Elon Musk unveils Grok 5 with 1.5 trillion parameters, learning programming skills from Cursor.
[Introduction] Elon Musk announced late at night: The Grok V9 with 1.5 trillion parameters has completed training, three times that of the current model! Even more astonishing is that a large amount of Cursor programming interaction records were directly fed into the training data.
Elon Musk has revealed all his three programming tricks at once!
Just in the early morning of May 25th, Elon Musk personally announced on X:
The Grok base model V9-Medium with 1.5 trillion parameters has completed training, exactly three times that of the current model, and will be released to the public in 2 to 3 weeks.
Almost at the same time, even more explosive details emerged —
During the training process, xAI fed a large amount of Cursor programming data into the model.
In response, Musk directly stated that Grok's programming ability will be significantly improved.
Coupled with the $60 billion investment in Cursor and the launch of the Grok Build programming Agent, Musk's layout in the programming field has been fully revealed.
1.5 trillion parameters, Grok 5 is coming
V9-Medium is the internal version code of xAI, and the corresponding product name is almost certainly "Grok 5".
With a parameter scale of 1.5 trillion, it is exactly three times that of the current v8-small (500 billion parameters), and it has been specially optimized for NVIDIA's Blackwell architecture GPUs.
Even Musk himself is extremely dissatisfied with the current Grok — "There is too much junk."
V9-Medium is here to make up for this shortcoming.
The core means of making up for it lies in the phrase "a large amount of Cursor data" in the post.
Learning from Cursor: Reviewing for the exam with the answer sheet
There's no need to elaborate on the importance of Cursor.
More than 67% of Fortune 500 companies are using it, and it is expected to exceed $6 billion in annualized revenue by the end of 2026.
Jensen Huang publicly stated that Cursor is his favorite enterprise-level AI service.
Training Grok V9 with Cursor data is like reviewing for an exam with the answer sheet.
It's just that this exam is about writing code, and the answer sheet is the way engineers actually write code.
In a sense, this is both cheating and the right solution.
The reason behind this is that currently the bottleneck of LLMs is not writing basic syntax, but understanding complex engineering logic, multi-file code libraries, and real debugging workflows.
And Cursor has a large number of complete interaction records from real developers collaborating with AI in programming — prompt design, code modification, bug debugging, and multi-file collaboration.
Therefore, it can just fill this gap.
Netizen Peter Kuhar simply asked Grok directly "What exactly does the Cursor data contain", and Grok actually answered —
High-quality real programming interactions, including developers' prompts, code context, editing operations, and task completion records.
It's worth mentioning that in addition to the release of V9, the previous v8-small (0.5T) model will be open-sourced by the end of the year.
The programming suite is complete
Looking back at Musk's actions during this period, it's not hard to find that his layout in the programming field is a three - pronged attack.
Step 1: Lock in Cursor with $60 billion.
On April 21st, SpaceX announced an agreement with Cursor, obtaining an option to acquire Cursor for $60 billion. If the option is not exercised by the end of the year, a $10 billion cooperation fee will be paid.
SpaceX stated in a statement: "SpaceXAI and Cursor are working closely together to create the world's best AI for programming and knowledge work."
Locking in a programming tool company with $60 billion shows how much Musk values the AI programming field.
Step 2: Directly feed Cursor data into the model.
Public code can train basic skills, but Cursor's private programming interaction data is the core fuel for creating programming "top students".
Step 3: The Grok Build programming Agent is officially launched.
On May 14th, xAI released Grok Build —
A terminal-level AI programming Agent that runs on the command line and supports code generation, file editing, dependency management, and shell command execution.
The biggest selling point is that it supports up to 8 sub-Agents working in parallel.
The pricing is $300 per month for the SuperGrok Heavy subscription, with a promotional price of $99 for the first six months.
It's worth mentioning that Grok Build is natively compatible with the CLAUDE.md configuration file of Claude Code.
Yes, the tool developed by xAI natively supports the configuration format of its competitor, Claude.
Claude Code has become a de facto standard among developers, and xAI is willing to actively be compatible with the competitor's ecosystem to reduce the migration cost.
It's practical, but also a bit sad.
So far, all three steps have been completed — acquisition for distribution, data for ability, and Agent for entry.
All point to the same goal: AI programming.
How far is Grok from the programming throne?
Putting Grok in the coordinate system of the programming field, the gap is obvious.
SWE-bench Verified is the most concerned AI programming benchmark test in the developer community at present.
However, compared with OpenAI and Anthropic, Grok lags far behind —
- GPT-5.5: 88.7%.
- Claude Opus 4.6: 80.8%.
- Grok 4 series: 72% - 75%.
Even worse is the enterprise adoption rate.
Data from Enterprise Technology Research shows that as of March 2026, OpenAI accounts for 55% of enterprise users, Anthropic has soared from 20% a year ago to 47%, and Google accounts for 39%. While Grok only has a meager 6%.
Tripling the parameters and adding Cursor data, V9-Medium may indeed bring about a qualitative change.
However, there has never been a linear relationship between parameter scale and ability. The quality of training data and the accuracy of RLHF are often more critical.
Some analysts speculate that xAI may have fed a large amount of programming data in the later stage of pre - training, similar to the heavy mid - training strategy of Cursor Composer 2.5, so the post - training stage can indeed be compressed.
Showdown in June, a head - on confrontation among four
Musk chose to reveal his cards at this time, and the timing is very delicate.
SpaceX will be listed on the NASDAQ on June 12th, with a target valuation of $1.75 trillion. If successful, it will be the largest IPO in history. The $60 billion acquisition of Cursor is expected to be completed within 30 days after the IPO.
The public release of V9-Medium happens to be right before the IPO.
What's even more terrifying is that Musk is not the only one accelerating.
OpenAI's GPT - 5.6 has been leaked in the Codex background, codenamed iris - alpha, and the 1.5 million Token context window has been tested successfully. Polymarket predicts that the probability of its release before the end of June is over 85%.
Anthropic's Claude Opus 4.8 has appeared in the Google Vertex background . Sonnet 4.8 skipped 4.7 and went straight to 4.8. Mythos 1, once called "too dangerous to be made public", also briefly appeared in the Claude interface.
Google's Gemini 3.5 Pro is also scheduled for June.
Four leading laboratories will have a head - on confrontation in the same month. This June is destined to be the most brutal battle in the AI field in 2026.
The gap between leading laboratories has now shrunk to a weekly basis. A single release is no longer important; the rhythm is the key.
Musk is betting on this rhythm.
V9-Medium doesn't need to become the king overnight. It just needs to prove that xAI is still in the game and has more than one card in hand.
As for the Colossus 1 rented to Anthropic for $1.25 billion per month — while helping the competitor train the model, it's also helping itself survive.
In this AI arms race, survival is more important than victory.
And the prerequisite for survival is that someone is willing to pay for your computing power. Even if that person happens to be your competitor.
Reference materials:
https://x.com/testingcatalog/status/2058800215960719861
https://x.com/elonmusk/status/2058787384364265734
This article is from the WeChat official account "New Intelligence Yuan", author: ASI Revelation; editor: Moses, published by 36Kr with authorization.