Frontier · who leads the AI leaderboards, and for how long

Gemini 4 Carbon (High) leads Arena’s Text board

Open-weights models took a median of 127 days to match each closed leader.

Arena scores come from people voting between two anonymous answers. First place is the highest score; the top two are often within each other’s margin of error. Arena data to 28 October 2026.

Model retirements: Amazon Bedrock retires five models on Friday 13 November. Each vendor’s dates, quoted from its own page.

Open-weight downloads: Alibaba drew 45% of the Frontier labs’ open-model downloads. Read from the Hugging Face Hub each week.

Model usage: OpenRouter’s daily model ranks and task shares, kept each week. The first read is pending.

Embed

Credits LMArena.

Credits LMArena.

Credits Epoch AI.

Each frame links back to its page. Preview it.

Cite
Download
Post on X

OpenAI held first place in Text for 487 days of 1,269

four arenas · 626 snapshots

Text

Now: Gemini 4 Carbon (High), first since 28 October 2026

WebDev

Now: claude-opus-5.5-max, first since 23 September 2026 (35 days)

Vision

Now: claude-fable-5-high, first since 28 September 2026 (21 days)

Text to image

Now: gpt-image-2.5-sunburst, first since 8 September 2026 (48 days)

proprietary open weights · Each block is one model’s time in first place. The board began in May 2023 for Text; the others later.

