xAI / Grok
The frontier challenger the synthesis credited with leading on context but hadn’t paged. xAI (Elon Musk, founded 2023) makes Grok; it sits with openai and anthropic (cross-wiki) on the proprietary frontier pole — but with a distinctive open-then-closed licensing arc and a very large context window.
The licensing arc — open, then closed
- Grok-1 (Nov 2023) was released under Apache-2.0 — genuinely open weights.
- Every version since (Grok-1.5 → 2 → 3 → 4 → 4.1) is proprietary (xAI-specific license). So Grok traces the opposite trajectory from the open-weight labs: it started open and closed up as it scaled — a useful counter-example to the “open weights only move one direction” read, and another point on the synthesis’s “open is a spectrum” thread.
Context & capability
- Context window grew from 128K (Grok-1.5) to 2M tokens in recent enterprise variants — the basis for the synthesis’s “xAI leads context” claim on llm-benchmarks.
- Grok-3 (Feb 2025) was trained with ~10× the compute of Grok-2 and claimed to beat GPT-4o on math benchmarks (AIME); Grok-4 (July 2025)/4.1 (Nov 2025) are the current flagships.
- Distribution moat: wired directly into X (Twitter), standalone iOS/Android apps, embedded in Tesla vehicles, plus an API — a distribution channel no other frontier lab has.
Where it sits
Grok rounds out the frontier-premium pole next to openai: proprietary, reasoning-focused, premium-priced — but differentiated by context length and native X/Tesla distribution rather than the cheapest API. Its open→closed arc is the inverse of llama/qwen, making it a clean illustration that “open vs closed” is a strategy over time, not a fixed label. (Benchmark/context claims are vendor-stated, dated snapshots — the standing caveat.)
Related
openai · llm-benchmarks · llm-provider · open-weight-models · llama · qwen