SpaceXAI联手Cursor首款大模型今日亮相对标GPT-5.5与Opus 4.8

The first large model jointly trained by SpaceXAI and Cursor is reportedly launching today, with some capabilities able to go toe-to-toe with GPT-5.5 and Claude Opus 4.8. Although the $60 billion acquisition deal has not yet been finalized, the technical teams from both companies have already started collaborating.

Before the $60 billion acquisition deal has even completed its process, the model has already arrived.

According to sources cited today by The Information, the first large model jointly developed by SpaceXAI and Cursor could officially debut as early as today (local time, July 8). The positioning of this model is highly aggressive — some of its capabilities are being directly benchmarked against Anthropic’s Claude Opus 4.8 and OpenAI’s GPT-5.5.

This is a signal: over the past six months, Cursor — long regarded as a major “wrapper app” player — is making the leap from the application layer to the foundational model layer. And the force pushing it forward is Musk’s Colossus 2 data center, which has already expanded to nearly one million H100-equivalent GPUs.

An Acquisition That Hasn’t Closed Yet, but the Model Is Already Running

First, let’s sort out the timeline. At the end of last month, Musk confirmed on X that SpaceX had proposed a $60 billion stock-based acquisition offer for Cursor, and that the offer had been accepted. Converted to RMB, that’s roughly 408.1 billion yuan — a record figure in the AI coding tools sector, instantly making Cursor one of the most valuable assets in the SpaceX empire outside of Starlink.

What’s interesting is that the deal terms explicitly stated it was “conditional.” Cursor had previously publicly stated that it had no plans, for the time being, to jointly develop coding models with xAI. Clearly, though, that statement didn’t last more than a few weeks.

Judging from today’s leaked information, engineering teams from both companies had already begun technical integration during the acquisition negotiation phase. In a recent internal communication, Cursor CEO Michael Truell mentioned that the company was “training from scratch at the Colossus data center a model capable of directly competing with Anthropic and OpenAI.” That statement now has a concrete product behind it.

This kind of “the contract isn’t finalized, but the code is already merged” operation is uncommon in Silicon Valley, and it also reflects Musk’s intensity in pushing this project forward.

Technical Foundation: 1.5T Parameters + Starting with 100,000 GPUs

The full technical picture of the new model is still incomplete, but several key figures have already surfaced:

Base model: Derived from xAI’s V9 series, with 1.5 trillion parameters, previously used to train Grok 4.5
Compute: Training used the Colossus cluster, starting at 100,000 H100 GPUs, with the next generation expected to scale to nearly one million GPUs
Data contribution: Cursor contributed user interaction data accumulated from coding scenarios, code modification traces, and a large volume of long-context repository-level samples

If Grok 4.5 is viewed as the “preview version” of this collaboration — using Cursor’s data to supplement-train the V9 base model — then the model debuting today can be seen as the full production release: not patching an existing model, but retraining from scratch according to the needs of coding scenarios.

There’s an interesting technical point worth considering here. GPT-5.5 and Claude Opus 4.8 excel at coding tasks because they combine general reasoning ability with code knowledge. Cursor’s killer advantage is that it possesses the world’s richest “AI-assisted programming behavior data” — what suggestions users accepted, what they rejected, where they manually rewrote code, how long the cursor lingered, and more. This data is gold for training a model that truly understands developer intent.

In other words, OpenAI and Anthropic rely on general-purpose capability to tackle coding, while Cursor + xAI are attempting to tackle coding through specialized data plus hyperscale compute. The endpoints of these two paths may not be the same.

Benchmarking GPT-5.5 and Opus 4.8 — Based on What?

The wording “comparable in some performance areas” is actually quite subtle. Translated plainly: overall it still loses, but it can compete on several specific benchmarks.

