Parameter Update: 2025-25

"still missing claudius" edition

Parameter Update: 2025-25

Bit of a slower week again. Some Cursor drama and very little else going on right now.

Cursor: Updates & Rate limits

Just one week after Google rolled out the incredibly generous Gemini CLI free tier, it would appear that the days of VC-subsidized tokens may be coming to an end over at Cursor. While they are continuing on their journey to transition away from being "just" a VS Code fork by rolling out a mobile app (as well as other improvements like integrated TODO-lists, searchable PRs,...), they also rolled out the pricing changes first announced together with their $200/month "Ultra" plan.

Here's the summary as I understand it:

  • Before, they allowed a flat 500 requests/month to external models (anything that is not their atrocious "auto" mode). They then realized that people were overwhelmingly using these on Agent calls, which were much more costly than they anticipated.
  • Now, they are instead giving you "at least" $20 of API credits to spend on requests as you see fit. Unsurprisingly, those aren't getting people nearly as far as the old pricing model.

People are mad about this change, and I can fully understand why - Cursor spent the better part of two years getting people used to the full vibe-coding paradigm, only to now effectively charge them hundreds of dollars for that style of work. Personally, my Cursor Student credits ran out yesterday after renewing three days ago and, if I were to continue my usage behavior) I would be looking at ~$1000 in API spend. They also didn't do an amazing job communicating the change, leading to a bunch of people racking up surprise bills (though they have since announced they would refund affected customers). Somewhat surprisingly though, they are not walking the change back at all, sticking to their guns in their "apology" blog post.

Check out Simon Willison's blog for a more in-depth look.

Meta: Superintelligence Team

Meta has finally given us our first official look at the much-discussed "Superintelligence" team Zuckerberg has been cooking up for a while now.

Led by Alexandr Wang, co-founder of Scale.ai, the new team is composed of some incredibly high quality talent, mostly headhunted from other AI labs:

While Altman was quick to shrug off Meta's attempts before, proudly proclaiming that 'Missionaries will beat mercenaries', the reaction now seems more severe. OpenAI employees are apparently talking about the poaching feeling "like someone broke into our home" and the company gave almost everyone the week off to "recalibrate comp" (though the latter may or may not be normal practice for 4th of July?).

Perplexity Max

Following in the footsteps of OpenAI, Anthropic, Cursor and plenty more, Perplexity has announced a new $200/month "Max" plan, giving people:

  • Unlimited access to their "Labs" data visualization product
  • Early access to new product features
  • Advanced model options like Claude 4 Opus and o3-pro (!)

While I can see the value in some of what they're offering, I have a bit of a hard time imagining what the target demographic for this subscription might be. I highly doubt that people have the money for multiple $200/month subscriptions (one is already pushing it!), and for most people ChatGPT will probably provide better value, no?

Cloudflare: Pay-per-Crawl

In quite a momentous shift, Cloudflare (which control 1/4th of all internet traffic!) have announced two changes to how they plan to deal with AI crawlers:

  1. Crawlers will be blocked by default, unless the website owner explicitly opts-in
  2. Website owners will also have the ability to enable a "pay-per-crawl" model that appears to be a prominent step towards monetization of website content as AI training data

While both of these seem like great steps at first glance, some significant questions remain. For example, what would stop someone like Google from abusing their market position and simply down-rank websites that refuse their AI crawlers? Additionally, what would stop someone from just taking my website content and re-hosting it with a crawler-paywall? (this feels illegal and might genuinely be, idk?)

Either way, pay-per-crawl is currently in private beta while the blocking-by-default appears to already be rolled out globally(?)

xAI: Grok 4

Apparently just missing the cut-off for this week, xAI is currently preparing to launch Grok 4, planned for "just after July 4th". So unless Musk's new 'America' party has an impact on that timeline, we should get more info any minute now. So far, we have seen some very funny changes to the Twitter bots behavior and some leaked benchmarks that seem extremely promising. Seems like it will be an interesting week!