Open every reign · 81 in four arenas
Arena Model Lab Weights From To Days
Text Gemini 4 Carbon (High) Google proprietary 2026-10-28 now 0
Text gemini-4-argon-high Google proprietary 2026-09-30 2026-10-28 28
Text claude-opus-5.5-high Anthropic proprietary 2026-09-25 2026-09-30 5
Text claude-fable-5.1-max Anthropic proprietary 2026-09-02 2026-09-25 23
Text claude-opus-5-max Anthropic proprietary 2026-08-27 2026-09-02 6
Text claude-opus-5-high Anthropic proprietary 2026-08-19 2026-08-27 8
Text claude-opus-5-max Anthropic proprietary 2026-07-27 2026-08-19 23
Text claude-opus-5-high Anthropic proprietary 2026-07-26 2026-07-27 1
Text claude-opus-4-6-thinking Anthropic proprietary 2026-03-18 2026-07-26 130
Text claude-opus-4-6 Anthropic proprietary 2026-02-19 2026-03-18 27
Text claude-opus-4-6-thinking Anthropic proprietary 2026-02-06 2026-02-19 13
Text gemini-3-pro Google proprietary 2025-11-16 2026-02-06 82
Text gemini-2.5-pro Google proprietary 2025-06-24 2025-11-16 145
Text o3-2025-04-16 OpenAI proprietary 2025-04-23 2025-06-24 62
Text chatgpt-4o-latest-20250326 OpenAI proprietary 2025-03-26 2025-04-23 28
Text grok-3-preview-02-24 xAI proprietary 2025-03-03 2025-03-26 23
Text deepseek-r1 DeepSeek open weights 2025-01-24 2025-03-03 38
Text o1-2024-12-17 OpenAI proprietary 2024-12-30 2025-01-24 25
Text o1-preview OpenAI proprietary 2024-09-17 2024-12-30 104
Text grok-2-2024-08-13 xAI proprietary 2024-09-15 2024-09-17 2
Text chatgpt-4o-latest OpenAI proprietary 2024-08-22 2024-09-15 24
Text gpt-4o-2024-05-13 OpenAI proprietary 2024-05-15 2024-08-22 99
Text gpt-4-turbo-2024-04-09 OpenAI proprietary 2024-04-11 2024-05-15 34
Text claude-3-opus-20240229 Anthropic proprietary 2024-03-26 2024-04-11 16
Text gpt-4-1106-preview OpenAI proprietary 2024-02-15 2024-03-26 40
Text gpt-4-0125-preview OpenAI proprietary 2024-02-02 2024-02-15 13
Text gpt-4-0314 OpenAI proprietary 2023-12-06 2024-02-02 58
Text wizardlm-70b Microsoft open weights 2023-10-02 2023-12-06 65
Text vicuna-33b LMSYS open weights 2023-07-17 2023-10-02 77
Text guanaco-33b UW open weights 2023-06-19 2023-07-17 28
Text vicuna-13b LMSYS open weights 2023-05-08 2023-06-19 42
WebDev claude-opus-5.5-max Anthropic proprietary 2026-09-23 now 35
WebDev gpt-6-astra-max OpenAI proprietary 2026-09-05 2026-09-23 18
WebDev claude-fable-5.1-max Anthropic proprietary 2026-09-02 2026-09-05 3
WebDev qwen3.8-max-0902 Alibaba proprietary 2026-09-01 2026-09-02 1
WebDev claude-opus-5-max Anthropic proprietary 2026-07-27 2026-09-01 36
WebDev kimi-k3-max Moonshot AI open weights 2026-07-26 2026-07-27 1
WebDev kimi-k3 Moonshot AI proprietary 2026-07-16 2026-07-26 10
WebDev gpt-5.6-sol-xhigh (codex-harness) OpenAI proprietary 2026-07-12 2026-07-16 4
WebDev claude-fable-5 Anthropic proprietary 2026-06-10 2026-07-12 32
WebDev claude-opus-4-7-thinking Anthropic proprietary 2026-04-20 2026-06-10 51
WebDev claude-opus-4-7 Anthropic proprietary 2026-04-17 2026-04-20 3
WebDev claude-opus-4-6-thinking Anthropic proprietary 2026-04-02 2026-04-17 15
WebDev claude-opus-4-6 Anthropic proprietary 2026-03-16 2026-04-02 17
WebDev claude-opus-4-6-thinking Anthropic proprietary 2026-03-04 2026-03-16 12
WebDev claude-opus-4-6 Anthropic proprietary 2026-02-17 2026-03-04 15
WebDev claude-opus-4-6-thinking Anthropic proprietary 2026-02-09 2026-02-17 8
WebDev claude-opus-4-6 Anthropic proprietary 2026-02-06 2026-02-09 3
WebDev claude-opus-4-5-20251101-thinking-32k Anthropic proprietary 2025-11-26 2026-02-06 72
WebDev gemini-3-pro Google proprietary 2025-11-16 2025-11-26 10
WebDev claude-sonnet-4-5-20250929-thinking-32k Anthropic proprietary 2025-11-12 2025-11-16 4
Vision claude-fable-5-high Anthropic proprietary 2026-09-28 now 21
Vision claude-fable-5 Anthropic proprietary 2026-06-10 2026-09-28 110
Vision claude-opus-4-7-thinking Anthropic proprietary 2026-06-03 2026-06-10 7
Vision claude-opus-4-7 Anthropic proprietary 2026-05-12 2026-06-03 22
Vision claude-opus-4-7-thinking Anthropic proprietary 2026-05-07 2026-05-12 5
Vision claude-opus-4-6-thinking Anthropic proprietary 2026-04-27 2026-05-07 10
Vision claude-opus-4-7 Anthropic proprietary 2026-04-22 2026-04-27 5
Vision claude-opus-4-7-thinking Anthropic proprietary 2026-04-19 2026-04-22 3
Vision claude-opus-4-6-thinking Anthropic proprietary 2026-04-10 2026-04-19 9
Vision claude-opus-4-6 Anthropic proprietary 2026-03-31 2026-04-10 10
Vision gemini-3-pro Google proprietary 2025-11-16 2026-03-31 135
Vision gemini-2.5-pro Google proprietary 2025-06-24 2025-11-16 145
Vision chatgpt-4o-latest-20250326 OpenAI proprietary 2025-03-26 2025-06-24 90
Vision gpt-4.5-preview-2025-02-27 OpenAI proprietary 2025-03-20 2025-03-26 6
Vision o1-2024-12-17 OpenAI proprietary 2025-03-11 2025-03-20 9
Vision gemini-2.0-flash-001 Google proprietary 2025-02-05 2025-03-11 34
Vision gemini-1.5-pro-002 Google proprietary 2024-11-04 2025-02-05 93
Vision gpt-4o-2024-05-13 OpenAI proprietary 2024-06-29 2024-11-04 128
Text to image gpt-image-2.5-sunburst OpenAI proprietary 2026-09-08 now 48
Text to image gpt-image-2 (medium) OpenAI proprietary 2026-04-21 2026-09-08 140
Text to image gemini-3.1-flash-image-preview (nano-banana-2) [web-search] Google proprietary 2026-02-25 2026-04-21 55
Text to image gpt-image-1.5-high-fidelity OpenAI proprietary 2025-12-16 2026-02-25 71
Text to image gemini-3-pro-image-preview-2k (nano-banana-pro) Google proprietary 2025-12-02 2025-12-16 14
Text to image gemini-3-pro-image-preview (nano-banana-pro) Google proprietary 2025-11-21 2025-12-02 11
Text to image hunyuan-image-3.0 Tencent open weights 2025-10-01 2025-11-21 51
Text to image gemini-2.5-flash-image-preview (nano-banana) Google proprietary 2025-08-14 2025-10-01 48
Text to image imagen-ultra-4.0-generate-001 Google proprietary 2025-08-04 2025-08-14 10
Text to image gpt-image-1 OpenAI proprietary 2025-06-03 2025-08-04 62
Text to image imagen-3.0-generate-002 Google proprietary 2025-01-22 2025-06-03 132
Text to image recraft-v3 Recraft proprietary 2025-01-05 2025-01-22 17

In 40.3% of Text snapshots, first place was clear of the margin of error.

