/
Published on 12.03.2026
TLDR: OpenAI is reportedly pulling Sora into ChatGPT after the standalone video generation app saw installs plummet by 45 percent. Meanwhile, NVIDIA open-sourced a 120-billion parameter model that cleverly activates only 12 billion at inference time, and OpenAI gave developers full shell access for building agent workflows through the Responses API.
Let us start with the Sora situation because it is a fascinating case study in product strategy gone sideways. OpenAI launched Sora as a standalone video generation app, presumably betting that video AI deserved its own surface area, its own brand, its own install base. The market disagreed rather emphatically with a 45 percent drop in installs. Now the plan is to fold it back into ChatGPT. This is not surprising if you think about it from a user behavior standpoint. People do not want another app for a capability they expect to find in the tool they already use every day. The lesson here is one the industry keeps relearning: distribution matters more than novelty, and bundling usually wins over standalone plays for AI features. The question worth asking is whether Sora's capabilities will actually shine inside ChatGPT's interface, or whether they will get buried in an increasingly crowded feature set.
NVIDIA's release of Nemotron 3 Super is the more technically interesting story this week. It is a 120-billion parameter model with a one-million token context window, but here is the architectural trick that matters: it uses a hybrid approach where only 12 billion parameters are activated during inference. That is a 10x reduction in compute at runtime, which NVIDIA claims delivers 5x faster agent performance. This is a meaningful contribution to the open-source ecosystem because it directly addresses the practical bottleneck everyone faces: you want the reasoning quality of a large model without the inference cost that makes it impractical for real-time agent loops. If the benchmarks hold up in production settings, this could shift the calculus for teams deciding between hosted API calls and self-hosted inference. Worth watching closely.
The third piece worth your attention is OpenAI adding computer environments and shell tools to the Responses API. This gives developers hosted containers where their agents can execute code, manipulate files, and run arbitrary shell commands. Think about what this means architecturally: instead of building fragile tool-calling chains where the model requests actions and your code executes them, you can now drop an agent into a sandbox and let it operate more autonomously. The security implications are obvious and worth taking seriously, but the developer experience improvement is substantial. This is OpenAI making a clear bet that the future of their platform is not just chat completions but full agentic workflows with real compute environments.
In the broader ecosystem, Anthropic launched the Anthropic Institute under co-founder Jack Clark to study societal impacts and governance of powerful AI. This is a meaningful move because it signals that at least some labs are investing real institutional resources in governance research rather than just publishing position papers. Perplexity also launched Computer for Enterprise, which routes tasks across 20 different AI models with over 400 app connectors. The multi-model orchestration approach is interesting because it acknowledges what practitioners already know: no single model is best at everything, and the real value is in intelligent routing.
On the investment side, Cursor is reportedly in talks at a roughly 50-billion dollar valuation, nearly doubling from last fall. That is a staggering number for a code editor, but it reflects the market's conviction that AI-native developer tools are going to capture enormous value. Rhoda AI raised 450 million in a Series A for robot foundation models, and OpenAI acquired Promptfoo, an AI security platform, to integrate into their Frontier offering. The Promptfoo acquisition is particularly telling because it suggests OpenAI recognizes that enterprise adoption hinges on security tooling being built into the platform rather than bolted on afterward.