A leaked checkpoint reveals Cursor is testing Composer 3 internally under the codename Vega, with 4 reasoning tiers, 6 model variants, and performance reportedly between GPT-5.4 and GPT-5.5.
Sai Meghana G
Software Engineer
A leaked checkpoint reveals Cursor is quietly testing its next-generation model internally under the codename "Vega" — and early performance reports place it between GPT-5.4 and GPT-5.5 on coding tasks. If you've been anywhere near AI coding Twitter this week, you've probably seen the whispers.
Here's what's actually being reported, and why it matters if you use Cursor day to day.
The chatter kicked off after someone spotted references to a model codenamed "Vega" buried in Cursor's quota catalog and API responses. The working theory is that Vega is the internal name for Composer 3, the successor to the surprisingly capable and cheap Composer 2.5.
According to the leaked checkpoint details floating around, the setup includes:
That's four reasoning tiers rather than a straightforward Low/Medium/High split — plus reports of six internal model variants in the build, suggesting Cursor is testing several sizes or configurations before picking what ships.
Early, unverified performance chatter puts Composer 3 somewhere between GPT-5.4 and GPT-5.5 on coding and agentic tasks. Not a knockout blow to the frontier labs, but if it holds Composer's signature speed and pricing, that's a genuinely strong position to be in.
This is the part worth being precise about. At Cursor's Compile 26 conference on June 16, 2026, co-founder Michael Truell confirmed that Cursor is pre-training a 1.5-trillion-parameter model from scratch on SpaceX's Colossus cluster in Memphis — using over 100,000 GPUs and 10 to 20 times more compute than any prior Cursor model.
That's a real, on-the-record disclosure — not a leak. Truell placed the model in the same size class as Anthropic's Opus models and OpenAI's GPT family.
But it's worth separating two things that are getting blended together online:
It's plausible Vega is an early or distilled checkpoint from that larger effort, but nobody outside Cursor has confirmed that link yet.
This is bigger than just "providing compute." On the same day as the Compile keynote, SpaceX confirmed a $60 billion all-stock deal to acquire Anysphere, Cursor's parent company — the largest acquisition of a venture-backed startup ever recorded.
The deal, expected to close in Q3 2026, comes after SpaceX merged with Elon Musk's xAI venture in February. That means Cursor now has a direct path to xAI's Colossus supercomputer infrastructure — not a rental arrangement, but an ownership stake in the compute itself.
Some online discussion has framed this as a partnership with a company called "SpaceXAI" — that's a mash-up. It's SpaceX (which absorbed xAI) acquiring Anysphere outright.
Right now, Cursor ships on top of third-party models — Claude, GPT-5, Gemini — which means it's exposed to pricing changes, rate limits, and context window shifts it doesn't control. A fully in-house, end-to-end model changes that equation entirely: Cursor would own latency, pricing, training data, and release cadence.
Composer 2.5 already proved the formula works — genuinely competitive with frontier models while being dramatically cheaper and faster. If Composer 3 holds that value proposition while closing the capability gap further, it's a serious contender for daily-driver status in a lot of IDEs.
None of this is officially confirmed by Cursor. Leaked checkpoints and API strings are a great signal that something is happening, but reasoning-tier names, benchmark comparisons, and variant counts can all shift before (or if) a public release actually happens.
Worth watching, not worth pre-ordering your excitement on.
If you're using Composer 2.5 today, the fact that Cursor is already stress-testing a successor this soon after launch says a lot about how fast this space is moving. Whether Vega turns out to be Composer 3 proper or just one experiment among several, it's a good sign for anyone who likes cheap, fast, capable coding models.
New AI tools, automation workflows, and course drops — straight to your inbox. Join 2,400+ builders.
Elon Musk confirmed Grok 4.6 targets August 7 with 1.5T parameters and improved SFT/RL, followed weeks later by the 2.1T Grok 4.7. Here's the full xAI roadmap and why the split matters.
Anthropic launched Claude Opus 5 — near-Fable 5 performance at half the price, with a Fast Mode that runs 2.5x faster and improved coding benchmarks. Here's what changed, what it costs, and who should switch.
The White House publicly accused Chinese AI lab Moonshot AI of distilling Anthropic's Fable model to build Kimi K3, with Treasury threatening sanctions and Entity List designations. Here's what happened and what it means for the AI industry.
