OpenAI’s DevDay, Google’s First Gemini 4 Model, Black Forest Labs Dives Into Robots
The Batch Analytics DeepLearning.AI ● Covered by 20 sources
OpenAI cut the price of a strong new model and shipped always-on agents. Google answered with a long-context model for risky work, but only for trusted defenders.
Based on reporting by The Batch, Analytics DeepLearning.AI — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
OpenAI used DevDay to push a simple message: the cheaper model is getting close to the expensive one. GPT-6.1 Sol lands at one-fifth of GPT-6 Astra’s standard per-token price, and on Artificial Analysis’ Intelligence Index it came within a single point of Astra while beating GPT-6 Sol. That matters because the people chasing agentic workflows care about cost almost as much as raw capability. If a model can think for less, it gets used for more steps, more often, and in more places.
The company paired that with Dots, always-on personal agents built on GPT-6 Astra. They can connect to more than 4,000 apps, run through their own cloud computer and browser, and take instructions in ChatGPT, Slack, and Microsoft Teams. Users can set rules about what they may do on their own, and sensitive actions like password changes need approval. It’s a neat product pitch, but it also shows where OpenAI is placing its bet: not just better models, but systems that keep working after the chat window closes.
There’s a catch. OpenAI says GPT-6.1 Astra was pulled before launch because it didn’t meet the company’s security bar. That makes Dots look even more interesting, because they run on GPT-6 Astra, not the cheaper Sol model that just impressed on benchmarks. The result is a company that is both celebrating guardrails and quietly admitting that the guardrails still decide what ships.
Google answered with Gemini 4 Argon, a long-context model aimed at software engineering, law, finance, and cybersecurity. It can take in text, images, and video up to 1 million tokens, and its outputs can stretch to 1 million tokens too, using Long Decode Continuation when responses would otherwise time out. For now, though, wider access is tightly fenced off: only organizations in Google’s Fairwind cybersecurity program can use it. That’s the recurring pattern here. The models keep getting more capable, and the companies keep making access narrower when the stakes go up.
My take — AI-written commentary, not fact-checked reporting
The real story is not that these models are smarter; it’s that everyone is suddenly talking like security is a product feature and not a bill you eventually pay. The closed-model crowd loves to ship “guardrails” after the demo is already booked, which is a bit rich for software that wants to run your life. Open systems still have the better argument: let outsiders test the thing before it starts making decisions in your Slack.
Read more about this at: The Batch