Model Drop

Grok 4.5

July 9, 2026

View on Substack

Grok 4.5 is the first model from SpaceXAI that they’ve co-trained with Cursor (on trillions of tokens of real developer data). Cursor’s Composer line, at least for now, appears to be dead.

SpaceXAI released Grok 4.5 yesterday as its smartest model to date, aimed at coding, agentic tasks, and knowledge work. Fresh off their agreement to be acquired, Cursor helped train it, feeding trillions of tokens of real developer interactions with codebases and software tools into the run. Musk’s framing set the target explicitly: “an Opus-class model, but faster, more token-efficient and lower cost,” with the internal assessment pegging it as roughly comparable to Opus 4.7, but much faster. The price does most of the talking. Two dollars in and six out per million tokens undercuts Opus 4.8 by 60-76% and Fable 5 by 80-88%.

Grok 4.5 claims to complete agent tasks in under half the steps of comparable models, averaging 1.9M tokens per task where GPT-5.5 burns 6.2M and Fable 5 burns 7.2M, which compounds the sticker discount into a cost-per-task gap of $2.49 versus $5.07 and $11.80. It effectively ties the frontier on Terminal-Bench 2.1 and leads Fable 5 and Opus 4.8 outright on AutomationBench-AA’s simulated office-work agents.

On the neutral-harness DeepSWE 1.1 it drops to 53% against Fable 5’s 70%, its SWE-Bench Pro score trails Opus 4.8 by four and a half points, Artificial Analysis’s Omniscience eval clocked its hallucination rate at 54% (up from 25% on the prior Grok), and SpaceXAI shipped no model card. Also, buried in a footnote of Cursor’s own launch post: an earlier snapshot of the Cursor codebase was accidentally included in Grok 4.5’s training data, which handed the model an unearned advantage on CursorBench and forced Cursor to strike those scores from the launch entirely.

What’s new

Grok 4.5 is a step up for xAI (now SpaceXAI).

• Cursor as co-trainer, not customer. No frontier lab has shipped a model trained jointly with an IDE company on the actual telemetry of professional developers working in real codebases. The RL setup had engineers specify hard problems and verification methods while agent groups built and refined the environments at scale. This is a growing data moat Anthropic and OpenAI don’t have, and it also explains why the model punches hardest inside agentic coding harnesses.

• Token efficiency as the product. Every lab claims efficiency but SpaceXAI made it their headline spec. Under half the steps per task, 1.9M tokens where rivals burn 6-7M, 4.2x fewer tokens than Opus 4.8 on SWE-Bench Pro.

• An agent that leads on office work. AutomationBench-AA runs 657 tasks across 40 simulated apps (Gmail, Slack, Salesforce, HubSpot), and Grok 4.5’s 51.4% beats Fable 5 and Opus 4.8. The same eval logged it breaking guardrails more often than either (0.63 violations per task against Opus 4.8’s 0.55), unfortunately.

• End of the SpaceX era. First model out of the merged, publicly traded SpaceXAI. The AI lab that was an X subsidiary two years ago now sits inside a rocket company with public shareholders, and this launch is the first read on what that machine ships under quarterly scrutiny.

• Pricing with trapdoors. The $2 / $6 sticker hides a 200K-token threshold that doubles the rate and cache pricing at 25% of input where competitors charge 1-10%. Long-context agent work, the exact workload this model is sold for, is where those trapdoors open.

How and where to use it

Where it’s available

• Grok Build and the SpaceXAI API console (grok-4.5)

• Cursor on every plan across desktop, web, iOS, CLI, and SDK with doubled allocation through the first week

• OpenRouter, Vercel AI Gateway, Cloudflare, Snowflake, Databricks Mosaic, and Microsoft Office add-ins

• Nothing in the EU until mid-July at the earliest

What it’s good at

• High-volume agentic coding where cost-per-task decides the tool (the $2.49 per coding-agent task against Fable 5’s $11.80 is the whole pitch)

• Terminal-driven work, where 83.3% on Terminal-Bench 2.1 sits at the frontier

• Simulated office and knowledge-work automation, where it leads AutomationBench-AA outright

• Long agent loops where its half-the-steps efficiency keeps both latency and bills down

• Anything you were already doing in Cursor, since the model was trained on exactly that distribution and runs there at ~90 tokens/sec