Take Cursor’s Composer 2.5, released last month — the precursor to its self-developed model effort. It achieved 69.3% on Terminal Bench 2.0 and 79.8% on SWE-Bench Multilingual, while also performing strongly on Cursor’s own more difficult internal evaluation set. Composer 2.5 used a relatively modest compute budget yet already achieved top-tier response speed and accuracy.

So what about this new model? According to leaked information, its focus is in three areas:

Information processing speed: This may be the first noticeable difference. Cursor has consistently emphasized “achieving the same quality reasoning with lower latency,” and the new model reportedly shows significant improvements in long-context throughput.
Repository-level understanding: Not just understanding individual files, but performing cross-file refactoring in projects containing hundreds of thousands of lines of code.
Agent capabilities: The ability to autonomously execute multi-step tasks, evolving from “code completion” to “fulfilling requirements.”

If these three claims hold up, then it has the credentials to directly compete with GPT-5.5’s Coding mode and Opus 4.8’s Extended Thinking. Of course, the benchmarks will ultimately decide.

The Real Significance of This

My assessment is that the key point of this news is not how strong the model itself is, but two broader developments:

First, the AI coding tools war has entered the “self-developed model” phase. Over the past two years, tools like Cursor, Windsurf, and Cline competed over product experience and prompt engineering. Now Cursor has moved furthest ahead: it no longer wants to merely be a “wrapper,” but intends to own the underlying model layer as well. This is bad news for Anthropic — Claude has consistently been the most-used model among Cursor users. If users switch to Cursor’s in-house model, Anthropic loses a major distribution channel.

Second, Musk’s AI ecosystem has formed a closed loop for the first time. xAI has foundational models and compute (Colossus), Cursor has distribution and data, X has social scenarios and real-time corpora, and Tesla has physical-world data. Acquiring Cursor gives Musk, for the first time, a true “developer entry point,” which may ultimately prove far more valuable than Grok serving as a chatbot inside X.

What will the developer community’s attitude be? Most likely: wait and see. Cursor users are extremely sensitive to stability, and any model switch triggers complaints. If the new model launches with poor latency, hallucination issues, or refusal rates, users may still switch back to Claude.

API Usage

If you plan to test the new SpaceXAI model as soon as its API officially becomes available, OpenAI Hub will integrate support shortly after launch. At that point, a single API key will allow you to call it alongside GPT-5.5, Opus 4.8, and the Gemini series for comparative testing, without registering separate accounts or going through multiple vendors’ KYC processes.

Here’s an example call after integration goes live (model name subject to official release):

from openai import OpenAI

client = OpenAI(
    api_key="your-openai-hub-key",
    base_url="https://api.openai-hub.com/v1"
)

response = client.chat.completions.create(
    model="cursor-composer-x1",  # Official model name may differ after release
    messages=[
        {"role": "system", "content": "You are a senior engineer."},
        {"role": "user", "content": "Implement a concurrent LRU cache in Rust with O(1) read/write complexity."}
    ],
    temperature=0.2,
    max_tokens=4096
)

print(response.choices[0].message.content)

For comparative testing, simply replace model with gpt-5.5, claude-opus-4.8-20260601, or gemini-3.0-pro to run the same prompt and directly compare output differences.

Several Questions That Still Need Verification

After the new model debuts today, several issues deserve close attention:

Context length: Composer 2.5 supports 200K. If the new model doesn’t reach beyond 1M, it will struggle to create meaningful separation in repository-level tasks.
Inference cost: 1.5 trillion parameters implies expensive inference. Will Cursor dare to provide it free within subscription plans?
Open-source possibility: Musk previously open-sourced parts of the Grok series. Will that continue this time?
Relationship with Grok 5: xAI is also training Grok 5. Are these two models different distilled versions of the same system, or entirely independent branches?

The answers will gradually emerge over the next few weeks. What is certain is that the coding model race became more crowded starting today — OpenAI, Anthropic, Google, and xAI+Cursor are now all seated at the table, and none of them will win this round easily.