Grok 3.xAI's reasoning flagship — AIME 93.3%, trained on Colossus.

Grok 3 is xAI's flagship large language model, announced by Elon Musk and the xAI team on February 17, 2025. It was trained on Colossus, xAI's Memphis supercomputer with 200,000 Nvidia H100 GPUs — roughly ten times the compute used for Grok 2. Grok 3 ships in four variants: the flagship Grok 3, the smaller Grok 3 Mini, a Think mode with visible chain-of-thought reasoning, and a Big Brain mode that allocates extra test-time compute. xAI positions Grok 3 as a frontier model that matches or beats GPT-4o and Claude 3.5 Sonnet on math, science, and code benchmarks.

On xAI's published benchmarks, Grok 3 (Think) scores 93.3% on AIME 2024 versus GPT-4o's 9.3% and Claude 3.5 Sonnet's 16.0%. On GPQA Diamond it scores 84.6% versus GPT-4o's 53.6% and Claude 3.5 Sonnet's 65.0%. On LiveCodeBench it scores 79.4% versus GPT-4o's 32.3% and Claude 3.5 Sonnet's 40.9%. On the LMArena chatbot leaderboard in late February 2025, Grok 3 briefly held the #1 slot at 1402 Elo, narrowly ahead of GPT-4o and Gemini 2.0 Pro.

Think mode is Grok 3's reasoning variant — it streams visible chain-of-thought before producing a final answer, trained with reinforcement learning in the same family as OpenAI's o-series and DeepSeek R1. Big Brain mode is an escalated Think variant that allocates additional test-time compute for the hardest prompts; xAI describes it as the highest-quality configuration Grok 3 can run in. Both modes boost Grok 3's AIME score from about 52% (base) to 93.3% (Think).

Grok 3's API exposes a 1,000,000-token context window, one of the largest of any frontier model in early 2025 — matched only by Gemini 2.0 Pro's 2M token window. The consumer interface on X and grok.com typically exposes a smaller context, around 128K tokens, consistent with other chat-tier deployments.

Three tiers. API: Grok 3 is priced at $3.00 per million input tokens and $15.00 per million output tokens on the xAI API; Grok 3 Mini is cheaper. Consumer: Grok 3 is free with daily limits on grok.com and X, and included in X Premium+ starting at $22/month. On ZeroTwo, Grok 3 is included alongside Grok 4, Claude, GPT-5, and Gemini on the free tier and unlimited on Pro at $29.99/month.

Colossus is xAI's Memphis, Tennessee supercomputer, brought online in September 2024. It started at 100,000 Nvidia H100 GPUs — at launch the largest single AI training cluster in the world — and was expanded to 200,000 H100s before Grok 3 finished training. Musk publicly stated Grok 3 used roughly ten times the compute of Grok 2, making it the most compute-intensive open frontier model training run disclosed at the time.

Grok 3 launched with DeepSearch, xAI's agentic research mode that browses the live web and X, synthesizes findings, and returns cited answers — conceptually similar to OpenAI's Deep Research and Perplexity Pro. Combined with Think or Big Brain, it can run multi-step research plans, cross-check sources, and produce long-form reports. API developers can also wire Grok 3 into tool-calling agents via the xAI function-calling endpoints.

Yes. You can use Grok 3 on grok.com with a free daily-limited tier, on the xAI API (billed pay-as-you-go), via OpenRouter, or inside ZeroTwo's multi-model workspace. On ZeroTwo you get Grok 3, Grok 3 Mini, Grok 4, Claude Sonnet 4.6, GPT-5, and Gemini 3 in one subscription — cheaper than X Premium+ and without requiring an X account.

Yes, for three reasons. First, price: Grok 3 remains the cheaper tier on the xAI API and is often the better cost-performance pick for chat, extraction, and RAG. Second, availability: Grok 3 Mini sees wider integration across OpenRouter, LangChain, and aggregators than the latest Grok 4 variants. Third, reasoning: Grok 3 (Think)'s AIME and GPQA scores are still competitive with most frontier reasoning tiers. For bulk agentic work, Grok 3 Mini often beats Grok 4 on $/task.