When a company itself removes its own benchmark from a presentation — it's either honesty or a sign that the results are saying the wrong thing. In the case of Grok 4.5, it turned out to be both at once.
What is this model
Grok 4.5 — the first model under the SpaceXAI brand following xAI's rebranding, developed jointly with code editor Cursor. In terms of architecture — a mixture-of-experts based on the V9 foundation with 1.5 trillion parameters. For comparison: the previous version Grok 4.3 used approximately 0.5 trillion parameters.
The model is available in Cursor on all plans, in Grok Build and via SpaceXAI API at $2 per million input tokens. Exception — the EU: the model is not yet available there; access is promised by mid-July.
The benchmark trick
Cursor immediately announced in its announcement that it excluded its own CursorBench benchmark from the comparative results. The reason — an early snapshot of Cursor's code base accidentally ended up in the training dataset, giving the model an unfair advantage specifically on this test.
"Grok 4.5 has an advantage on CursorBench because an early snapshot of Cursor's code base accidentally made it into the training data. The exact impact is unclear. These data have been removed for future models."
Cursor, official announcement
Independent benchmarks paint a contradictory picture. On DeepSWE 1.0 (infrastructure provider benchmark), Grok 4.5 outperforms Claude Opus 4.8. But on the neutral DeepSWE 1.1 — it noticeably falls behind both Opus 4.8 and Fable 5. Such gaps between "proprietary" and independent measurements are typically the main argument of skeptics.
What's actually new
- Broader specialization. The previous Composer 2.5 was trained solely on code. Grok 4.5 received a wider mix: STEM tasks, scientific articles, financial and legal documents — the model is positioned as an agent for "knowledge work," not just a programmer.
- Reduced context window. The window dropped from 1 million tokens in Grok 4.3 to 500,000 in Grok 4.5. xAI provided no public explanation. For teams analyzing large code bases or documents, this is a real limitation.
- Token efficiency. On practical tasks, Grok 4.5 consumes fewer output tokens than competitors — this directly impacts the cost of agent sessions.
- Reinforcement learning in realistic environments. The model was trained to recover from errors and verify its own results — not just generate code, but also verify it.
Cursor data as asset and risk
The training dataset includes trillions of tokens from Cursor — developer interactions with code bases, patterns of agent work, bug fixes in real projects. This gives the model practical "development experience" that models trained only on static code don't have.
But this raises the question of dependency. The more Cursor data in the model, the tighter the connection between the editor and the provider. Teams choosing Grok 4.5 because of integration with Cursor are essentially betting that both products will remain compatible — and that SpaceXAI won't change API terms.
If the next iteration (rumored to be 2T V-series, expected in August) shows the same gap between provider and neutral benchmarks, the question of trust in evaluation methodology will become central — and not just for Grok.
