Something just flipped. According to expense data from fintech firm Ramp, 34.4% of participating businesses are now paying for Anthropic services. OpenAI sits at 32.3%. That’s not a huge gap, but the direction matters. Six months ago, OpenAI was the default enterprise AI vendor. Now it isn’t, at least by this measure.

Ramp’s methodology is worth noting: this is real spend from their clients’ expense reports, not a survey asking what tools people prefer. Companies are actually paying for Anthropic. That’s a harder signal than any usage poll.

The same week, Anthropic launched Claude for Small Business, a new tier aimed at smaller teams. The timing isn’t subtle. If you’re going to announce you’ve taken the business customer lead from OpenAI, launching a product that makes it easier for more businesses to become customers is a smart move to make in the same breath.

The broader context: OpenAI has been pushing hard on the enterprise side with GPT-4o and its operator tooling, while Anthropic has been quietly building a customer base that apparently just grew past them. Clio, the legal tech company, hit $500M ARR this week partly on the back of AI adoption, and Anthropic is one of the beneficiaries of that legal-tech spending wave.

Notion Tries to Become an Agent Platform

Notion shipped a developer platform this week that lets teams plug AI agents, external data sources, and custom code directly into their workspace. The pitch is that Notion becomes the hub where agents live and operate, not just a place to take notes.

This is worth watching carefully, not because the concept is new, but because Notion has actual distribution. Millions of teams already live in Notion. If they can make it genuinely easy to wire up an agent that reads your project docs, updates your task list, and talks to your other tools, that’s useful. If they can’t, it’s another checkbox on a features page.

The skeptic’s take comes from Boris Mann, who noted this week that “11 AI agents” is as meaningful a phrase as “11 spreadsheets” or “11 browser tabs.” Having a lot of agents doesn’t mean much on its own. The question for Notion is whether their platform makes the agents you actually need easier to build and run, or just easier to count.

For developers, this is worth a look if your team already runs on Notion and you’re building internal tooling. The integration surface is already there. Whether the developer platform is robust enough to build on is a different question.

Adaption’s AutoScientist: Automated Fine-Tuning

Adaption launched AutoScientist, which takes an automated approach to model fine-tuning. The idea is that instead of a human ML engineer manually iterating on fine-tuning runs, AutoScientist automates the process of adapting a model to specific capabilities. Think of it as fine-tuning that runs experiments on its own.

The positioning is ambitious. “Models training themselves” is a phrase that should make you read carefully before getting excited, because it usually oversells what’s happening. What Adaption is describing sounds more like automated hyperparameter search and data pipeline management for fine-tuning workflows, which is genuinely useful for teams doing a lot of customization work, even if it’s not quite the recursive self-improvement the headline implies.

If you’re an ML team running fine-tuning at scale and spending significant engineering time managing those workflows, this is worth looking at. If you’re an app developer who occasionally fine-tunes a model, it’s probably not for you yet.

OpenAI Built a Real Sandbox for Codex on Windows

OpenAI published a post this week about how they built the security sandbox that makes Codex work on Windows. The short version: running a coding agent safely on Windows required building controlled file access and network restrictions from scratch, because Windows doesn’t have the same container primitives Linux does.

This is the kind of infrastructure post that doesn’t get enough attention. Coding agents that can write and run code are only useful if you can trust that they won’t do something you didn’t ask for. The sandbox work is what makes that promise real. The fact that OpenAI documented it publicly is useful for anyone building similar tools.

Separately, Simon Willison built the Datasette project blog using OpenAI Codex desktop, and noted that Codex’s Markdown session transcript export was a feature he’d been wanting for a long time. It’s a small thing, but the ability to export a full record of what an AI coding session did is genuinely valuable for reproducibility and review.

Who Should Pay Attention to What

The Anthropic/OpenAI market share story matters if you’re evaluating which platform to standardize on for your team. The numbers suggest Anthropic’s quality reputation is translating to real enterprise adoption, and Claude for Small Business lowers the floor for getting started.

Notion’s developer platform is worth a prototype if you’re already embedded in Notion and want to build something for your team. Don’t bet a production system on it until you see how the ecosystem develops.

AutoScientist is for ML infrastructure teams. If that’s not your job, skip it.

The Codex sandbox post is worth reading if you’re building any kind of code-execution agent and thinking about the security model. It’s practical and specific, which is rare for this kind of infrastructure writing.