Breaking

xAI Releases Coding-Focused Grok 4.5, Trained With Cursor Data

July 8, 2026 at 14:09 EDT

Foundation Models
Software Dev & Coding
AI Agents

xAI officially released Grok 4.5 on July 8, its latest model built for coding, agentic tasks and knowledge work. The company calls it its most capable model, saying it rivals Anthropic's Claude Opus class while running faster and at lower cost.

July 8 · xAI

Grok 4.5 Ships — Built for Coding, Priced to Undercut Claude Opus

xAI calls Grok 4.5 its most capable model yet, claiming it rivals Anthropic's Claude Opus class while running faster and at lower cost. Battle-tested first inside SpaceX and Tesla.

1.5T

parameters (V9 foundation model)

~80 TPS

inference speed (tokens/sec)

4.2×

fewer output tokens vs Opus 4.8

Output token use on SWE-Bench Pro

Average output tokens per task — lower is cheaper & faster

15,954

Grok 4.5

67,020

Claude Opus 4.8

Same task, ~4.2× leaner — the efficiency story behind the cost claim.

API INPUT

$2 / M tokens

API OUTPUT

$6 / M tokens

EU AVAILABILITY

Delayed until mid-July

How it was trained

Pre-train

NVIDIA GB300 GPU cluster

→

Refine

Cursor developer workflow data

→

RL + Agentic

Reinforcement learning & Grok Build rollouts

What developers praise

Handles large codebases & long engineering tasks
End-to-end app building from one prompt
Strong on Rust/C++ and Excel/PowerPoint/Word work

What critics caution

Claims rest largely on internal evaluations
Limited independent benchmarking so far
Still being tuned as RL improvement continues

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.