Table of Contents
Grok 4.5 in Cursor: benchmarks, pricing, and which model to run in July 2026
Grok 4.5's pitch is price, not peak capability. On the benchmark run xAI conducted in its own harness, the model edges Claude Opus 4.8; on the neutral third-party harness it trails both GPT-5.5 and Fable 5, per kingy.ai's benchmark writeup. That spread, plus a base price of $2 input / $6 output per million tokens, per Cursor, decides who should switch.
If you are a solo developer on Cursor Pro: set Grok 4.5 base for long-running agent loops and terminal work. It posts 83.3% on Terminal Bench 2.1 and takes first place on SWE Marathon at 29%, per kingy.ai's benchmark writeup and onemetrik's summary. Keep a Claude model pinned for code review, where the cleaner DeepSWE 1.1 test shows Grok at 53% against Fable's 70% and GPT-5.5's 67%.
If you run a team on Teams Standard: Cursor doubled included usage for the first week, per Cursor's Grok 4.5 announcement. Treat that week as a paid pilot. Route your actual sprint work through Grok 4.5 and read the credit burn before you commit a model policy.
If you are in the EU: Grok 4.5 is not available to you yet as of July 10, 2026, per onemetrik. Cursor's picker will not surface it on EU-geolocated accounts, and Auto routes to other models. Skip to the EU section for what to run in the meantime.
What Grok 4.5 is, and why it shows up in Cursor now
SpaceX agreed to buy Cursor for $60 billion in an all-stock deal announced June 16, 2026, with regulatory close expected in Q3, per TechCrunch. The deal is meant to close the gap between SpaceX's xAI division and the frontier labs. Grok 4.5 is the first model shipped under that arrangement, and it arrived inside Cursor on July 8, 2026, per The Next Web.
"Trained with Cursor" is doing a lot of work in the marketing copy, so here is the mechanism. Grok 4.5 was trained on Cursor interaction data, not just fine-tuned on it after the fact. Elon Musk framed the result as an "Opus-class model, but faster, more token-efficient and lower cost," per The Next Web, which also notes the model does not beat competitors' largest systems across the board.
The training story carries one caveat you need before reading any leaderboard. Cursor states plainly in its own post that "Grok 4.5 has an advantage on CursorBench because an earlier snapshot of the Cursor codebase was accidentally included in training," per Cursor. A model that saw the benchmark's underlying code during training will score high on that benchmark for reasons that have nothing to do with generalization. Discount the CursorBench figure entirely. It is contaminated, and Cursor says so.
The benchmark grid: harness choice changes the story
Four coding benchmarks tell most of the story, and the pattern in them is the whole point. The numbers below are the neutral third-party figures compiled by kingy.ai, cross-checked against onemetrik for SWE Marathon, both fetched 2026-07-10. Fable refers to Fable 5.
| Benchmark | Grok 4.5 | Claude Opus 4.8 | GPT-5.5 | Fable 5 |
|---|---|---|---|---|
| DeepSWE 1.0 (provider harness) | 62.0% | 55.75% | 64.31% | 66.1% |
| DeepSWE 1.1 (neutral harness) | 53% | 59% | 67% | 70% |
| SWE Marathon | 29% | 26% | n/a | 24% |
| Terminal Bench 2.1 | 83.3% | 78.9% | 83.4% | 84.3% |
Read the two DeepSWE rows together. DeepSWE 1.0 runs inside each provider's own agent harness, which lets a vendor tune the scaffolding around its model; DeepSWE 1.1 uses a shared mini-swe-agent harness run by DataCurve, per kingy.ai, so it is the cleaner cross-model comparison. Grok 4.5 sits third on the provider-run test and last of four on the neutral one. That drop from 62.0% to 53% is the signal: some of Grok 4.5's headline strength lives in the harness, not the model. On the two tasks that reward long-horizon agent behavior rather than single-shot patch accuracy, Terminal Bench 2.1 and SWE Marathon, Grok 4.5 clears Opus 4.8 and stays within a point of the leaders. The practical reading: Grok 4.5 is a strong agent-loop and terminal model and a middling code-review model, and no benchmark here justifies making it your only model.
Pricing: where Grok 4.5 base lands in the picker
Price is where Grok 4.5 has an unambiguous edge. Base pricing is $2 input / $6 output per million tokens, and the fast variant is $4 input / $18 output, per Cursor. Claude Opus 4.8 runs $5 input / $25 output per million tokens, per The Next Web. Grok 4.5 base undercuts Opus 4.8 by 60% on input and 76% on output. For scale, The Next Web lists GPT-5.6 Luna at $1 input / $6 output, so Grok 4.5 is cheap for an Opus-class model but not the cheapest token in the picker.
| Model | Input ($/Mtok) | Output ($/Mtok) | Draws from included pool? |
|---|---|---|---|
| Grok 4.5 (base) | $2 | $6 | Yes |
| Grok 4.5 (fast) | $4 | $18 | Yes, at the higher rate |
| Claude Sonnet 4.5 | see Cursor picker | see Cursor picker | Yes |
| Claude Opus 4.8 | $5 | $25 | Yes |
| GPT-5.5 | see Cursor picker | see Cursor picker | Yes |
Two cells stay blank on purpose. We pulled the current Cursor and third-party pricing pages during this writeup and could not source exact per-token rates for Claude Sonnet 4.5 or GPT-5.5 inside Cursor, so those read "see Cursor picker" rather than a guessed number. Check the live rate in the model dropdown before you route heavy traffic to either.
