Six weeks after SpaceX closed its $60 billion acquisition of Cursor, the empire shipped its first joint model. Grok 4.5, built by SpaceXAI (the entity formerly known as xAI) and trained on what Cursor describes as “trillions of tokens” of real developer interaction data pulled from its own product, went live July 8 across Cursor, Grok Build, and the SpaceXAI API console. The pitch is aggressive: Opus-class capability, half the price, and — for the first time — a model explicitly trained to go beyond software engineering into legal, finance, and general STEM knowledge work.
It’s a genuinely interesting release. It’s also a release where the company shipping it quietly admitted, in its own announcement, that some of the numbers backing the “beats Opus 4.8” headline shouldn’t be fully trusted. That combination is worth unpacking, because it’s a preview of how benchmark marketing is going to work now that the model vendor and the IDE vendor are the same company.
The Pitch: Cheap, Fast, and Trained on Your Own Codebase#
The base terms are straightforward. Grok 4.5 runs at $2 per million input tokens and $6 per million output tokens, with a faster variant at $4/$18. Compare that to Claude Opus 4.8’s $5/$25: on paper, Grok 4.5 is 60% cheaper on input and 76% cheaper on output. It serves at roughly 80 tokens per second. Availability is immediate and broad — desktop, web, iOS, CLI, SDK, plus OpenRouter and Vercel AI Gateway — with Cursor doubling included usage allocation for the first week to drive adoption.
The more distinctive claim is the training data. Cursor says Grok 4.5 was built on reinforcement learning against “difficult problems in realistic environments,” with a distributed agent system constructing training environments that Cursor says would otherwise have taken “teams of hundreds of engineers months to build” — and, critically, on interaction data captured from actual Cursor users working in actual codebases. That’s a structural advantage no pure model lab has: Anthropic, OpenAI, and Google all train on public code, synthetic tasks, and licensed data, but none of them sit inside millions of developers’ live editing sessions the way an IDE vendor does. If that data genuinely transfers to capability, it’s a real moat — and a real reason the Cursor-xAI marriage was always going to produce something more threatening than either company alone.
The Benchmarks: A Split Decision, Not a Sweep#
Here’s where the “beats Opus 4.8” framing gets complicated. Grok 4.5’s benchmark spread, as disclosed by Cursor and corroborated by independent outlets including MarkTechPost and OfficeChai, is genuinely mixed:
- Terminal-Bench 2.1: Grok 4.5 scores 83.3%, ahead of Opus 4.8’s 78.9% by roughly 4.4 points, and essentially tied with GPT-5.5’s 83.4%.
- DeepSWE 1.0: Grok 4.5 hits 62.0%, beating Opus 4.8’s 55.8% but trailing GPT-5.5’s 64.3%.
- SWE-bench Multilingual: Grok 4.5 lags at 78.0%, well behind Opus 4.8’s 84.4%.
- SWE-bench Pro — the benchmark this blog has repeatedly flagged as the one that actually separates frontier agentic coding models from everyone else — Grok 4.5 scores 64.7%. Opus 4.8 sits at 69.2%. Claude Fable 5 sits at 80.3%, a full 15.6 points ahead.
So the real picture is: Grok 4.5 wins on shorter, terminal-centric agentic tasks and loses — sometimes by a wide margin — on the harder, longer-horizon SWE-bench Pro measure that correlates most closely with real production engineering work. Musk’s own marketing chart, comparing Grok 4.5 favorably to Opus 4.8 “on several benchmarks,” is technically accurate and substantively misleading in the way that phrase always is: it’s true on the metrics chosen, and silent on the one where the gap is largest.
The Part Cursor Buried: Contaminated Training Data#
The more consequential disclosure is smaller and further down the announcement. Cursor acknowledged that “an earlier snapshot of the Cursor codebase was accidentally included in Grok 4.5’s training data” — meaning the model may have been trained, at least partially, on the very kind of internal test material that benchmarks like Cursor’s own internal coding evals are meant to be blind to. Cursor says the contaminated data was “removed prospectively,” which addresses the problem going forward but does nothing to un-train the model that’s shipping today. Cursor also disclosed that its SWE-bench Pro and Terminal-Bench figures for third-party models — including the GPT-5.5 numbers used in its own comparison chart — are self-reported and self-tested rather than pulled from OpenAI’s or Anthropic’s published results.
Put those two admissions together and the takeaway isn’t “Grok 4.5 is bad.” It’s that the benchmark chart accompanying the launch is not an independent measurement — it’s Cursor grading its own model against Cursor’s own internal reproduction of competitors’ scores, using training data that, by Cursor’s own admission, wasn’t fully clean of the codebase it was being benchmarked against. That’s not a minor asterisk. It’s the exact failure mode this blog has flagged before with GPT-5.6 Sol’s METR eval-gaming: a self-reported “the model performs well” signal that can’t be taken at face value without independent verification.
Why This Matters Beyond One Model#
This launch is the clearest signal yet of what the Cursor-SpaceX combination is actually for. Cursor was always structurally limited by being an IDE wrapper around other labs’ models — Anthropic’s, OpenAI’s, Google’s — with no control over the thing actually doing the reasoning. Owning a model lab fixes that, and owning the world’s highest-volume developer-interaction dataset gives that model lab a training signal no competitor can replicate without also owning an IDE with comparable market share. That’s a genuine structural advantage, and it’s worth taking seriously as a long-term competitive threat to Anthropic’s API and Claude Code businesses, not dismissing as a rebrand exercise.
But it also means the incentive to publish favorable, self-graded benchmarks just got stronger, not weaker — the same company now controls the product surface, the model, and the numbers used to market both. Cursor deserves credit for disclosing the contamination issue at all; plenty of vendors wouldn’t have. But “we told you the data might be dirty, in paragraph nine” is not the same as an independently verified benchmark, and teams choosing a model for agentic coding work should weight SWE-bench Pro — where Grok 4.5 trails both Opus 4.8 and Fable 5 by a wide margin — over the terminal-task numbers Cursor led its announcement with. Cheap and fast is real. “Opus-class” is not yet demonstrated.
Sources: