Breaking

xAI Releases Grok 4.5 and Enters the Agent Arena

July 8, 2026 at 15:09 EDT

Foundation Models
AI Agents
Software Dev & Coding

On July 8, 2026, xAI unveiled Grok 4.5, a new model focused on coding, agentic tasks and knowledge work. The same day it was added to the "Agent Arena" on evaluation platform Arena.ai, where it will be measured on real-world, long-horizon tasks.

July 8, 2026 · xAI

Grok 4.5 bets on efficiency to crack the coding-agent race

xAI's first model built for coding, agentic tasks and knowledge work — co-trained with Cursor — enters the tool-using Agent Arena where Anthropic's Claude family has dominated.

4.2×

fewer output tokens than Opus 4.8 on SWE Bench Pro

80 tok/s

generation speed cited as a key differentiator

on Harvey's Legal Agent Benchmark

Output tokens on SWE Bench Pro

Doing the same job with far less text

15,954

Grok 4.5

~67,000

Opus 4.8 (max)

Roughly 4.2× fewer output tokens — cheaper and faster per task.

Benchmarks — Grok 4.5 vs leaders (%)

Competitive on engineering, trailing on the hardest suites

Terminal Bench 2.1

Grok 4.5

83.3

Fable (max)

84.3

DeepSWE 1.0

Grok 4.5

62.0

Fable (max)

66.1

SWE Bench Pro

Grok 4.5

64.7

Fable (max)

80.4

Agent Arena — tool-using leaderboard

Where Grok has to climb

#1Claude Fable 5 (High)

#22Grok 4.3 High

#23Grok Build 0.1

#29Grok 4.3

TBDGrok 4.5 — score being tallied

1.1M+ real user sessions logged across 29 models.

Continue reading

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