Grok 4.5 Ships: Cursor-Trained, Token-Cheap, and One Rung Short

For a month, SpaceXAI's coding strategy was an argument about data. In June, Elon Musk's lab told the world it had finished training a 1.5-trillion-parameter foundation model on Cursor's developer workflows, and had struck a $60-billion right-to-acquire deal for the editor's parent company to lock that pipeline in. The thesis was blunt: you cannot out-research your way to a coding lead, but you might be able to buy the data that produces one. On July 8, the argument stopped being theoretical. Grok 4.5 shipped.

The model is exactly what the June disclosures promised. Grok 4.5 is a mixture-of-experts model built on the 1.5-trillion-parameter V9 foundation, trained jointly with Cursor on trillions of tokens of real editor interaction data. It's live in Grok Build, inside Cursor on every plan, and from the SpaceXAI console. And it arrives priced to move: $2 per million input tokens and $6 per million output tokens, running at roughly 80 tokens per second. For a frontier-tier coding model in mid-2026, that is aggressive.

Third place on its home turf

Here is the part SpaceXAI's launch materials handle carefully. On Terminal-Bench 2.1 — the agentic-coding benchmark that has become the industry's shorthand for "can this model actually operate a terminal and finish real work" — Grok 4.5 scored 83.3%. That is a genuinely strong number. It is also, by SpaceXAI's own reported figures, third.

Anthropic's Claude Fable (max) leads at 84.3%. GPT-5.5 (xhigh) sits at 83.4%. Grok 4.5 trails both by a hair, then opens a real gap on the field: Opus 4.8 (max) lands at 78.9%. So the model that was purpose-built on the world's richest corpus of coding behavior — the entire strategic justification for a $60-billion transaction — came within a percentage point of the top, and still didn't reach it.

Whether that's a disappointment depends entirely on what you thought the bet was for. If the goal was to win the coding benchmark outright, third is a miss, and it's a pointed one given the setup. If the goal was to close a gap that looked structural six months ago — when Claude's coding lead read as a moat rather than a margin — then a hair behind the two most expensive labs on the planet is a vindication. SpaceXAI didn't have a credible coding model in this weight class a year ago. Now it has one that trades blows at the top of the table.

The number that actually matters

The benchmark headline undersells the more interesting result, which is efficiency. On SWE-Bench Pro, SpaceXAI reports Grok 4.5 using roughly 4.2× fewer output tokens than Opus 4.8 (max) to reach comparable results. Read that slowly, because it reframes the whole comparison.

Coding agents don't bill by benchmark score. They bill by tokens. An agent that grinds through a multi-file refactor, reads logs, retries failed commands, and narrates its reasoning can burn output tokens by the hundreds of thousands in a single session — and output tokens are where the money is. A model that reaches the same finish line with a quarter of the token spend isn't just cheaper per million; it's cheaper per task, compounded across every step of an agentic loop. At $6 per million output tokens and a 4× efficiency edge, Grok 4.5's effective cost-to-complete undercuts the frontier by a margin that a benchmark leaderboard simply doesn't capture.

This is the quiet shift in how coding models compete in 2026. The frontier has flattened — the top four models are separated by low single digits on the benchmarks that matter, and no one is running away with it. When capability converges, the competition moves to cost, latency, and token economy. Grok 4.5's pitch isn't "we're the smartest." It's "we're within noise of the smartest, we run at 80 tokens a second, and we'll finish the job for a fraction of the token bill." For a developer paying real money to keep an agent running, that's often the more persuasive argument.

The Cursor loop, closing

The strategic logic also looks sharper now that the model is real. Grok 4.5 doesn't just train on Cursor data — it ships inside Cursor, on every plan, which means every session run through the editor is both a product surface and a fresh data-collection channel for the next model. That's a flywheel other labs can't easily copy without owning a top-tier editor of their own. The $60-billion deal bought SpaceXAI a coding-data pipeline that renews itself with usage, and the July release is the first turn of that wheel with a competitive model attached.

There's an obvious risk in the arrangement, and it's the one that made observers uneasy in June: a model tuned on one editor's telemetry, distributed through that same editor, can start to optimize for the benchmark of "feels good in Cursor" rather than general coding ability. The Terminal-Bench and SWE-Bench numbers are reassuring precisely because they're independent of the training surface — Grok 4.5 performs on tasks that aren't Cursor sessions. But the incentive to overfit the home turf will only grow as the flywheel spins, and it's worth watching whether future Grok releases keep clearing external benchmarks or quietly start winning only where SpaceXAI controls the arena.

For now, the verdict is narrower and more interesting than either the hype or the skepticism predicted. The data bet worked well enough to produce a real frontier-class coding model, priced and tuned to compete on economics rather than raw capability. It didn't take the crown. It didn't need to. In a market where the top four are within a point of each other, the model that finishes the job cheapest may be the one that actually gets deployed — and Grok 4.5 was built, quite deliberately, to be that model.