Grok 3

Retired · Language, Reasoning · Milestone

Grok 3 is a language model from xAI with a reasoning mode, Grok 3 (Think). It launched as a beta on 19 February 2025 for X Premium and Premium+ users on X and grok.com, became generally available on the xAI API on 3 April 2025 and was retired from the API on 15 May 2026. xAI says it was trained on its Colossus supercluster. [1] [2] [3] Primary source

Timeline of Grok 3 →

Claims and evidence

  • Released The post says Grok 3 is now available to X Premium and Premium+ users on X and grok.com, as a beta; it also says the rollout to users would take place over the coming days. Primary source[1]closing section
  • Retired Effective 12:00 PM PT on the xAI API. Primary source[3]first paragraph
  • Generally available on the xAI API Primary source[2]April 2025, Apr 3 [4]Our path of progress, 2025: Grok 3 API
  • Status RetiredRetired from the xAI API; requests to grok-3 are served by grok-4.3. Primary source[3]list of retired models
  • Replaced by Grok 4.3grok-3 redirects to grok-4.3 with reasoning effort none. Primary source[3]Recommended Replacements
  • Successor of Grok-2 (December 2024 revision) EditorialEditorial link along the main Grok line: the announcement names no predecessor model, and the December 2024 Grok-2 was the current flagship before it. Primary source[1]
  • Change · Context length Context window of 1 million tokens, which xAI called eight times larger than its previous models; the grok-3 API model page listed 131,072 tokens.Compared with Grok-2 (December 2024 revision) Primary source[1]Pretraining on a Massive Scale [5]At a glance
  • Change · Reasoning Adds a beta reasoning mode, Grok 3 (Think), trained with large-scale reinforcement learning to think for seconds to minutes before answering.Compared with Grok-2 (December 2024 revision) Primary source[1]Thinking Harder: Test-time Compute and Reasoning
  • Change · Other xAI says it trained Grok 3 on its Colossus supercluster with ten times the compute of earlier leading models.Compared with Grok-2 (December 2024 revision) Primary source[1]Pretraining on a Massive Scale
  • Change · Tool use Launched together with DeepSearch, an xAI agent that searches for information and writes a summary report.Compared with Grok-2 (December 2024 revision) Primary source[1]
  • Input textAs listed for the API model grok-3 (drawn as icons: text to text). The announcement also reports image and video understanding evaluations, which are not reflected here. Archived primary[5]At a glance, Modalities
  • Output text Archived primary[5]At a glance, Modalities
  • Feature Reasoning modeGrok 3 (Think), a beta reasoning mode at launch, switched on with the Think button. Primary source[1]Thinking Harder: Test-time Compute and Reasoning
  • Context window 1 million tokens tokensThe xAI API model page listed a context window of 131,072 tokens for grok-3 (src.xai.grok-3-api-model-page). Primary source[1]Pretraining on a Massive Scale
  • Parameters Not disclosedNot disclosed by xAI. No claim madeNo source
  • Access consumer app, API Primary source[1]closing section [2]April 2025, Apr 3

Lineage

Predecessors

Successors

Based on

Not derived from another model.

Variants and derived

None recorded.

Siblings

None recorded.

All ancestors

All descendants

Variants

No variants recorded in this record.

Related AI Radar coverage

AI Radar coverage starts in June 2026; no coverage linked yet.

All model releases from xAI on AI Radar →