Breaking

SpaceXAI's Grok 4.5 Debuts at No. 13 on New Agent Arena Leaderboard After Big Jump in Terminal Tasks

July 13, 2026 at 14:01 EDT

AI Agents
Foundation Models
Software Dev & Coding

SpaceXAI's newly released Grok 4.5 has entered the Agent Arena leaderboard at No. 13, a marked climb from its predecessor's No. 29 ranking, powered by sharp gains in terminal recovery and task completion measured across nearly 10,000 live agentic sessions.

July 8, 2026 · SpaceXAI

Grok 4.5 Vaults to No. 13 on the Agent Arena

SpaceXAI's first model trained specifically for coding and agents climbs 16 places over its predecessor across nearly 10,000 live agentic sessions — a faster, cheaper challenger closing ground on the frontier.

Grok 4.3

No. 29

→

Grok 4.5

No. 13

A 16-place leap on the Agent Arena leaderboard, measured over ~9,800 live agentic sessions.

+9.7%

Bash Recovery gain — now 9th

+6.9%

Confirmed task success — now 6th

+5.0%

Overall improvement vs Grok 4.3

Token efficiency: fewer output tokens on SWE Bench Pro

SpaceXAI claims ~4.2× fewer output tokens than Opus 4.8 on the same tasks — the size of each column IS the token spend.

Opus 4.8

~4.2× tokens

Grok 4.5

1× baseline

$2 / $6

Per M input / output tokens

~80 t/s

Generation speed

64.7%

SWE Bench Pro resolve rate

What builders praise

Speed + cost fit real builder workflows
Holds context, completes complex tasks with minimal supervision
Clearer communication reduces cognitive load

Where it lags

Hovers ~4th, not top of every signal
May trail Opus-class depth on edge cases
Weekly token limits hit during heavy use; EU access delayed to mid-July

Continue reading

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