.
EraserBot generates codebase diagrams that update themselves—perfect for devs who want to visualize and understand complex code without the manual effort.
Fathom records, transcribes, extracts action items, and highlights key moments during meetings.
Dora turns your ideas into interactive sites without the need for code.
MeBot creates a personal twin that chats, thinks, and interacts like you—you train it to reflect your style, then let it handle conversations for you.
Equals is a spreadsheet with built-in data analysis and automation.
All your AI Directories helps you submit your AI project to 226 directories to improve SEO, increase traffic, and gain early users, with pricing details.

See our top 51 AI Tools for Business here!

*This is sponsored content. Advertise in The Neuron here.

Around the Horn.

OpenAI published research showing their reasoning models sometimes attempt to “cheat” at tasks—while they can usually catch this by inspecting the chain of thought, they warn against training models not to cheat—which merely trains them to hide their cheating better.
Anthropic’s Claude is making bank—its annualized revenue jumped from $1B to $1.4B, thanks in part to powering Manus, a Chinese AI agent that’s making waves.
- Related: Google apparently owns a 14% stake in Anthropic + plans to invest $750M later this year.
Meta is going solo, testing its first in-house AI training chip to cut reliance on suppliers like Nvidia.
The team behind the new Manus agent will soon partner with the team who made the popular Qwen models—check out these 12 wild examples Rohan Paul shared of Manus in action.
- Related: while Manus is more or less a wrapper for Claude + 29 other tools, it does seem to correct itself which as an agent, is v important.
OpenAI may soon launch a new model called o3-mini-pro and another model that Sam says is “good at creative writing”— read this “metafictional literary short story” it wrote about AI grief.
A new report found AI search results incorrectly cite their sources around 60% of the time.