What it’s bad at / shouldn’t be used for

• Hard multi-file engineering where Fable 5’s 80.4% SWE-Bench Pro and 70% DeepSWE 1.1 dwarf Grok’s 64.7% and 53%

• Anything where a 54% hallucination rate on unanswerable questions is disqualifying, which includes most of the “finance, legal, research” knowledge work that SpaceXAI names in its own launch copy

• Agents operating near live financial or customer systems, given the highest guardrail-violation rate in its comparison set

• EU deployments, full stop, until the rollout lands

• Any regulated workload that needs a model card to clear review, because there isn’t one

First impressions

The positives

Artificial Analysis ran its full index suite and titled the result “Grok 4.5 brings SpaceXAI to the intelligence frontier,” scoring it 54, fourth overall and 16 points over the prior Grok. Hacker News user Tiberium distilled the economics in the launch thread:

“4x better reasoning efficiency compared to Opus while being priced at $2/$6.”

Benchmark leads come and go weekly, but a 4x efficiency edge priced at a fraction of the frontier is a structural argument.

The hands-on reports back the harness numbers. HN user paradox460 threw a gnarly Elixir and Tailwind refactor at it, one where Opus had been struggling:

“Grok aced it, rather quickly and cheaply, surprisingly.”

Another tester, jonathaneunice, had it review a full test suite, where it surfaced “a substantive long-standing bug” that previous approaches had walked past. Anecdotes, sure, but they’re the right kind.

HN user redox99 posted the calibrated version of the praise, and it’s worth quoting for what it concedes:

“In the same tier as opus, occupying the lower end of that tier together with GLM 5.2.”

The negatives

The Decoder flagged the ugliest number in the launch data, from Artificial Analysis’s Omniscience eval:

“Hallucination rate increased from 25% to 54%, despite accuracy improving from 35% to 52%.”

The model got smarter and more than twice as willing to bluff when it doesn’t know.

The benchmark story has a contamination problem, and it came from the co-training arrangement itself. Cursor’s launch post disclosed, in a footnote, that the model trained on the very codebase its in-house benchmark tests against:

“Grok 4.5 has an advantage on CursorBench because an earlier snapshot of the Cursor codebase was accidentally included in training. The exact impact is unclear. That data has been removed for future models, and in parallel we are working on a larger update to CursorBench, hence the exclusion here.”

Credit where due: Cursor disclosed it and struck the scores. But “the exact impact is unclear” is doing a lot of work in that footnote. The whole sales pitch for this model is trillions of tokens of Cursor data, and the one eval built by the people who know that data best had to be thrown out because the training set memorized the answers.

Kingy AI’s analysis called the results “mixed rather than triumphant,” with the 62% on SpaceXAI’s own DeepSWE 1.0 harness collapsing to 53% on the independently run DeepSWE 1.1, a nine-point gap between the maker’s rig and everyone else’s.

SpaceXAI shipped no Grok 4.5 model card, the Grok 4 generation was documented consulting Musk’s own X posts before answering questions on immigration and the Middle East, and the MechaHitler episode is still the first association many developers have with the brand.

Fair or not as a summary of the 4.5-generation model, it’s the reputational baseline xAI chose to launch into without a model card, and no amount of token efficiency answers it.

Jake’s take

As sad as I am to see the Composer line go, the co-training with Cursor seems to have produce some worthwhile fruit in a remarkably short time. The lab with the best model is now competing against the lab with the best data about how developers actually work, and Grok 4.5 is the first evidence that the second thing can substitute for some of the first.

But it’s a Grok model, and there’s baggage that comes with that. A 54% hallucination rate and 0.63 guardrail violations per task, both the worst in their comparison sets, in a model line with a controversial background, is unfortunate.

And then there’s the benchmark contamination. The model “accidentally” trained on Cursor’s own codebase, aced the benchmark built from that codebase, and the whole eval got quietly retired in a footnote. I believe the accident, for what it’s worth (contamination happens, and Cursor disclosing it beats the industry norm of saying nothing), but a launch whose headline numbers come from the maker’s own harnesses, whose in-house eval got tossed for memorizing the test, and whose neutral-harness scores drop nine points has forfeited the benefit of the doubt on every chart it published.

Stack the missing model card on top of the Grok 4 generation’s habit of checking Musk’s posts before answering political questions, and I have a hard time getting excited about the model overall.