SpaceXAI has launched Grok 4.5, a model trained alongside Cursor that targets coding, agentic workflows, and knowledge work. The company describes it as its smartest model to date.
Performance and benchmarks
SpaceXAI published scores across four coding benchmarks. Fable (max) leads on all four. Grok 4.5 performs closest to the top score on Terminal Bench 2.1.
Quick reference: “pass@1” counts only first-attempt passes; “resolve rate” is the share of tasks fixed.
| Benchmark (harness) | Grok 4.5 | Top listed | Others |
|---|---|---|---|
| DeepSWE 1.0 — pass@1 (each provider’s harness) | 62.0% | Fable (max) 66.1% | GPT 5.5 (xhigh) 64.31%; Opus 4.8 (max) 55.75% |
| DeepSWE 1.1 (mini-swe-agent harness, DataCurve) | 53% | Fable (max) 70% | GPT 5.5 (xhigh) 67%; Opus 4.8 (max) 59%; GLM 5.2 44% |
| Terminal Bench 2.1 | 83.3% | Fable (max) 84.3% | GPT 5.5 (xhigh) 83.4%; Opus 4.8 (max) 78.9% |
| SWE Bench Pro — resolve rate | 64.7% | Fable (max) 80.4% | Opus 4.8 (max) 69.2%; GLM 5.2 62.1%; GPT 5.5 (xhigh) 58.6% |
On SWE Bench Pro, Grok 4.5 resolved tasks with 15,954 output tokens on average. Opus 4.8 (max) used 67,020 on the same benchmark. This is about 4.2× fewer output tokens.
Training and speed
Training ran across tens of thousands of NVIDIA GB300 GPUs. SpaceXAI used training and stability techniques designed for large-scale runs. Beyond raw token volume, the team invested in data filtering and curation. This included deduplication, quality scoring, and domain-focused selection.
SpaceXAI team then scaled reinforcement learning with a focus on per-token intelligence. RL covered hundreds of thousands of tasks. Most centered on multi-step software engineering and other technical work. Grading combined automated and model-based methods. The stack supports highly asynchronous training. Agentic rollouts can run for many hours while learning continues.
Grok 4.5 is served at 80 TPS. SpaceXAI reports roughly twice the token efficiency of leading models. Fewer output tokens usually means lower output cost and latency per task.
Pricing
Grok 4.5 costs $2 per million input tokens and $6 per million output tokens. SpaceXAI says it solves tasks in under half the number of steps. Confirm current pricing in the SpaceXAI console before budgeting.
Use cases
- Codebase repair: find a bug, fix it, then explain the root cause.
- App prototyping: build a Three.js solar-system simulation from one prompt.
- Legal agent tasks: Grok 4.5 ranks #1 on Harvey’s Legal Agent Benchmark.
- Spreadsheet work: build multi-sheet Excel models that pull in web research.
- Documentation: turn an outline into slides and a Word report.
Getting started
Grok 4.5 is available in Grok Build, in Cursor on all plans, and from the SpaceXAI console. Grab an API key and call the responses endpoint. The model ID is grok-4.5.
curl -s https://api.x.ai/v1/responses \
-H "Authorization: Bearer $XAI_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "grok-4.5",
"input": "Find and fix the bug, then explain it: function median(a){a.sort();return a[a.length/2]}"
}'To use Grok Build from the terminal, install the CLI:
curl -fsSL https://x.ai/cli/install.sh | bashAvailability and limits
Grok 4.5 is live in Grok Build and in Cursor on all plans. It is also available via the SpaceXAI console. It is not yet available in the EU. SpaceXAI expects EU availability in mid-July. Free usage is offered for a limited time in Grok Build and Cursor.
What it means
Developers using Cursor will see Grok 4.5 as the default model. The focus is on reducing the number of tokens required to solve a problem. This lowers cost and speeds up the time to get a working fix. The model is also positioned for legal work and spreadsheet modelling tasks.
