HomeFull speed ahead. Keep track of advancements in every frontier
ColossusxAI—GrokxAI—GrokxAI—GrokxAI—ColossusxAI—xAI—ColossusxAI—GrokxAI—ColossusxAI—

Grok

by xAIUS

ScalingStage 5 of 5

Grok 4.x runs on X, the Grok app and for developers. Work continues inside SpaceX.

Updated 19d agoChecked 25 Sep0 updates this week

Milestones

Next · Grok 5 release
  1. Grok-1 announced to X Premium usersNov 2023Complete.
  2. Grok-1 weights open-sourced17 Mar 2024Complete.
  3. Grok-213 Aug 2024Complete.
  4. Grok 3, trained on Colossus17 Feb 2025Complete.
  5. Grok 4 and Grok 4 Heavy9 Jul 2025Complete.
  6. Grok 4.117 Nov 2025Complete.
  7. Grok 5 releaseNowCurrent milestone.

Most important updates

  • 19d ago
  • 12 Aug 2026
  • 8 Jul 2026
  • 9 Jul 2025
  • 17 Feb 2025

Upcoming

  1. 2026Grok 5, which Musk said in August 2026 should arrive within 2026 (next)

Current obstacles

  • Safety and moderationPublic incidents show how hard it is to make a less restricted model avoid harmful content.
  • Staff turnoverAfter the SpaceX merger xAI restructured, and about half its co-founders had left, reports say.

Physics limits

  • Only checkable skills train wellReasoning is learned by rewarding answers a program can grade, such as maths and code. Where no automatic checker exists (strategy, taste, open research) the signal is weak.
  • Errors compound over long tasksIf each step succeeds 99% of the time, a 100-step task succeeds only about 37% of the time (0.99^100). Agents need near-perfect per-step reliability to work for hours.
  • Returns shrink as a power lawError falls only as a small power of training compute: each fixed step of improvement needs roughly 10x more compute, energy and money than the one before.

How it works

3 parts
Open cabinets of GPU compute blades (the Alps supercomputer in Switzerland): Grok is trained on racks of this kind in Colossus
Open cabinets of GPU compute blades (the Alps supercomputer in Switzerland): Grok is trained on racks of this kind in ColossusPhoto: CSCS · CC BY-SA 3.0 (opens commons.wikimedia.org)
Model

Big model, heavy practice

Pretrained on text, then given unusually heavy reinforcement learning (rewarded trial and error) on checkable problems, on xAI's Colossus cluster.

Heavy

Several agents at once

Grok 4 Heavy runs several agents on a problem in parallel and compares their answers.

Search

Live search

Grok searches X and the web when asked and cites posts, so its information is up to date.

Papers & demos

2 items
  1. Aug 2025blog
    Grok 4 model cardDescribes how Grok 4 was trained and safety-tested.
  2. Mar 2024code
    Grok-1 open weights (314B MoE)Released the weights of its first model for anyone to use.

Update log

9 updates

Mon 21 Sep

  • Major: BlogSoftware

Wed 12 Aug

  • Minor: BlogSoftware

Wed 8 Jul

  • Minor: BlogSoftware

Mon 17 Nov 2025

  • Minor: BlogSoftware

Wed 9 Jul 2025

  • Major: BlogSoftware

Tue 8 Jul 2025

  • Minor: PressOther

Mon 17 Feb 2025

  • Major: BlogSoftware

Tue 13 Aug 2024

  • Minor: BlogSoftware

Sun 17 Mar 2024

  • Minor: BlogSoftware

About xAI

The team behind Grok

xAI

Grok models, Colossus supercomputer

xAI, founded by Elon Musk in 2023, makes the Grok models used on X, which it absorbed in 2025. It built its Colossus supercomputer in Memphis unusually fast. SpaceX bought xAI in February 2026, valuing it at $250B.

  • Bought by SpaceX on 2 February 2026, in a deal valuing SpaceX at $1T and xAI at $250B.
  • Colossus went from empty building to 100,000 H100 GPUs in about four months in 2024.
  • Grok 4 (July 2025) and Grok 4.1 (November 2025) ranked near the top on reasoning tests.
Founded
20233 yrs
Headquarters
United States
Status
Subsidiaryof SpaceX
Valuation
$250BSpaceX-ownedFeb 2026
Raised
$32B3 rounds
Last round
Series E · $20BJan 2026
Works in
AIFrontier modelsComputingSpaceflight
Coverage
2 programs · 18 updateslatest 19d agochecked 25 Sep
Partners
XTesla
Suppliers
NVIDIA
Acquired by
SpaceX
People
Elon MuskFounder
Primer