Grok 4.5 in Cursor: the agentic loop holds, the token-speed claim doesn't
Grok 4.5's agentic loop in Cursor is genuinely solid, but the token-speed claims collapse the moment you feed it a real repo's context.
I wired Grok 4.5 into both Cursor and Grok Build over a weekend against a real 40k-line TypeScript monorepo. Verdict up front: the agentic loop is genuinely solid — multi-file edits, tool calls, and self-correction hold together better than I expected. But the "faster tokens" marketing falls apart the moment you feed it a real repo with real context.
SpaceXAI (yes, xAI rebranded — Elon being Elon) shipped Grok 4.5 on July 8 as its first model tuned specifically for coding and agentic work rather than a general capability bump. That framing is accurate. This is a developer model, not a chatbot that also codes.
Does the agentic loop actually work?
Yes, mostly. In Cursor's agent mode, Grok 4.5 planned a refactor across six files, ran the test suite via the tool interface, read the failures, and patched its own mistake without me nudging it. That's the loop working end to end.
Where it holds up: it doesn't hallucinate imports as aggressively as I'm used to, and it actually reads the error output instead of guessing. When a Vitest run failed on a mock, it went back and fixed the mock rather than deleting the assertion — which is the lazy move cheaper models make.
Where it wobbles: on the third or fourth tool-call round it started losing track of earlier decisions. It re-introduced a type it had already removed. Context management under long agent chains is still the weak point, same as everyone else.
What broke?
The token-speed claim. On short prompts, sure, tokens fly. But under a real repo — 15k+ tokens of context from Cursor's codebase indexing — throughput dropped hard. First-token latency in Grok Build was consistently 4-6 seconds on large-context requests, and the stream stuttered mid-generation more than once.
That matters because agentic work is exactly the case where context is always large. You're never running an agent on a two-line prompt. So the speed numbers in the original launch coverage describe a workload nobody doing agentic coding actually runs.
The other gotcha: Grok Build's console and Cursor gave me slightly different outputs on identical prompts. I suspect different system prompts or context truncation between the two surfaces. If you're evaluating, pick one surface and stick to it, or your benchmarks are noise.
Should you use it?
If you already live in Cursor and want to A/B a coding model against Claude or GPT, Grok 4.5 is worth a real trial — the loop quality is there. If you're chasing the speed pitch, skip it; under real context it's middle of the pack, not fast.
I'm keeping it in rotation for now but not switching my default. It's a credible coding model with an overstated latency story. Solid engineering, honest-adjacent marketing.