Breaking

xAI Debuts Coding-Focused Grok 4.5, Hitting Elo 1328 for 5th on Website Arena

July 8, 2026 at 14:06 EDT

Foundation Models
Software Dev & Coding
AI Agents

xAI on July 8, 2026 unveiled Grok 4.5, described as its first model built specifically for coding and agents. Trained on Cursor-derived data, it logged Elo 1328 and a 5th-place ranking on the crowdsourced Website Arena benchmark, a 25-rank jump over the company's previous best model.

July 8, 2026 · xAI

Grok 4.5 pivots to code — and claims "Opus-level" web building

xAI's first model built for coding and agents jumped 25 ranks on Website Arena to 5th place, trained on Cursor-derived data — while running faster and far cheaper on tokens.

↑25

rank jump on Website Arena over prior best model

1328

Elo, landing 5th place on Website Arena

80 tok/s

inference speed, ~2× token efficiency

$2/$6

per M input / output tokens (xAI API)

Token efficiency · SWE Bench Pro

~4.2× fewer output tokens to resolve a task

15,954

Grok 4.5
avg output tokens

67,020

Opus 4.8
max output tokens

Software engineering benchmarks

Competing with leading models (%)

DeepSWE 1.0

Grok 4.5 · 62.0

Opus 4.8 · 55.75

GPT-5.5 · 64.31

Terminal Bench 2.1

Grok 4.5 · 83.3

Opus 4.8 · 78.9

GPT-5.5 · 83.4

SWE Bench Pro (resolve rate)

Grok 4.5 · 64.7

Opus 4.8 · 69.2

Developers point to

Cursor-based training synergy
Speed and token-cost efficiency
One-prompt end-to-end app builds
1st on Harvey's Legal Agent Benchmark

Open questions

Many figures provider-supplied
Limited independent public benchmarks
Past Grok trailed Claude on UI polish
EU rollout delayed to mid-July 2026

Continue reading

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