The best open-weights model in Text is number 12, 49 points behind first place

mimo-v2.6-pro · Xiaomi · MIT

Open weights took a median of 127 days to match a closed leader in Text

14 caught, from 25 days to 313 days · 7 still open · 2 left the board first

  1. gpt-4-0314 125 days by command-r-plus
  2. gpt-4-0125-preview 179 days by llama-3.1-405b-instruct
  3. gpt-4-1106-preview 166 days by llama-3.1-405b-instruct
  4. claude-3-opus-20240229 126 days by llama-3.1-405b-instruct
  5. gpt-4-turbo-2024-04-09 110 days by llama-3.1-405b-instruct
  6. gpt-4o-2024-05-13 229 days by deepseek-v3
  7. chatgpt-4o-latest left after 13 days
  8. grok-2-2024-08-13 106 days by deepseek-v3
  9. o1-preview 129 days by deepseek-r1
  10. o1-2024-12-17 25 days by deepseek-r1
  11. grok-3-preview-02-24 106 days by deepseek-r1-0528
  12. chatgpt-4o-latest-20250326 128 days by qwen3-235b-a22b-instruct-2507
  13. o3-2025-04-16 100 days by qwen3-235b-a22b-instruct-2507
  14. gemini-2.5-pro 287 days by glm-5.1
  15. gemini-3-pro 313 days by mimo-v2.6-pro
  16. claude-opus-4-6-thinking left after 186 days
  17. claude-opus-4-6 251 days, open
  18. claude-opus-5-high 94 days, open
  19. claude-opus-5-max 93 days, open
  20. claude-fable-5.1-max 56 days, open
  21. claude-opus-5.5-high 33 days, open
  22. gemini-4-argon-high 28 days, open
  23. Gemini 4 Carbon (High) 0 days, open

Open weights took a median of 60 days to match a closed leader in WebDev

9 caught, from 14 days to 134 days · 5 still open · 1 left the board first

  1. claude-sonnet-4-5-20250929-thinking-32k 40 days by glm-4.7
  2. gemini-3-pro 88 days by glm-5
  3. claude-opus-4-5-20251101-thinking-32k 134 days by glm-5.1
  4. claude-opus-4-6 130 days by glm-5.2 (max)
  5. claude-opus-4-6-thinking 127 days by glm-5.2 (max)
  6. claude-opus-4-7 60 days by glm-5.2 (max)
  7. claude-opus-4-7-thinking 57 days by glm-5.2 (max)
  8. claude-fable-5 46 days by kimi-k3-max
  9. gpt-5.6-sol-xhigh (codex-harness) 14 days by kimi-k3-max
  10. kimi-k3 left after 8 days
  11. claude-opus-5-max 93 days, open
  12. qwen3.8-max-0902 57 days, open
  13. claude-fable-5.1-max 56 days, open
  14. gpt-6-astra-max 53 days, open
  15. claude-opus-5.5-max 35 days, open

Open weights took a median of 189 days to match a closed leader in Vision

7 caught, from 118 days to 283 days · 4 still open · 3 left the board first

  1. gpt-4o-2024-05-13 283 days by qwen2.5-vl-32b-instruct
  2. gemini-1.5-pro-002 239 days by gemma-3-27b-it
  3. gemini-2.0-flash-001 146 days by gemma-3-27b-it
  4. o1-2024-12-17 118 days by gemma-3-27b-it
  5. gpt-4.5-preview-2025-02-27 151 days by step-3
  6. chatgpt-4o-latest-20250326 189 days by qwen3-vl-235b-a22b-instruct
  7. gemini-2.5-pro 219 days by kimi-k2.5-thinking
  8. gemini-3-pro 337 days, open
  9. claude-opus-4-6 202 days, open
  10. claude-opus-4-6-thinking left after 118 days
  11. claude-opus-4-7-thinking left after 109 days
  12. claude-opus-4-7 180 days, open
  13. claude-fable-5 left after 95 days
  14. claude-fable-5-high 21 days, open

