There's a sense of relief across the developer community on social media. At the last moment, Anthropic decided to pause their billing change that would have forced all non-interactive sessions into per-token billing. The change, expected to happen yesterday, would have excluded third-party tool use like Conductor and programmatic work like GitHub PR reviews from the Pro and Max plans. While I understand the elation that I'm seeing online as people discover that they can still use their preferred tools, it's a bit distressing that developers and even entire companies are basing the way they write software on the promise of virtually unlimited inference for a low fixed cost.
In their announcement of the pause, Anthropic didn't give any specific guidance around when the change will actually happen and what its impact will actually be. They said "We’re working to update the plan to better support how users build with Claude subscriptions." The most charitable way to read this is that they're reconsidering the change entirely, having recognized that the ability to automate work is a core piece of Claude Code's value. But they didn't explicitly say that, and I'm not sure they've earned the most charitable read from anyone on this.
If we step back and think about it, the now aborted change was well considered and well communicated. Having struggled for months to find a way to segment what they consider to be beneficial automated use from abuse, they decided to just shut it all down. Customers were given plenty of notice to decide whether to continue using their tools for these automated workflows or migrate to something else. We could dislike the change in policy, but it was handled well. Brilliantly, when compared against several of their other recent changes.
Users of OpenClaw and OpenCode weren't given any notice period when they were cut off. Anthropic argued that the use of subscriptions within these products was a violation of their terms of use, and that's fair. But their terms of use weren't (and still aren't) written in a way that adequately defines appropriate use in a space where new ways of using it are invented pretty much every day.
Whenever we adopt a tool and make it a core part of our business, we are exposing our businesses to a level of risk. We are allowing ourselves to rely upon a third party to form a piece of our own foundation. This isn't new, and it's not unique to software. But it's a reality that, until recently, business owners have taken very seriously. There's something unique about AI that invites us to ignore the risk of bolting this thing onto our companies and entrusting it with mission-critical work. We allow ourselves to ignore the fact that it could be yanked out from under us at any moment. The price could go up by orders of magnitude overnight. We dismiss these thoughts with a vague self-assurance that it'll never happen because "AI is the future."
AI is the future. We can safely take that for granted. What we shouldn't take for granted is that Anthropic is the future of AI, or inference subscriptions are the future. The very loud subtext of the message from Anthropic today is that the end of subsidized automated inference is coming. It's just not coming today. We need to stop taking for granted that virtually unlimited access to frontier AI models will always be available and affordable.
Today's pause is a brief reprieve, and a reminder that we need to own our future. We should be leveraging the subsidized pricing right now and making as much use of these affordable tokens as we can. But we need to keep one eye on the future and ensure that as we entrust automated AI with mission-critical pieces of our business, we're doing it in a way that doesn't cripple us if a particular lab or vendor significantly changes their offering.