On July 16, SpaceXAI released Grok 4.5, and its positioning fits in one sentence: their smartest model to date, built specifically for coding, agentic tasks, and knowledge work. But the interesting part is not any single benchmark score — it is the overall shape. Grok 4.5 tops some boards and clearly trails on others, while pairing that record with a claimed 2x token efficiency and one-quarter the output volume. The pitch quietly moves the battlefield from “which model is strongest” to “which model is cheapest per unit of intelligence.”
Positioning: a Cursor-co-trained engineering workhorse
The capability range is concrete: engineering work in Rust and C/C++ on one end, building a complete application from a single prompt on the other. This is not a generalist flagship; it is a model converging on execution. Training happened alongside Cursor — that phrase matters, because the scoring contexts come from how a real development tool is actually used, not from purely academic tests. It also explains why on launch day the model landed in every Cursor plan at once instead of sitting in an API console waiting for users to wire it up. For teams already running coding agents heavily, there is one more consequence worth spelling out: model and harness are evaluated as a pair. Swapping the model means re-testing how your harness behaves with it, not just checking API compatibility.
The full benchmark card: mixed wins, not a clean sweep
The official scorecard has to be read as a whole. The wins are clear: 29.0% on SWE Marathon pass@1, first place ahead of Opus 4.8’s 26.0% and Fable’s 24.0%; 83.3% on Terminal Bench 2.1, one point behind Fable’s 84.3% and 0.1 behind GPT 5.5’s 83.4%. DeepSWE tells a different story. On 1.0, Grok 4.5’s 62.0% trails Fable(max) at 66.1% and GPT 5.5(xhigh) at 64.31%, ahead of Opus 4.8(max) at 55.75% and Opus 4.7(max) at 40.12%. On 1.1, which swaps in the mini-swe-agent harness, it drops to 53% — behind Fable at 70%, GPT 5.5 at 67%, and Opus 4.8 at 59%, beating only GLM 5.2’s 44%. Its 64.7% on SWE Bench Pro also sits mid-pack, behind Fable’s 80.4% and Opus 4.8’s 69.2%.
| Benchmark | Grok 4.5 | Best rival on that board | Grok position |
|---|---|---|---|
| SWE Marathon pass@1 | 29.0% | Opus 4.8 26.0% | First |
| Terminal Bench 2.1 | 83.3% | Fable 84.3% | Third, 1.0 and 0.1 points behind the top two |
| SWE Bench Pro | 64.7% | Fable 80.4% | Upper middle |
| DeepSWE 1.0 | 62.0% | Fable(max) 66.1% | Behind the top two |
| DeepSWE 1.1 | 53% | Fable 70% | Clearly behind |
The point of the table is not the ranking but the mixture itself. The two DeepSWE versions deserve a side-by-side look: the same model goes from 62.0% on 1.0 to 53% on 1.1, and the only thing that changed is the harness (1.1 swaps in mini-swe-agent) — a nine-point swing, larger than the gap between most of the models listed. Change the harness, the repositories, or the scoring method, and relative positions move; that is a statement about the whole agentic benchmark ecosystem, not just this release. The data’s provenance matters too — every number comes from SpaceXAI’s own announcement, with competitor figures pulled from their respective system cards and leaderboards, not rerun under identical conditions by one team. Reasonable for shortlisting candidates; reckless as a purchasing ranking. The comparison that actually settles anything is your repository, your test suite, and your time limits.
Training: hundreds of thousands of tasks and hours-long rollouts
The training investment is not small: tens of thousands of NVIDIA GB300 GPUs, by the company’s account. RL covered hundreds of thousands of tasks, concentrated on multi-step software engineering, scored with a mix of automated metrics and model-based grading, with rollouts running for hours at a stretch. That detail matters more than it sounds: an agent’s unit of work is not one generation but a full loop of editing, testing, and fixing. A training environment has to permit trajectories that long for the model to learn discipline across steps. Read against the benchmark spread, this suggests another lens: on tasks that resemble the training shape — terminal operations, marathon-style long repairs — it wins; under more constrained harnesses with leaner tooling, it loses. How closely your workload matches the training shape is the best available predictor of where it will land for you.
The efficiency economics: 80 TPS, a quarter of the tokens, $2/$6
Here is where the real story seems to live. The fast-model variant runs at roughly 80 TPS, which the company calls faster than flash models. The token math is more direct: on SWE Bench Pro, Grok 4.5 averaged 15,954 output tokens — about a quarter (1/4.2) of Opus 4.8(max)’s 67,020. The announcement further claims roughly 2x overall token efficiency and less than half the task steps. Note that those two sentences are vendor claims, not third-party findings.
Add pricing and the picture completes: $2/M input and $6/M output. If token consumption is a quarter of a flagship rival’s while speed sits at fast-model level, then even without topping every board, intelligence per unit cost may be the actual product. For teams running high volumes of agent tasks, total cost per completed task was always closer to the purchase decision than any single leaderboard crown; Grok 4.5 simply productizes that angle. But the ledger cuts both ways: if the claimed 2x efficiency and halved steps hold, what you save is tokens and round-trips. If your workload happens to fall into the categories it loses — long multi-step edits under a harness like DeepSWE 1.1’s — the extra retries will eat the efficiency advantage straight back. The economics only work if your tasks sit in the categories it wins.
How to get it: default, every plan, free for now
Grok 4.5 is already the default model in Grok Build, with plugins for Excel, PowerPoint, Word, and Outlook — a reminder that its stated scope is broader than code: knowledge work and office documents are in the official script too, though the evidence in this launch sits firmly on the coding side. As of launch day it is available in Grok Build, every Cursor plan, and the SpaceXAI console; usage is free for a limited time in Grok Build and Cursor, while the console API bills at list price — take “limited time” literally. It is a launch-window promotion, not the standing price. Budget against the $2/$6 list price, because the bill that arrives when the free window closes will not include a grace period.
Sources
AI-assisted summary compiled from the sources above, reviewed by a human before publishing.