Simon Willison writes about the recent Cursor changes:

Firstly, the era of VC-subsidized tokens may be coming to an end, especially for products like Cursor which are way past demonstrating product-market fit.

First, have some epistemic disclaimers:

1. I am a happy and paying user of Cursor (the past two months I averaged ~$5 of overages above the $20/mo plan).
2. Even predating this latest situation, I am fairly bearish on Cursor’s prospects as a company: I haven’t found any of the arguments as to how they’ll protect their marketshare as meaningfully cogent or persuasive.
3. I also pay for Claude’s chungus-sized $200/mo plan 1 ^[1], and use the hell out of it.

That aside: I love — and find especially apt — the phrase “VC-subsidized tokens”.

In 2015, Lyft would charge me $7 for a ride from West Seattle to Belltown; Uber would charge me $5. It was not clear to me if this was because Uber had better magic or because they were more aggressively loss-leading. I certainly did not (nor did anyone I know) have "loyalty" to one or the other — we simply opened both apps and chose the cheaper quote.

In 2025, both companies would charge me $30 — a price that is less beggaring of disbelief, and also a price that wouldn’t cause me to say “okay, time to call a cab company” but would cause me to look up how delayed the C Line has been lately 2 ^[2]. It’s unambiguous that the ridehailing wave resulted in an objectively better user experience for getting a cab; it’s also very clear that a lot of the temporary shift towards these services was more about taking advantage of very generous a16z subsidy and less about enduring behavioral change. 3 ^[3]

I use these tools every day; I find myself starting to shift from a “the marginal costs round down to zero!” mindset to scrutinizing the bill of a random Claude Code session as if I ordered a burger and a beer at an airport bar and the bottom of the check reads $80. (“$12 to fix two linting errors in the OpenAPI spec? Really?”)

As always, I remain particularly interested in the field of open source models, not just for the obvious moral and custodial benefits but increasingly for the economic benefits. If you subscribe to the broad idea that open source models will lag a year or so behind frontier models, a lot of these conversations shift towards discussions of trading price for speed, rather than accuracy. (And for much of my Claude Coding, which is of the genre “hey, go run off and do this thing that I thought of while walking the dog”, I do not care about the wall clock time so much as I care about the quality of the final artifact.)