Kimi dropped an open-weight model good enough to make Box's CEO publicly congratulate a Chinese AI lab, and the enterprise AI stack is quietly cracking open. Meanwhile, OpenAI's coding story is being rewritten in real time, and Google just made NotebookLM's name official after 30 million people started using it without one.
01
Kimi's open model is rewriting the enterprise cost calculation
Box CEO Aaron Levie posted that the performance coming from open-weight models like Kimi is "truly wild," and he's not being polite. The argument is straightforward: every time frontier-quality intelligence gets cheaper, the list of enterprise workflows worth automating gets longer. Workflows that were too expensive to run at scale six months ago are now viable. Levie is specifically pointing to token cost as the gating factor, not model quality, not integration complexity. ---
Why it matters: If your company shelved an AI project last year because the API costs didn't pencil out, the math is worth running again. Kimi-quality open models mean you can run more inference for the same budget, or run the same workload for less. The projects that got killed in procurement review are candidates for a second look.
An enterprise AI strategist's three-step checklist for surviving open-model chaos
Separately, Madhu Guru posted a more tactical take on the same Kimi moment: open-weight models are about to force a rethink of how enterprises choose and lock in AI vendors. His prescription is build rigorous evals before committing, distinguish between "regression" evals (the things that must always work) and "aspirational" evals (the harder cases you're trying to crack), and maximize model optionality so you're not stranded when the next open release reshuffles the rankings. ---
Why it matters: The enterprises that signed long-term contracts with a single AI vendor based on last year's benchmarks are going to feel this first. If your team hasn't built internal evals that are specific to your actual use cases, you have no way to know whether the next open model is better for you than what you're paying for.
The inside story of how OpenAI lost the coding race and then won it back
Dan Shipper, who runs Every, posted a detailed retelling of OpenAI's coding product arc. The short version: OpenAI launched GPT-5 in summer 2025 as a pair programmer and missed the shift toward agentic coding that Claude Code was pioneering. They bet on browser-based coding and vibe coding in ChatGPT, which was too early. A small internal team broke off to build a separate Codex model line. That bet is now paying off, and Shipper says OpenAI is "firing on all cylinders." ---
Why it matters: The lesson isn't that OpenAI recovered, it's that the agentic coding shift happened faster than the market leader anticipated. If you're building developer tools and your product assumes humans are still in the loop for each edit, that assumption is worth pressure-testing.
Boris Cherny on what it actually takes to run Claude at full autonomy
Boris Cherny, who works on Claude Code at Anthropic, posted a detailed breakdown of the features you need to actually unlock high-autonomy Claude usage. The list: end-to-end self-verification, auto mode for permissions, automated code and security review, multi-agent interfaces, and tools like /loop, /batch, dynamic workflows, and worktree isolation for subagents. The point is that getting to higher levels of autonomy isn't about one feature, it's about combining the right set of them correctly. ---
Why it matters: This is the most specific public accounting yet of what "agentic coding" actually requires in practice. If you're using Claude Code and wondering why your setup feels brittle, the gap is probably somewhere on this list.
Google makes NotebookLM's name official after 30 million users
Josh Woodward, who leads Gemini at Google, posted that what the team has internally called "Notebook" for years is now officially named NotebookLM externally. The product has reached 30 million users and 600,000 organizations. The announcement is more naming formality than feature news, but the scale is worth noting: this started as a small experiment inspired by watching author Steven Johnson describe his writing process.
Why it matters: 30 million users on a product that most people outside the productivity world couldn't have named six months ago is a quiet signal about where AI-assisted research and writing is landing. Your company's knowledge management problem has a product with real traction now.