Agent Arena
Signals from real agent sessions on Arena (task success, user feedback, tool errors) turned into an estimated net improvement per orchestrator model, with 95% confidence intervals.
- Unit
- net improvement
- Results
- 53 results48 Tokscale models matched, 51 in the source
- Last update
- Oct 19, 2026Newest publish date
- Retrieved
- Oct 21, 2026Source data through Oct 19, 2026
Conditions. Arena's Agent Mode harness. Reasoning effort is part of each entry's name.
Attribution. LMArena leaderboard dataset (lmarena-ai/leaderboard-dataset on Hugging Face), CC BY 4.0
Results
One row per model: its best reported result. The rank is the position in this list, by score; the source's own ranking can differ.
Tokscale usage counts accounts that used the model in public Tokscale submissions, Oct 14 – Nov 12, 2026. It is shown next to the score and does not explain it.
| Rank | Model | Score (net improvement) | Conditions | Date | Tokscale usage (30 days) | Source link |
|---|---|---|---|---|---|---|
| 1 | Claude Opus 5.5Listed as Claude Opus 5.5 (Max) | +0.20Confidence interval 0.18 to 0.2311,182 sessions | Max effort · Arena Agent Mode | Oct 19, 2026Published | 709active users | Source for Claude Opus 5.5 (opens in a new tab) |
| 2 | GPT-6 AstraListed as GPT-6 Astra (Max) | +0.15Confidence interval 0.13 to 0.1713,792 sessions | Max effort · Arena Agent Mode | Oct 19, 2026Published | 562active users | Source for GPT-6 Astra (opens in a new tab) |
| 3 | Claude Fable 5.1Listed as Claude Fable 5.1 (Max) | +0.14Confidence interval 0.12 to 0.1616,607 sessions | Max effort · Arena Agent Mode | Oct 19, 2026Published | 361active users | Source for Claude Fable 5.1 (opens in a new tab) |
| 4 | GPT-6.1 SolListed as GPT-6.1 Sol (Max) | +0.14Confidence interval 0.12 to 0.1514,746 sessions | Max effort · Arena Agent Mode | Oct 19, 2026Published | 659active users | Source for GPT-6.1 Sol (opens in a new tab) |
| 5 | Claude Sonnet 5.5Listed as Claude Sonnet 5.5 (Max) | +0.13Confidence interval 0.10 to 0.167,769 sessions | Max effort · Arena Agent Mode | Oct 19, 2026Published | 555active users | Source for Claude Sonnet 5.5 (opens in a new tab) |
| 6 | Gemini 4 ArgonListed as Gemini 4 Argon (High) | +0.12Confidence interval 0.11 to 0.1335,505 sessions | High effort · Arena Agent Mode | Oct 19, 2026Published | 82active users | Source for Gemini 4 Argon (opens in a new tab) |
| 7 | GPT-6 SolListed as GPT-6 Sol (Max) | +0.09Confidence interval 0.07 to 0.119,246 sessions | Max effort · Arena Agent Mode | Oct 19, 2026Published | 245active users | Source for GPT-6 Sol (opens in a new tab) |
| 8 | Claude Fable 5Listed as Claude Fable 5 (High) | +0.08Confidence interval 0.07 to 0.1043,116 sessions | High effort · Arena Agent Mode | Oct 19, 2026Published | 41active users | Source for Claude Fable 5 (opens in a new tab) |
| 9 | Claude Opus 5Listed as Claude Opus 5 (Max) | +0.08Confidence interval 0.06 to 0.1024,050 sessions | Max effort · Arena Agent Mode | Oct 19, 2026Published | 163active users | Source for Claude Opus 5 (opens in a new tab) |
| 10 | GPT-5.6 SolListed as GPT-5.6 Sol (xHigh) | +0.07Confidence interval 0.05 to 0.0837,049 sessions | XHigh effort · Arena Agent Mode | Oct 19, 2026Published | 217active users | Source for GPT-5.6 Sol (opens in a new tab) |
| 11 | Claude Sonnet 5Listed as Claude Sonnet 5 (High) | +0.04Confidence interval 0.02 to 0.0633,569 sessions | High effort · Arena Agent Mode | Oct 19, 2026Published | 138active users | Source for Claude Sonnet 5 (opens in a new tab) |
| 12 | Muse Wonder 1.0Listed as Muse Wonder (Max) | +0.04Confidence interval 0.02 to 0.059,689 sessions | Max effort · Arena Agent Mode | Oct 19, 2026Published | 17active users | Source for Muse Wonder 1.0 (opens in a new tab) |
