SpaceXAI just released Grok 4.5. The company calls it its smartest model to date. It targets coding, agentic tasks, and knowledge work. SpaceXAI says Grok 4.5 was trained alongside Cursor, an AI coding editor.
TL;DR
- Grok 4.5 targets coding, agentic tasks, and knowledge work, trained alongside Cursor.
- In SpaceXAI’s own chart, Fable (max) leads all four benchmarks; Grok 4.5 is closest on Terminal Bench 2.1.
- Token efficiency is the standout: about 4.2× fewer output tokens than Opus 4.8 (max) on SWE Bench Pro.
- Pricing is $2/M input and $6/M output, served at 80 TPS.
- It ranks #1 on Harvey’s Legal Agent Benchmark and is the default model in Grok Build.
What is Grok 4.5?
Grok 4.5 is a general-purpose model tuned for real engineering work. SpaceXAI trained it on datasets spanning coding, science, engineering, and math. The research team describes its reasoning as both intelligent and efficient. It scored #1 on Harvey’s Legal Agent Benchmark, which SpaceXAI cites as office-work strength.
How SpaceXAI trained Grok 4.5?
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.
Benchmark performance
SpaceXAI team published scores across four coding benchmarks. Competitor figures come from published system cards or leaderboards. SpaceXAI’s prose says Grok 4.5 exceeds comparable leading models. Its own chart is more mixed. Fable (max) posts the top score on all four benchmarks. Grok 4.5 stays closest 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% |
Speed and token efficiency
Grok 4.5 is served at 80 TPS. SpaceXAI reports roughly twice the token efficiency of leading models. On SWE Bench Pro, Grok 4.5 resolved tasks with 15,954 output tokens on average. SpaceXAI reports Opus 4.8 (max) used 67,020 on the same benchmark. That is about 4.2× fewer output tokens. 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 with examples
- 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 (working code)
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 | bash
Grok 4.5 at a glance
| Attribute | Detail |
|---|---|
| Vendor | SpaceXAI |
| Focus | Coding, agentic tasks, knowledge work |
| Training partner | Cursor |
| Hardware | Tens of thousands of NVIDIA GB300 GPUs |
| Serving speed | 80 TPS |
| Token efficiency | ~4.2× fewer output tokens than Opus 4.8 (max) on SWE Bench Pro |
| Input price | $2 / million tokens |
| Output price | $6 / million tokens |
| Access | Grok Build, Cursor (all plans), SpaceXAI console |
| Model ID | grok-4.5 |
Availability 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.
Interactive Explorer
Check out the Technical details here. Also, feel free to follow us on Twitter and don’t forget to join our 150k+ML SubReddit and Subscribe to our Newsletter. Wait! are you on telegram? now you can join us on telegram as well.
Need to partner with us for promoting your GitHub Repo OR Hugging Face Page OR Product Release OR Webinar etc.? Connect with us
Michal Sutter is a data science professional with a Master of Science in Data Science from the University of Padova. With a solid foundation in statistical analysis, machine learning, and data engineering, Michal excels at transforming complex datasets into actionable insights.
Credit: Source link
