# Grok 4.5 Private Beta: 1.5T Params, Cursor Data, xAI’s Machine > Source: > Published: 2026-06-29 09:11:07+00:00 On June 28, 2026, Elon Musk confirmed Grok 4.5 entered private beta at SpaceX and Tesla. No public API. No external benchmarks. Just 1.5 trillion parameters running inside two of the most data-rich industrial operations on the planet. That’s not a soft launch — it’s a reinforcement learning pipeline wearing a product rollout costume. ## V9 Is Not a Fine-Tune — It’s a Ground-Up Rebuild Grok 4.5 runs on xAI’s new V9 foundation, and the parameter jump is significant enough to matter. The current Grok 4.3 you can call via API today sits at 500 billion parameters (V8-small). Grok 4.4 — released in late May — stepped to 1 trillion (V8-medium). V9 lands at 1.5 trillion: three times the current production model’s size, and 50% larger than the prior private beta. More important than the number: V9 was trained from scratch. Not fine-tuned. Not adapter-layered. [Colossus 2](https://x.ai/colossus) — the world’s first 1.5-gigawatt training cluster, operational since April 2026 — ran a full ground-up training run. xAI is simultaneously training seven model variants on that cluster right now, ranging from 1T to 10T parameters. Grok 5’s path to 10 trillion parameters has already started. ## Cursor Data: Training on How Code Gets Written, Not What It Looks Like Here is the genuinely novel part of the V9 story. SpaceX acquired Cursor for $60 billion on June 16, 2026 — the largest acquisition of a venture-backed startup in history. The day that deal was formalized, Cursor announced it had trained a 1.5 trillion-parameter model from scratch on Colossus. That is Grok 4.5. Standard coding models — including GPT-5.5 and Gemini 3 Flash — train primarily on GitHub repositories. That means they learn from finished code: the output. [Cursor’s data is different](https://www.techtimes.com/articles/318974/20260624/cursor-trains-first-frontier-model-scratch-colossus-15-trillion-parameters.htm). With over one million paying users and 50,000 corporate clients, Cursor captures what happens before the code is committed. Which AI-generated suggestion got accepted. Which one got rewritten. How a developer iterated through three refactors before settling on a pattern. That signal — what researchers are calling “human developer thought mapping” — is qualitatively different from static repo training. No other frontier lab has this dataset. OpenAI doesn’t own a code editor with eight-figure user counts. Anthropic doesn’t. That’s xAI’s actual moat here, and it matters more than the parameter count. ## SpaceX and Tesla as Live RLHF Loops Running the model at SpaceX and Tesla before public release is not just dogfooding. It’s reinforcement learning at industrial scale. Every time a SpaceX engineer accepts or rejects a Grok 4.5 suggestion, that signal goes back into the training pipeline. The same is true at Tesla’s software teams. No competitor has a comparable live industrial deployment generating this kind of feedback. OpenAI uses ChatGPT users. Anthropic uses Claude.ai. Both are consumer-grade signal. xAI is running RLHF inside operations that build rockets and autonomous vehicles. The signal density is different in kind, not just degree. Based on Grok 4.4’s timeline — which went from private beta to public API in roughly four weeks — expect Grok 4.5 general availability around late July or early August 2026. ## About That “Rivals Claude Opus” Claim Musk stated that internal evaluations put Grok 4.5’s performance close to, or exceeding, Anthropic’s Claude Opus. Worth unpacking what that claim rests on. Claude Opus 4.8 currently holds the top SWE-bench Verified score at 88.6% — the highest public benchmark score among all frontier models as of June 2026. Grok 4.3, the current public API model, has no official published SWE-bench number from xAI. [Third-party estimates](https://lmcouncil.ai/benchmarks) put it around 75%. That’s a 13-point gap to close. xAI has claimed parity at launch before. Grok 4’s release announcement implied GPT-5.5 equivalence on coding tasks; independent testing found meaningful gaps. “Rivals Opus” may be accurate — or it may be aspirational. There is no way to verify until public benchmarks are published. Treat the claim accordingly. ## What Developers Should Do Right Now Grok 4.5 is not available through the API today. Grok 4.3 ($1.25 per million input tokens, $2.50 per million output tokens) remains the production-ready xAI option. Based on parameter-scale pricing precedent from 4.3 to 4.4, expect Grok 4.5 API pricing in the $2–3 per million input token range. If you’re evaluating the xAI stack: xAI still offers $175 per month in free API credits via their data-sharing program — the most generous free tier among major providers. Enroll at the [xAI developer console](https://docs.x.ai/developers/models) under Settings > Data Sharing. Watch three things: the xAI API changelog for a Grok 4.5 GA announcement, the [LM Council benchmark leaderboard](https://lmcouncil.ai/benchmarks) for V9 public results, and the Cursor IDE roadmap — post-acquisition close in Q3 2026, Cursor will likely begin preferencing Grok V9 models over its current multi-provider support. Grok 5, built on a 6–10 trillion parameter architecture, is already in training at Colossus 2. At xAI’s current pace, it arrives before the end of 2026. The monthly release machine is running.