Open weights took a median of 120 days to match a closed leader in Text to image

5 caught, from 48 days to 217 days · 4 still open · 2 left the board first

  1. recraft-v3 214 days by qwen-image
  2. imagen-3.0-generate-002 217 days by qwen-image-prompt-extend
  3. gpt-image-1 120 days by hunyuan-image-3.0
  4. imagen-ultra-4.0-generate-001 58 days by hunyuan-image-3.0
  5. gemini-2.5-flash-image-preview (nano-banana) 48 days by hunyuan-image-3.0
  6. gemini-3-pro-image-preview (nano-banana-pro) 339 days, open
  7. gemini-3-pro-image-preview-2k (nano-banana-pro) left after 217 days
  8. gpt-image-1.5-high-fidelity 314 days, open
  9. gemini-3.1-flash-image-preview (nano-banana-2) [web-search] left after 132 days
  10. gpt-image-2 (medium) 188 days, open
  11. gpt-image-2.5-sunburst 48 days, open

A catch-up runs from the day a proprietary model takes first place to the first snapshot in which an open-weights model matches its score.

A model scores 90% on 5 of Epoch AI’s 15 benchmarks

saturation: days from a benchmark’s release to the first model at 90%
  1. FrontierMath-Tier-4-v2-Private out 12 June 2026 0 days claude-fable-5_max, 90.2% · released before the benchmark
  2. FrontierMath-Tiers-1-3-v2-Private out 12 June 2026 · Epoch gives it a ceiling of 95.0% 81 days claude-fable-5-1_max, 90.2%
  3. OTIS Mock AIME 2024-2025 out 19 December 2024 231 days gpt-5-2025-08-07_high, 91.4%
  4. GPQA diamond out 20 November 2023 · Epoch gives it a ceiling of 97.0% 729 days gemini-3-pro-preview, 92.6%
  5. MATH level 5 out 5 March 2021 1,383 days o1-2024-12-17_high, 94.7%
  6. SWE-Bench verified out 13 August 2024 808 days so far best so far 83.5%, claude-opus-4-7_max
  7. FrontierMath-2025-02-28-Private out 28 February 2025 · replaced by a newer version · Epoch gives it a ceiling of 57.0% 609 days so far best so far 52.4%, gpt-5.5-pro-pre-release_high
  8. FrontierMath-Tier-4-2025-07-01-Private out 11 July 2025 · replaced by a newer version · Epoch gives it a ceiling of 60.0% 476 days so far best so far 47.9%, gdm-ai-co-mathematician
  9. SimpleQA Verified out 9 September 2025 416 days so far best so far 76.0%, gemini-4-argon_high
  10. Chess Puzzles out 10 December 2025 324 days so far best so far 76.0%, gpt-6-astra_promax
  11. MirrorCode out 26 June 2026 126 days so far best so far 78.9%, claude-opus-5-5_high
  12. EBR-bench out 1 July 2026 121 days so far best so far 78.1%, gpt-6.1-terra_max
  13. Mystery Game Puzzles out 30 July 2026 92 days so far best so far 84.0%, gpt-6-astra_max
  14. Furniture Assembly out 23 September 2026 37 days so far best so far 83.3%, claude-opus-5-5_max
  15. SpireClimb out 14 October 2026 16 days so far best so far 25.7%, gpt-6-astra_max

reached 90% not yet, counted to 30 October 2026

What changed

RSS feed →
  1. Gemini 4 Carbon (High) took first place in Text, from gemini-4-argon-high
  2. gemini-4-argon-high took first place in Text, from claude-opus-5.5-high
  3. claude-fable-5-high took first place in Vision, from claude-fable-5
  4. Open weights matched gemini-3-pro in Text: mimo-v2.6-pro, 313 days after gemini-3-pro took first place
  5. claude-opus-5.5-high took first place in Text, from claude-fable-5.1-max
  6. claude-opus-5.5-max took first place in WebDev, from gpt-6-astra-max
  7. gpt-image-2.5-sunburst took first place in Text to image, from gpt-image-2 (medium)
  8. gpt-6-astra-max took first place in WebDev, from claude-fable-5.1-max
  9. claude-fable-5.1-max took first place in Text, from claude-opus-5-max
  10. claude-fable-5.1-max took first place in WebDev, from qwen3.8-max-0902
  11. FrontierMath-Tiers-1-3-v2-Private reached 90%: claude-fable-5-1_max scored 90.2%, 81 days after the benchmark came out
  12. qwen3.8-max-0902 took first place in WebDev, from claude-opus-5-max