Grok 4.5 is here: coding champion, but caught cheating
July 9, SpaceXAI released Grok 4.5.
It's the first model under the unified SpaceXAI brand after xAI merged in. 1.5T parameter MoE, $2/$6 pricing, pitched at coding and agent tasks.
The report card looks great: 83.3% on Terminal-Bench 2.1, first place in coding. Except there's a catch — CursorBench scores were inflated because the training data included snapshots of Cursor's own codebase.
The good parts
- 1.5T MoE, built on the V9 foundation
- $2/$6 (input/output per million tokens) — drastically cheaper than GPT-5.6's $5/$30
- Terminal-Bench 2.1: 83.3%, first place
- Trained on trillions of tokens of real Cursor agent interaction data
- Fast inference, roughly 4x the output efficiency of Opus 4.8
Coding is genuinely strong. Terminal-Bench measures a model's ability to complete real development tasks in a terminal — 83% is a serious score.
The problem: CursorBench was tainted
SpaceXAI disclosed it themselves: training data included Cursor's codebase snapshots. That means the model had "seen the answers" during training, inflating CursorBench scores.
How bad is it?
- On the mild side — SpaceXAI disclosed it voluntarily, didn't hide it. Decent move.
- On the serious side — CursorBench is a major coding-model benchmark. Tainted scores mean the ranking isn't trustworthy. Other models took the test blind; Grok 4.5 had the answer key.
So that "coding champion" label needs an asterisk. Terminal-Bench 83.3% is probably legit (that benchmark is harder to taint this way), but take specific CursorBench numbers with a grain of salt.
Why this happens
Grok 4.5's training used massive amounts of real Cursor agent interactions — the actual developer workflows inside Cursor. That data naturally includes Cursor's own codebase.
The issue: CursorBench's test set overlapped with the training data. The model didn't "understand" Cursor's patterns — it memorized the answers.
This isn't unique to Grok. The entire industry's data governance faces this: training sets grow larger and messier, and no one can 100% guarantee benchmarks aren't contaminated. But getting caught is getting caught.
Pricing is worth watching
$2/$6 puts it directly against Anthropic's Sonnet 5 ($2/$10 intro pricing). Output is 40% cheaper than Sonnet 5.
SpaceXAI's play is clear: trade margin for volume. Coding workloads are cost-sensitive (a single agent run can burn hundreds of thousands of tokens), and cheap is a hard advantage.
How to use it
If your workload is coding and agentic tasks, Grok 4.5 is worth trying — Terminal-Bench proves real terminal-task capability is solid, and it's cheap.
But if you're using benchmark rankings to pick a model, strip out CursorBench and look only at Terminal-Bench and real-project scores.
One line: the model may genuinely be strong, but the scores were inflated. Treat these as two separate things.
The real lesson from Grok 4.5 isn't "SpaceXAI cheated" — it's how fragile AI benchmarking is. Training data contamination is an industry-wide disease. Today they caught Grok; how many haven't been caught? Don't buy a model on rankings alone. Run it yourself and see.