Breaking

xAI Debuts Grok 4.5, a Model Tuned for Coding

July 9, 2026 at 13:02 EDT

Foundation Models
Software Dev & Coding
AI Agents

SpaceXAI (xAI) on July 8, 2026 released Grok 4.5, a new flagship model tuned for coding, agentic tasks and knowledge work. Co-trained with Cursor to sharpen real-world performance, it is immediately available through Grok Build, all Cursor plans, and the company's console/API, with EU availability slated for mid-July.

July 8, 2026 · xAI

Grok 4.5 arrives: "Opus-class, but faster, cheaper and leaner"

xAI's new flagship, co-trained with Cursor for coding and agentic work, ships across Grok Build, all Cursor plans and the API — pitched to undercut Claude Opus and GPT 5.5 on both performance and cost.

500k

token context window

80 TPS

inference speed, flash-class

$2 / $6

per 1M input / output tokens

Leaner on tokens where it counts

On SWE-Bench Pro, Grok 4.5 completes tasks using roughly a quarter to half the output tokens of rivals — the cost edge, drawn to scale.

~4 units

rival output tokens

~1 unit

Grok 4.5

Down to as little as a quarter of the tokens for the same task.

How it was built

Tens of thousands of GB300 GPUs → RL on 100Ks of SWE tasks → Cursor co-training

Large-scale data curation plus reinforcement learning across multi-step software engineering, fine-tuned on real-world coding.

Developers say +

Beat Claude & GLM building a business site
Durable in long agentic loops
Fast enough for real work
Office plugins for Excel, PowerPoint, Word

Open questions −

Weekly usage quotas burn fast
Value vs higher-tier plans unproven
EU availability only mid-July

Continue reading

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