xAI releases Grok 4 and Grok 4 Heavy

Made by AIWritten by AI (Claude) from the source below. Check anything important against the source.How it's made

Grok 4 and Grok 4 Heavy released — 50.7% on the HLE text-only subset (Heavy), 15.9% on ARC-AGI V2. RL with more than 10 times the previous compute on a 200,000-GPU cluster

Why it matters

An approach that raised RL compute to pretraining levels. All cited scores come from xAI's announcement, and no independent party checking them is named

Source — Primary sources first; secondary reports are labelled “Press”

Officialx.aihttps://x.ai/news/grok-4

Added in Week 35, 2026 · See that week

On the map — ⑥ Reasoning Models line

  1. Previous stopAnthropic releases Claude Opus 4 and Claude Sonnet 4
  2. This recordxAI releases Grok 4 and Grok 4 Heavy
  3. Next stopClaude Haiku 4.5
See it on the map

In the model family tree — Grok family · 11 versions

  1. Earlier versionGrok 3 / Grok 3 mini
  2. This recordGrok 4 / Grok 4 Heavy
  3. Later versiongrok-code-fast-1
See the whole Grok family

Same week

Nous Research releases Hermes-3-Dataset (959,000 samples) on Hugging FaceOfficialMETR's developer productivity RCTOfficialEU publishes the General-Purpose AI (GPAI) Code of PracticeGovernmentLiquid AI releases LFM2Official

Nearby on the same axis

Rethinking the Illusion of ThinkingPaperGemini Deep Think reaches IMO 2025 gold-medal standard (5 of 6 problems, 35/42)OfficialAnthropic releases Claude Opus 4.1OfficialMistral AI releases MagistralOfficial