Did Anthropic Just Lose Its Biggest Customer?
Cursor was Anthropic's biggest customer. Cursor co-founder Aman Sanger said so in January 2025: "We are without a doubt their largest customer."…
Since SpaceX bought Cursor in August, it has been pushing Grok on every Cursor user, and I think that is slowly killing Cursor. The new plans from October 12 include usage for Grok only. On those plans Claude costs full API price from the first request, and the newest OpenAI models never came to Cursor at all. Composer, Cursor's own cheap coding model that people loved, hasn't had a new version since May. Grok is a decent model, but Cursor users came for the choice of models, and that choice is going away.
Grok is good, the others are better
I wrote in July that Grok was back, and I still think so. But look at CursorBench, Cursor's own benchmark built from real Cursor sessions. Grok 4.7 scores 46.3% at its best setting and costs $6.01 per task. Claude Opus 5.5 scores 56.0% at High effort and costs $3.97 per task. Sonnet 5.5 and Fable 5.1 also score higher than Grok at their top settings. If you can choose, you pick one of those for hard work.
Everything except Grok now costs extra
On the old plans, a Cursor subscription included usage for outside models. On the new X Subscription plans, Cursor staff confirmed: "There's no included allowance for third-party models on Premium, Super, or Ultra." Claude is billed at API list prices "starting with the first request." The $8 Plus plan only has Grok. Current subscribers can keep their old plans for now, but Cursor says they will be moved over later.
OpenAI models are worse off. OpenAI plans to end Cursor's access on November 12 and said it would not provide future models to Cursor. Cursor's model list stops at GPT-5.6, so GPT-6 Astra and GPT-6.1 Sol never came to Cursor. After November 12 you can bring your own OpenAI API key, but only for local Chat and Agent.
Composer is the bigger loss
What Cursor users seem to miss most is Composer. Composer 2.5 came out on May 18. It is Cursor's own coding model, built on Moonshot's open Kimi K2.5, and it costs $0.50 per million input tokens and $2.50 per million output tokens. It is no frontier model. It is cheap, though: a CursorBench task costs $0.68 with it, against $3.49 for Grok 4.7 at Medium. For everyday coding, a lot of people used it as their daily driver.
Since then Grok got three new versions in Cursor: 4.5 in July, 4.6 in August and 4.7 in September. Composer got none. A forum thread called Please release Composer 3 soon has more than 5,000 views. Cursor staff answered that "it is still our plan to release new models of this size going forward!" That is not a promise of a Composer 3, and I don't think one is coming. One user put it well: "If Cursor becomes nothing more than a Grok harness, see you later alligator."
Choice was the whole point
Cursor's homepage still says: "Choose between every cutting-edge model from OpenAI, Anthropic, Gemini, SpaceXAI, and Cursor." That was the reason to pay for Cursor. A long-time Ultra subscriber wrote that Cursor earned its place next to Codex and Claude Code because it was "one editor where you can use the best models."
That promise was never cheap to keep, and for SpaceX it makes business sense to push its own model. For users it sucks. If you want Claude or GPT, Claude Code and Codex CLI give you their lab's newest models with a subscription. Other coding tools usually let you pick the model and pay the provider. Cursor without the choice is a Grok harness, and I think SpaceX is slowly killing it that way.
What I'm buildingDelegate 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