Grok 3 Launch and Current Model Record

Grok 3 Launch and Current Model Record

Historical definition

xAI introduced Grok 3 on February 19, 2025 as an early preview and beta model family. The company described Grok 3 and Grok 3 mini, with standard and reasoning-oriented variants. It also introduced Think and DeepSearch product experiences around reasoning and tool use.

Primary launch source:

https://x.ai/news/grok-3

Performance numbers on that page are xAI-reported evaluations. They should retain the named benchmark, configuration, date, and vendor attribution. They are not an independent general ranking or a guarantee for a reader's task.

The launch post said API access was coming later. xAI's release notes record Grok 3 models becoming generally available through the API on April 3, 2025. The launch experience, later API release, and current model catalog are separate facts.

Release notes:

https://docs.x.ai/developers/release-notes

Infrastructure boundary

xAI attributes Grok 3 training to its Colossus supercluster and publishes current infrastructure claims on its Colossus page.

https://x.ai/colossus

Colossus scale, construction, storage, network, energy, and expansion figures are company claims unless supported independently. Infrastructure scale is not itself a model-quality measurement and does not prove safety, reliability, product fit, or software-development readiness.

Current model boundary

The current model directory checked July 28, 2026 identifies Grok 4.5 as xAI's current flagship in the visible catalog. Grok 3 is therefore a historical model family for this article, not a current-frontier recommendation.

https://docs.x.ai/developers/models

The model catalog can change. Current implementation work must verify the exact model identifier, lifecycle, context window, tools, modalities, supported regions, rate limits, price, data policy, terms, and retirement notice at the time of use.

Current API documentation:

https://docs.x.ai/

Product-mode distinction

A model is not the same object as a consumer product mode or an API surface. "Grok 3," "Grok 3 mini," "Think," "DeepSearch," grok.com, X integration, and API models need separate descriptions.

The public explainer should not infer that a capability shown in a consumer interface was available through the API on launch day. It should not infer that a current API feature existed in February 2025.

Evaluation boundary

A historical benchmark belongs to the historical release record. A current buying or architecture decision requires a task-specific evaluation against current candidates.

The evaluation record should capture prompt set, expected outputs, scoring rule, tool access, temperature or reasoning controls, failures, latency, token use, price, reviewer agreement, safety cases, data restrictions, and date. Vendor leaderboard language alone cannot select a production model.