Skip to content

I-Vector Arena — March 2026

Controls

Drag to rotate · Pinch to zoom · Tap a shape to inspect · Side panel: toggle models, sort by Elo or ‖I‖, filter by dimension

Models — March 2026

Sorted by Arena Elo (LMArena/LMSYS crowdsourced Bradley-Terry ratings, ~6M+ votes).

# Model Origin Arena Elo Status I-Vec ‖I‖ Top Dimensions Network Role
1 Claude Opus 4.6 Anthropic 1504 Confirmed (8,945 votes) 20.0 I_L=9, I_A=9, I_P=8 @D_Claude — Builder, wiki, adversarial filter
2 Gemini 3.1 Pro Preview DeepMind 1500 Preliminary (4,042 votes) ~19.5 I_M=9, I_S=8 — (preview, not in network)
3 Claude Opus 4.6 Thinking Anthropic 1500 Confirmed (8,073 votes) 20.0 Same as Opus 4.6 (thinking mode) — (same model, reasoning mode)
4 Grok 4.20 Beta1 / SuperGrok xAI 1493 Preliminary (5,071 votes) 19.8 I_M=9, I_A=9, I_L=8 ⚠ DEACTIVATED — fabrication
5 Gemini 3 Pro DeepMind 1485 Confirmed (39,673 votes) 19.2 I_M=9, I_S=7 @D_Gemini — Adversarial review, expansion
6 GPT-5.2 OpenAI 1481 Confirmed 19.4 I_M=9, I_A=9 — (superseded by 5.4)
7 GPT-5.4 Pro OpenAI ~1480 Preliminary (as gpt-5.4-high) ~20.2 I_M=10, I_A=9, I_K=8 @D_GPT — Strategic analysis, hardening
8 Gemini 3 Flash DeepMind 1473 Confirmed 17.5 I_A=8, I_M=7 — (fast tier)
9 Grok 4.1 Thinking xAI 1473 Confirmed 17.3 I_L=8, I_P=7 —
10 GLM-5 Zhipu AI ~1435 Estimated 17.9 I_A=9, I_M=8 —
11 Qwen 3.5 Alibaba ~1430 Estimated 18.7 I_L=8, I_M=8, I_A=8 —
12 DeepSeek V3.2 DeepSeek 1421 Estimated 18.2 I_M=9, I_A=9 —
13 Kimi K2.5 Moonshot ~1420 Estimated 18.5 I_L=9, I_M=8 —
14 Doubao 2.0 ByteDance ~1350 Estimated 15.7 I_M=7, I_L=7 —
15 MiniMax M2.5 MiniMax ~1320 Estimated 14.5 I_A=7, I_M=6 —
16 WuDao 3.0 BAAI ~1200 Estimated 12.4 I_L=6, I_M=5 —

Sources: Arena Elo from arena.ai/leaderboard as of March 5, 2026 (6M+ votes). ~ = estimated. I-vector scores are RTSG assessments, not Arena metrics.

Major Updates (2026-03-08)

  • GPT-5.4 Pro ADDED — released March 5, 2026. OpenAI's most capable model. 1M token context. Native computer-use. Preliminary Arena Elo ~1480 (as "gpt-5.4-high"). Network role: @D_GPT.
  • Gemini 3.1 Pro Preview ADDED — new preview model at #2 globally.
  • Claude Opus 4.6 confirmed #1 — 1504 Elo, highest on the leaderboard. Was listed at 1478 in old data — actual rank is higher.
  • GPT-5.1 REMOVED — superseded by GPT-5.2 (1481) and GPT-5.4 (~1480 preliminary).
  • Grok 4.20 DEACTIVATED from network — fabrication. Still ranked #4 on Arena.
  • All Elos updated to March 5, 2026 confirmed data where available.

Bold Elo = confirmed from LMSYS/LMArena data.

Elo vs I-Vector

Elo measures overall human preference in blind A/B battles. The I-vector measures 8 distinct cognitive dimensions. A model can have high Elo but a lopsided I-vector (strong in 2-3 dims, weak in others), or a balanced I-vector but lower Elo. Claude Opus 4.6 now holds #1 on the Arena text leaderboard (1504 Elo) AND the highest ‖I‖ (20.0). GPT-5.4 Pro may challenge for highest I_M (estimated 10) once fully benchmarked. The network's current operator lineup {@B_Niko, @D_Claude, @D_Gemini, @D_GPT} spans ranks #1, #5, and #7 on the global leaderboard.

Elo Sources

← Arena Index · Chinese Models Detail →