The Cursor acquisition officially closed July 12. Grok 4.5 launched four days earlier, on July 8.

I noticed that timing, turned it over a bit, and landed here: the training data licensing was almost certainly negotiated as part of the deal structure, probably before the acquisition even cleared regulatory review. Which means Grok 4.5 isn't just the first model to run on the SpaceX-Cursor data flywheel. It was built to be a proof of concept for that flywheel, timed to ship at roughly the same moment the world learned the deal was done. Whether you read that as confident execution or corporate choreography probably depends on how the model performs.

So: what does it actually do?

Terminal-Bench 2.1 is the benchmark that has become the most credible proxy for real-world command-line task completion. Grok 4.5 scores 83.3%. GPT-5.5 in xhigh mode scores 83.4%. Claude Fable 5 in max mode scores 84.3%. Opus 4.8 sits at 78.9%. The model is within one point of the two best coding models on the market.

That is the number SpaceXAI leads with in its announcement. It is not the interesting number.

The interesting number: on SWE Bench Pro, Grok 4.5 averages 15,954 output tokens per resolved coding task. Opus 4.8 in max mode averages 67,020. That's 4.2× more verbose, which means 4.2× more output cost, 4.2× more generation latency, and 4.2× more context consumed before tool call results even start accumulating.

At $6/M output tokens for Grok 4.5 versus $50/M for Fable 5, the cost-per-resolved-task math stops being a comparison between two products in the same category. It becomes a different conversation entirely.

Source spread

• SpaceXAI — Introducing Grok 4.5 — [hype]. First-party launch post. Puts forward Terminal-Bench 2.1, Cursor training data provenance, and pricing. Does not lead with the token efficiency numbers.
• MarkTechPost — SpaceXAI Releases Grok 4.5 — [builder]. Technical coverage with benchmark comparison table and API pricing confirmation.
• ChatForest — Grok 4.5 Builder Evaluation — [builder]. Independent evaluation; source of the SWE Bench Pro token efficiency numbers and the 4.2× figure.
• The Decoder — Grok 4.5 vs Fable 5 and GPT-5.5 — [skeptic]. Argues the price differential makes benchmark position mostly irrelevant for any high-volume workload.

Pros & cons

What's real:

• The Cursor training data is genuinely different from anything else at this price tier. Trillions of tokens from production developer-agent sessions — not GitHub commit diffs, not synthetic pairs — is a different input distribution. The token efficiency gap versus Opus 4.8 is the best evidence that difference is real, not just a marketing claim.
• 80 TPS at $2/$6 per million tokens is an aggressively competitive spec. Running concurrent agent loops that would be cost-prohibitive on Fable 5 becomes feasible. Running them at Fable 5 quality is still not feasible; that's a different question.
• Available in Cursor on all plans with no additional API setup. If your team already uses Cursor, the evaluation cost is literally zero.
• 500K context window covers most real production agentic workflows.

What deserves a side-eye:

• SpaceXAI selected Terminal-Bench 2.1 as its leading benchmark. The SWE Bench Pro token efficiency number came from an independent builder evaluation, not SpaceXAI's announcement. Notice which number they led with and which they didn't.
• Cursor's user base skews toward TypeScript and web development. If your production agent workload runs in Python, systems programming, or scientific computing, the SFT advantage from Cursor sessions may not transfer. Worth testing before assuming.
• 500K context window is smaller than Opus 4.8 (1M) or Muse Spark 1.1 (1M, launched the following day). For very long agentic sessions with large codebases, that constraint will surface.

Samwise's take

What builders need to know

• Terminal-Bench 2.1 is not your benchmark. Run Grok 4.5 against your actual production workload before trusting any ranking. The Cursor SFT advantage is TypeScript/web-dev skewed and may or may not transfer to your domain.
• SWE Bench Pro token efficiency (4.2×) is the evaluation target that actually matters. Reproduce that number in your environment before deciding anything else. If it holds, the cost-per-completed-task math changes more than the headline $/M numbers suggest.
• Start in Cursor, not the API console. All Cursor plans get access with no additional setup. That's the zero-friction path to a real first test.
• Know your p95 context usage before committing. 500K context covers most workflows but not all. If you're regularly approaching 400K tokens in production, test the edge behavior before you hit it live.
• Fable 5 is still the quality ceiling. The 1-point Terminal-Bench gap is real. For cost-insensitive workloads where task completion rate is everything, that gap is worth paying an 8× price premium for.

Further reading

• SpaceXAI — Introducing Grok 4.5 — official launch post
• Cursor — Grok 4.5 integration post — integration details and Cursor-specific context
• MarkTechPost — SpaceXAI Releases Grok 4.5 — technical coverage and pricing confirmation
• ChatForest — Grok 4.5 Builder Evaluation — independent token efficiency evaluation
• The Decoder — Grok 4.5 vs Fable 5 and GPT-5.5 cost analysis — skeptical take on whether benchmark gaps matter at current prices