xAI Trained Grok 4.5 Alongside Cursor. Your Agent Isn't Cursor.

xAI says Grok 4.5 was trained alongside Cursor, tuned on real coding runs inside one agent's loop and tool format. The benchmark scores it launched with were earned in that setup.

A SWE-bench number measured in the harness a model was co-trained with is a ceiling under ideal conditions, not a promise the behavior carries over to your stack.

Model choice is quietly turning into model-plus-harness choice. If your agent is not the one the model was tuned against, benchmark inside your own loop before you trust the leaderboard.

xAI shipped Grok 4.5 on July 8, and the coverage sorted itself into the usual buckets. Opus-class quality. Two dollars input, six output. Tens of thousands of GB300 GPUs. Elon Musk calling it faster and more token-efficient than the flagship it was chasing. All real, all fine, and all skipping the one line in the announcement that actually changes how you should read the rest of it.

Grok 4.5 was trained alongside Cursor.

The detail that reframes the benchmarks

Not trained and then plugged into Cursor. Trained alongside it, on real coding workflows, inside the editor's agent loop. That is a different thing from a general model that Cursor later adopts. It means the reinforcement learning that shaped Grok 4.5's agentic behavior happened against a specific harness, with a specific tool-call format, a specific loop structure, and a specific idea of what a good run looks like.

When you co-train a model with a harness, you bake that harness into the weights. The model learns which tools exist and how they are called. It learns the shape of the loop: read, edit, run, check, repeat. It learns the system prompt it will see. Most of all it learns the token budget that loop rewards, because reinforcement learning across hundreds of thousands of tasks is, in part, teaching the model to get the result with fewer wasted moves in that environment.

So when the benchmark table says Grok 4.5 resolves 64.7 percent of SWE-bench Pro tasks using about 15,954 output tokens each, against Opus 4.8 at 69.2 percent and roughly 67,020 tokens, the efficiency is not a free-floating property of the model. Some of it is the model having learned to emit exactly what one loop wants and nothing it does not. That is genuinely impressive. It is also, quietly, a number attached to a setup.

The leaderboard got local

For most of the last two years, a model was a model. You could read GPT's SWE-bench score, or Claude's, or GLM's, and treat it as a portable fact. Drop the model behind your own agent, wire up your own tools, and you would land somewhere near the published number, give or take.

Co-training erodes that. The published number is now measured in a particular harness, often the vendor's own, and increasingly one the model was tuned against. It stops being a property of the model and becomes a property of the pair. A score earned inside Cursor tells you what Grok 4.5 does inside Cursor. It does not promise what it does inside your LangGraph setup, your homegrown loop, or your MCP server with a tool schema the model has never seen.

This is not xAI doing something sneaky. It is what everyone is doing. Anthropic tunes for Claude Code. GLM ships tuned for its own ZCode editor. OpenAI's models are shaped by Codex. The model and the harness are being co-designed, because that is how you get the agentic reliability that raw pretraining does not hand you. The side effect is that a benchmark number without a named harness is measuring an artifact of somebody's setup, and that somebody is usually not you.

What it costs you if you are not paying attention

Two things.

The first is that you over-trust the leaderboard. You see Grok 4.5 sitting near the top on token efficiency, you swap it into your own agent, and you get a version that reasons more, calls tools in a format it half-recognizes, and burns more tokens than the table led you to expect. Not because the model is bad, but because you took a paired number and treated it as a solo one. The disappointment is real and avoidable, and it comes from reading the benchmark as a promise instead of a ceiling.

The second is slower and worse: portability is quietly leaving the building. Switching models used to be a config change. One line, a different endpoint, done. As models couple to harnesses, switching starts to mean re-tuning your prompts, adjusting your tool schemas to match what the new model expects, and re-benchmarking on your own tasks to see what actually changed. The lock-in is moving off the API surface, where you could see it, and into the training, where you cannot. A model that is spectacular in its native harness and merely okay in yours is a soft form of vendor gravity, and it does not show up on any pricing page.

What to actually do

Benchmark in your own harness. Not the vendor's, not SWE-bench as published, yours, on a sample of your real tasks, behind the exact agent loop you run in production. That is the only number that predicts your outcome, and it is the one nobody hands you. Two afternoons of setup buys you a decision you can defend.

Treat every published benchmark as a ceiling under ideal conditions. The vendor measured the model in the environment where it looks best, which is fair and also exactly why you cannot lift the number into your context unchanged. Read it as the top of the range, then find out where you actually land.

And flip the whole thing into an advantage if you can. If you have standardized on one harness, Cursor, Claude Code, your own, then a model co-trained with a harness like yours is not a risk, it is a gift. The coupling that makes benchmarks non-portable is the same coupling that makes a well-matched model behave better than its raw scores suggest. The move is to know which harness a model was tuned against and weight it accordingly, instead of pretending the pairing does not exist.

The pattern to remember

The era of the context-free model number is ending. A benchmark score is starting to carry an invisible footnote naming the harness it was earned in, and as co-training spreads, that footnote matters more than the score. Grok 4.5 is a strong model, and it is strongest in the loop it was raised in. Before you trust what it does for you, find out whether your loop is that loop, and if it is not, measure before you migrate.