|

SpaceXAI Releases Grok 4.5, a Cursor-Trained Model for Coding, Agentic Tasks, and Knowledge Work at $2/M Input

ByRicardo

SpaceXAI simply launched Grok 4.5. The firm calls it its smartest mannequin to this point. It targets coding, agentic duties, and information work. SpaceXAI says Grok 4.5 was educated alongside Cursor, an AI coding editor.

TL;DR

  • Grok 4.5 targets coding, agentic duties, and information work, educated alongside Cursor.
  • In SpaceXAI’s personal chart, Fable (max) leads all 4 benchmarks; Grok 4.5 is closest on Terminal Bench 2.1.
  • Token effectivity is the standout: about 4.2× fewer output tokens than Opus 4.8 (max) on SWE Bench Pro.
  • Pricing is $2/M enter and $6/M output, served at 80 TPS.
  • It ranks #1 on Harvey’s Legal Agent Benchmark and is the default mannequin in Grok Build.

What is Grok 4.5?

Grok 4.5 is a general-purpose mannequin tuned for actual engineering work. SpaceXAI educated it on datasets spanning coding, science, engineering, and math. The analysis staff describes its reasoning as each clever and environment friendly. It scored #1 on Harvey’s Legal Agent Benchmark, which SpaceXAI cites as office-work power.

How SpaceXAI educated Grok 4.5?

Training ran throughout tens of 1000’s of NVIDIA GB300 GPUs. SpaceXAI used coaching and stability strategies designed for large-scale runs. Beyond uncooked token quantity, the staff invested in information filtering and curation. This included deduplication, high quality scoring, and domain-focused choice.

SpaceXAI staff then scaled reinforcement studying with a give attention to per-token intelligence. RL lined lots of of 1000’s of duties. Most centered on multi-step software program engineering and different technical work. Grading mixed automated and model-based strategies. The stack helps extremely asynchronous coaching. Agentic rollouts can run for many hours whereas studying continues.

Benchmark efficiency

SpaceXAI staff printed scores throughout 4 coding benchmarks. Competitor figures come from printed system playing cards or leaderboards. SpaceXAI’s prose says Grok 4.5 exceeds comparable main fashions. Its personal chart is extra combined. Fable (max) posts the highest rating on all 4 benchmarks. Grok 4.5 stays closest on Terminal Bench 2.1.

Quick reference: “cross@1” counts solely first-attempt passes; “resolve fee” is the share of duties mounted.

Benchmark (harness) Grok 4.5 Top listed Others
DeepSWE 1.0 — cross@1 (every supplier’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 fee 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 effectivity

Grok 4.5 is served at 80 TPS. SpaceXAI stories roughly twice the token effectivity of main fashions. On SWE Bench Pro, Grok 4.5 resolved duties with 15,954 output tokens on common. SpaceXAI stories Opus 4.8 (max) used 67,020 on the identical benchmark. That is about 4.2× fewer output tokens. Fewer output tokens often means decrease output value and latency per process.

Pricing

Grok 4.5 prices $2 per million enter tokens and $6 per million output tokens. SpaceXAI says it solves duties in below half the variety of steps. Confirm present pricing within the SpaceXAI console earlier than budgeting.

Use instances with examples

  • Codebase restore: discover a bug, repair it, then clarify the basis trigger.
  • App prototyping: construct a Three.js solar-system simulation from one immediate.
  • Legal agent duties: Grok 4.5 ranks #1 on Harvey’s Legal Agent Benchmark.
  • Spreadsheet work: construct multi-sheet Excel fashions that pull in internet analysis.
  • Documentation: flip a top level view into slides and a Word report.

Getting began (working code)

Grok 4.5 is accessible in Grok Build, in Cursor on all plans, and from the SpaceXAI console. Grab an API key and name the responses endpoint. The mannequin ID is grok-4.5.

Copy Code
curl -s https://api.x.ai/v1/responses 
  -H "Authorization: Bearer $XAI_API_KEY" 
  -H "Content-Type: software/json" 
  -d '{
    "mannequin": "grok-4.5",
    "enter": "Find and repair the bug, then clarify it: perform median(a){a.kind();return a[a.length/2]}"
  }'

To use Grok Build from the terminal, set up the CLI:

Copy Code
curl -fsSL https://x.ai/cli/set up.sh | bash

Grok 4.5 at a look

Attribute Detail
Vendor SpaceXAI
Focus Coding, agentic duties, information work
Training companion Cursor
Hardware Tens of 1000’s of NVIDIA GB300 GPUs
Serving velocity 80 TPS
Token effectivity ~4.2× fewer output tokens than Opus 4.8 (max) on SWE Bench Pro
Input worth $2 / million tokens
Output worth $6 / million tokens
Access Grok Build, Cursor (all plans), SpaceXAI console
Model ID grok-4.5

Availability and limits

Grok 4.5 is reside in Grok Build and in Cursor on all plans. It can be out there through the SpaceXAI console. It shouldn’t be but out there within the EU. SpaceXAI expects EU availability in mid-July. Free utilization is obtainable for a restricted time in Grok Build and Cursor.

Interactive Explorer