The mechanics of the pool matter as much as the sticker price. Individual and team plans include what Cursor calls "significant usage" of Grok 4.5 as part of the first-party model pool, per Cursor. Auto mode selects a model for you and spends from that included pool; picking a model by hand spends from the same pool at the rates above. Once you exhaust included usage, you pay overage at these per-token rates. That is why a cheaper base rate stretches a Pro plan further only if your workload actually routes to Grok 4.5, which brings us to the switch decision.
When to switch off Auto
Auto is the right default for most sessions because it hides model routing and spends from the included pool. Override it when one of three conditions holds.
Switch to Grok 4.5 base when you are running multi-step agent loops that live in the terminal. The 83.3% Terminal Bench 2.1 score and the first-place SWE Marathon result are the tasks Grok 4.5 is built for. In the Cursor CLI, pin the model on the run:
# Run an agent task pinned to Grok 4.5 instead of letting Auto choose
cursor-agent -m grok-4.5 "refactor the retry logic in src/queue.ts and run the tests"Override to Grok 4.5 base when you are burning through Pro's included pool and want to stretch it. At $2/$6 base, per Cursor, Grok 4.5 costs less per token than the Claude options, so routing high-volume, lower-stakes generation to it leaves more of the pool for the tasks that need a stronger model. You can watch the burn in the CLI before and after a change:
# Check remaining included usage after a batch of Grok 4.5 runs
cursor-agent usage --planFor long-horizon document work, such as finance or legal analysis inside Cursor, Grok 4.5 base is also worth a pin. SpaceXAI positioned Grok 4.5 for those domains, and the SWE Marathon result is the closest coding proxy for that long-context, many-step behavior.
Stay on Auto, or pin a Claude model, when the job is code review or a large refactor where correctness on hard tasks decides the outcome. The DeepSWE 1.1 gap is the reason. To route review to Sonnet 4.5 while keeping Grok 4.5 for generation, select the model per chat in the picker (Cmd/Ctrl + / opens it) or set it explicitly in a review command:
# Pin a Claude model for the review pass; the DeepSWE 1.1 gap matters here
cursor-agent -m claude-sonnet-4.5 "review the diff on this branch for correctness and edge cases"EU users: what to run until availability lands
Grok 4.5 is not available in the EU as of July 10, 2026, per onemetrik. Cursor's model picker will not list it on an EU-geolocated account, and Auto will route to the other first-party models, primarily the Claude and GPT lines already in the pool. We could not source a dated, official Cursor statement naming the exact EU rollout date during this writeup, so treat the timeline as unconfirmed and watch the Cursor changelog for the availability flip.
Until then the decision is simpler for EU accounts because the low-cost Opus-class option, Grok 4.5 base at $2/$6 per Cursor, is off the menu. Run Auto for general work and pin Claude Opus 4.8 for the hard review and refactor passes where its 59% DeepSWE 1.1 beats Grok 4.5's 53% anyway. The one thing not to do is architect a workflow around Grok 4.5 that you cannot actually run yet.
Data and privacy: read the terms before you standardize on it
Cursor confirms training included Cursor interaction data, and the CursorBench contamination it disclosed was an accidental codebase snapshot in that training set, per Cursor. The company frames future training runs as excluding those accidental snapshots. If your code cannot be part of any model-training corpus, that distinction is the one to verify against your plan's data terms rather than the marketing summary. Enterprise plans carry separate data-handling commitments; review the Cursor Enterprise data policy before you make Grok 4.5 the default model on repositories with sensitive or regulated code.
The verdict, by who you are
Solo dev on Cursor Pro: run Grok 4.5 base as your agent-loop and terminal model, and pin a Claude model for review. You get a low-cost Opus-class token at $2/$6, per Cursor, where it is strongest and keep the harder review work on a model that clears the neutral benchmark.
Team on Teams Standard: use the doubled first-week usage as a paid pilot. Route a real sprint through Grok 4.5, measure the credit burn and the review-rework rate, then write your model policy from your own numbers rather than the launch benchmarks.
EU user: you cannot run Grok 4.5 yet. Stay on Auto, pin Opus 4.8 for hard tasks, and watch the Cursor changelog for the EU flip before you plan around it.
All three paths need a paid plan, because Grok 4.5 access and the included usage pool are gated behind them. If you are still on the free Hobby tier and want the model in your picker, upgrade to a paid Cursor plan and start on the doubled first-week usage while it lasts.