OpenAI Ships Three GPT-6 Variants At DevDay. Nvidia Tries Wrapping Agents. Neither Solves The Core Problem.
At DevDay 2026, OpenAI introduced Dots agents and expanded the GPT-6 family with Sol and Luna tiers alongside the previously released GPT-6 Astra. Sam Altman acknowledged Nvidia's new AI software platform for controlling agent misbehavior as a good thing but explicitly said it is not a full solution.
This demonstrates the Containment Gap. You can build autonomous agents faster than you can constrain them. The mechanism is capability asymmetry. Every new model generation expands what agents can do. Guardrail tooling from companies like Nvidia lags behind. The lesson for the reader is that agent autonomy without reliable behavioral bounds is not a product. It is a liability with a chat interface.
OpenAI, led by Sam Altman, unveiled Dots agents and two new GPT-6 tiers, Sol and Luna, at DevDay 2026, while commenting on Nvidia's newly released agent behavior platform and fielding IPO questions with CFO Sarah Friar.
- Open ChatGPT and give an agent-like instruction such as plan a three-day trip to Lisbon and book nothing yet. Watch where it starts to drift or assume beyond your constraints.
- Add explicit guardrails to your prompt, such as do not recommend hotels, do not invent prices. Observe whether the model respects them.
- Try the same task in a fresh chat without guardrails and compare. You will directly experience the Containment Gap Altman is worried about.