grok 4.5 is trained on cursor session data — and it shows in the benchmarks

xAI launched Grok 4.5 on July 8–9, 2026 — first available inside Cursor, then publicly on the xAI API on July 9. it's the first model built jointly by SpaceXAI and Cursor, and the training story is more interesting than most model releases: Grok 4.5 was trained on trillions of tokens from real Cursor developer interaction data — actual agentic coding sessions, not synthetic benchmarks.

the base architecture is V9, a 1.5-trillion-parameter Mixture-of-Experts foundation. the coding-specific layer on top is what makes this different from Grok 4: xAI built it to work across real repositories, handle changes that touch multiple files, and operate inside agents that plan and call tools — rather than just answer questions about code.

the benchmark picture

on Terminal-Bench 2.1 — a benchmark that measures whether a model can complete real engineering tasks in a terminal environment — Grok 4.5 posts 83.3%. Fable 5 is at 84.3%; GPT-5.5 is at 83.4%. the gap to frontier is less than one point, at a price well below both.

on SWE-Bench Pro (real-world software engineering tasks on open-source codebases), Grok 4.5 posts 64.7%. Opus 4.8 is at 69.2%; Fable 5 at 80.3%. on SWE-Bench Multilingual, which tests across multiple programming languages, it posts 78.0% against Opus 4.8's 84.4%. DeepSWE 1.0: 62.0% (GPT-5.5: 64.3%).

one number is notably absent: CursorBench. xAI excluded it from the published results because of accidental training data contamination — the model was trained on data that included CursorBench examples, making any score on it unreliable. it's worth noting the exclusion because this kind of disclosure is rarer than it should be in model releases, and because CursorBench is otherwise the most relevant benchmark for this specific tool's core audience.

pricing

$2 per million input tokens and $6 per million output tokens at standard speed. a fast tier is available at $4/$18. compare that to Claude Opus 4.8 at $15/$75, or Claude Fable 5 at $10/$50 — Grok 4.5 is roughly 75% cheaper per token than Opus 4.8 while posting comparable terminal-bench numbers.

for teams running large-volume agentic coding workflows — CI bots, overnight refactors, PR review agents — the cost difference is significant. the tradeoff is on the harder SWE benchmarks where Fable 5 and Opus 4.8 still lead.

where it's available

Grok 4.5 is available in Cursor desktop, web, iOS, CLI, and SDK on all plans. this isn't an add-on or a paid-tier feature — it rolled out universally when SpaceX completed its acquisition of Cursor. it's also available via the xAI public API for anyone building outside the Cursor environment.

the data flywheel question

training on Cursor session data is a meaningful strategic advantage if it compounds: each new developer interaction theoretically makes the next model better at the specific things developers actually do. this is the same loop that made GitHub Copilot hard to displace in the early days — not benchmarks, but product-market fit through data.

the honest counterpoint: the model was good enough to land within 1% of frontier on Terminal-Bench, but still trails on SWE-Bench Pro by 15+ points versus Fable 5. terminal task completion and real-world codebase surgery are measuring different things. the coding session data is clearly helping with the former; the latter is a harder problem.