Grok 4.1

heading: LMArena Text Leaderboard · quote_snippet: Grok 4.1 Thinking (code name: quasarflux) holds the #1 overall position with 1483 Elo

{
  "harness": null,
  "tools": null,
  "shots": null,
  "reasoning_effort": "Thinking mode",
  "temperature": null,
  "top_p": null,
  "token_budget": null,
  "turn_limit": null,
  "time_limit": null,
  "run_count": null,
  "aggregation": "leaderboard",
  "judge": null
}

Maps to existing benchmark arena (LMArena Text Arena). Page states a 31-point margin over the highest non-xAI model; codename quasarflux is the arena submission identity. Leaderboard Elo is a moving snapshot. Grok 4 by comparison ranked #33 overall per the same page.

打开官方来源

heading: LMArena Text Leaderboard · quote_snippet: Grok 4.1 in its non-reasoning mode (code name: tensor)... ranks #2 at 1465 Elo

{
  "harness": null,
  "tools": null,
  "shots": null,
  "reasoning_effort": "non-thinking (no thinking tokens, immediate response)",
  "temperature": null,
  "top_p": null,
  "token_budget": null,
  "turn_limit": null,
  "time_limit": null,
  "run_count": null,
  "aggregation": "leaderboard",
  "judge": null
}

Page highlights that this NON-thinking configuration surpasses every other model's full-reasoning configuration on the public leaderboard — a cross-effort comparison claim worth flagging when reading the ranking.

打开官方来源

heading: Emotional Intelligence / EQ-Bench · figure: EQ-Bench recharts SVG ("Emotional Intelligence Benchmark - Elo (Normalized)"); numeric labels present in the rendered SVG text of page.html · quote_snippet: EQ-Bench is a LLM-judged test... We report the rubric score and normalized Elo score by running the official benchmark repository

{
  "harness": "official EQ-Bench3 repository",
  "tools": null,
  "shots": null,
  "reasoning_effort": "Thinking mode",
  "temperature": "default sampling parameters prescribed by the benchmark",
  "top_p": null,
  "token_budget": null,
  "turn_limit": null,
  "time_limit": null,
  "run_count": null,
  "aggregation": "rubric score + normalized Elo",
  "judge": "Claude Sonnet 3.7 (prescribed judge)"
}

new-benchmark: eq-bench3 not yet in data/benchmarks/. Values recovered 2026-09-01 from the chart's own SVG text in the archived page.html (not an image OCR path): Grok 4.1 Thinking 1586, Grok 4.1 (non-thinking) 1585, Kimi K2 Instruct 1561, Horizon Alpha 1559, Gemini 2.5 Pro 1460, GPT-5 Chat 1364, Claude Opus 4 1304, Grok 4 1206. Row records the Thinking config; the 1-point non-thinking gap is within Elo noise. Page states scores were computed with default sampling parameters, prescribed judge (Claude Sonnet 3.7) and no system prompt, run from the official EQ-Bench3 repository; xAI notes it is working with the benchmark author to publish leaderboard numbers.

打开官方来源

heading: Creative Writing / Creative Writing v3 · figure: Creative Writing v3 recharts SVG ("Judging creative writing reliably - Elo (Normalized)"); numeric labels present in the rendered SVG text of page.html · quote_snippet: models generate responses to 32 distinct writing prompts across 3 iterations

{
  "harness": null,
  "tools": null,
  "shots": null,
  "reasoning_effort": "Thinking mode",
  "temperature": null,
  "top_p": null,
  "token_budget": null,
  "turn_limit": null,
  "time_limit": null,
  "run_count": null,
  "aggregation": "rubrics + model battle normalized Elo",
  "judge": null
}

new-benchmark: creative-writing-v3 not yet in data/benchmarks/. Values recovered 2026-09-01 from the chart's own SVG text in the archived page.html: Polaris Alpha (early GPT 5.1) 1756.2 / Grok 4.1 Thinking 1721.9 / Grok 4.1 (non-thinking) 1708.6 / o3 1696.4 / Claude Sonnet 4.5 1648.7 / Kimi K2 Instruct 1627.5 / Grok 3 1126. Row records the Thinking config. Scores computed via rubrics + model-battle normalized Elo (same machinery as EQ-Bench3).

打开官方来源

heading: Reduced Hallucinations / FActScore · figure: FActScore recharts SVG (two-bar chart with % labels) plus Grok 4 Fast (Non-Reasoning) / Grok 4.1 (Non-Reasoning) legend; labels present in the rendered SVG text of page.html · quote_snippet: We also evaluate FActScore, which is a public benchmark consisting of 500 biography questions on individuals

{
  "harness": null,
  "tools": [
    "web search"
  ],
  "shots": null,
  "reasoning_effort": "non-reasoning (Fast)",
  "temperature": null,
  "top_p": null,
  "token_budget": null,
  "turn_limit": null,
  "time_limit": null,
  "run_count": null,
  "aggregation": null,
  "judge": null
}

new-benchmark: factscore not yet in data/benchmarks/. Direction flag: LOWER is better per the page's own "Lower score is better" subtitle. Values recovered 2026-09-01 from the chart's SVG text in the archived page.html: Grok 4.1 (Non-Reasoning) 2.97% vs baseline Grok 4 Fast (Non-Reasoning) 9.89%; the paired Hallucination Rate chart on the same grid shows 4.22% vs 12.09% on internal production-traffic queries (internal metric, kept in notes only). Methodology per prose: non-reasoning model evaluated with web search tools; hallucination rate = macro-average of atomic claims with major/minor errors.

打开官方来源