Cursor is Anthropic’s Largest Customer (and Maxing Out Their GPUs)

Cursor (that VS Code-based IDE with AI features) is apparently hitting the limits of what Anthropic's GPU infrastructure can handle.

I noticed some threads about people having issues with Claude 3.5 Sonnet in Cursor lately, and it turns out there's an interesting reason. A Cursor dev straight up said in their forum that they're Anthropic's biggest customer and are basically maxing out their GPU capacity:

"Anthropic literally cannot sustain all of Cursor's traffic as they do not have enough GPUs... We are without a doubt their largest customer."

Anthropic has direct API customers, cloud provider partnerships, and presumably enterprise clients - yet a dev tool is their largest customer. That wasn't what I expected.

Some Background on how Cursor's AI Access Works

Every request to a "premium" model (GPT-4, GPT-4o, or Claude 3.5 Sonnet) goes through either their fast or slow pool:

• Free ("Hobby") users get 50 slow premium requests
• Pro users ($20/month) get 500 fast premium requests per month, plus unlimited slow requests
• You can buy extra packages of 500 fast requests for another $20/month
• When you run out of fast requests, you get queued into the slow pool

The slow pool is basically a fair queue system - you wait until a fast request slot opens up. Usually this works fine, but right now the slow pool for Sonnet is completely maxed out. Users are getting errors telling them to either enable usage-based pricing (costs $0.04 per request) or switch to a different model.

I'm not 100% sure what that means for "fast requests". Whenever it pops up I click the link to buy more fast requests/enable usage based pricing, but I still have issues with Sonnet not responding.

The kicker is that nobody wants to switch models. Developers are sticking with Sonnet, myself included.

Forum thread with all the details: https://forum.cursor.com/t/anthropic-cannot-sustain-additional-slow-request-traffic-on-claude-3-5-sonnet-please-enable-usage-based-pricing/41361

What I'm building

Delegate tasks. Get software.

Give Vroni a GitHub issue, bug report, spec, or rough idea. It reads the repo, plans the change, writes code, runs checks, and works toward a review-ready pull request.

Take a look at vroni.com