| 13 | MiMo V2.6 ProListed as MiMo-V2.6-Pro | +0.04Confidence interval 0.03 to 0.0536,102 sessions | Arena Agent Mode | Oct 19, 2026Published | 39active users | Source for MiMo V2.6 Pro (opens in a new tab) |
| 14 | Gemini 3.8 FlashListed as Gemini 3.8 Flash (High) | +0.04Confidence interval 0.03 to 0.0538,204 sessions | High effort · Arena Agent Mode | Oct 19, 2026Published | 66active users | Source for Gemini 3.8 Flash (opens in a new tab) |
| 15 | Muse Spark 1.3Listed as Muse Spark 1.3 (Max) | +0.03Confidence interval 0.02 to 0.0474,953 sessions | Max effort · Arena Agent Mode | Oct 19, 2026Published | 21active users | Source for Muse Spark 1.3 (opens in a new tab) |
| 16 | Grok 4.7Listed as Grok 4.7 (xHigh) | +0.03Confidence interval 0.02 to 0.0515,665 sessions | XHigh effort · Arena Agent Mode | Oct 19, 2026Published | 131active users | Source for Grok 4.7 (opens in a new tab) |
| 17 | GPT-5.5Listed as GPT-5.5 (xHigh) | +0.03Confidence interval 0.02 to 0.0454,351 sessions | XHigh effort · Arena Agent Mode | Oct 19, 2026Published | 30active users | Source for GPT-5.5 (opens in a new tab) |
| 18 | GPT-6 LunaListed as GPT-6 Luna (Max) | +0.03Confidence interval 0.02 to 0.03182,171 sessions | Max effort · Arena Agent Mode | Oct 19, 2026Published | 432active users | Source for GPT-6 Luna (opens in a new tab) |
| 19 | DeepSeek V4.1 FlashListed as DeepSeek-V4.1-Flash-20260910 (Max) | +0.03Confidence interval 0.02 to 0.03255,742 sessions | Max effort · Arena Agent Mode | Oct 19, 2026Published | 225active users | Source for DeepSeek V4.1 Flash (opens in a new tab) |
| 20 | MiMo V2.6 FlashListed as MiMo-V2.6-Flash | +0.03Confidence interval 0.01 to 0.0420,935 sessions | Arena Agent Mode | Oct 19, 2026Published | 23active users | Source for MiMo V2.6 Flash (opens in a new tab) |
| 21 | Hy4 PreviewListed as Hy4 preview | +0.02Confidence interval 0.02 to 0.0388,908 sessions | Arena Agent Mode | Oct 19, 2026Published | 21active users | Source for Hy4 Preview (opens in a new tab) |
| 22 | GLM-5.2Listed as GLM-5.2 (Max) | +0.02Confidence interval 0.01 to 0.0388,576 sessions | Max effort · Arena Agent Mode | Oct 19, 2026Published | 13active users | Source for GLM-5.2 (opens in a new tab) |
| 23 | Kimi K3Listed as Kimi K3 (Max) | +0.02Confidence interval 0.01 to 0.03136,966 sessions | Max effort · Arena Agent Mode | Oct 19, 2026Published | 56active users | Source for Kimi K3 (opens in a new tab) |
| 24 | Qwen3.8 MaxListed as Qwen3.8-Max | +0.02Confidence interval 0.01 to 0.0271,737 sessions | Arena Agent Mode | Oct 19, 2026Published | 16active users | Source for Qwen3.8 Max (opens in a new tab) |
| 25 | Claude Haiku 5.5Listed as Claude Haiku 5.5 (High) | +0.01Confidence interval 0.00 to 0.0226,311 sessions | High effort · Arena Agent Mode | Oct 19, 2026Published | 477active users | Source for Claude Haiku 5.5 (opens in a new tab) |
| 26 | GLM-5.3Listed as GLM-5.3 (Max) | +0.01Confidence interval 0.00 to 0.0287,071 sessions | Max effort · Arena Agent Mode | Oct 19, 2026Published | 138active users | Source for GLM-5.3 (opens in a new tab) |
| 27 | GPT-5.4Listed as GPT-5.4 (High) | +0.01Confidence interval 0.00 to 0.0281,847 sessions | High effort · Arena Agent Mode | Oct 19, 2026Published | 1active users | Source for GPT-5.4 (opens in a new tab) |
| 28 | Grok 4.6Listed as Grok 4.6 (xHigh) | +0.01Confidence interval -0.00 to 0.0229,247 sessions | XHigh effort · Arena Agent Mode | Oct 19, 2026Published | 41active users | Source for Grok 4.6 (opens in a new tab) |
| 29 | GPT-5.6 TerraListed as GPT-5.6 Terra (xHigh) | +0.00Confidence interval -0.01 to 0.0224,677 sessions | XHigh effort · Arena Agent Mode | Oct 19, 2026Published | 119active users | Source for GPT-5.6 Terra (opens in a new tab) |
| 30 | DeepSeek V4 ProListed as DeepSeek-V4-Pro (High, 20260813) | −0.00Confidence interval -0.02 to 0.027,070 sessions | High effort · Arena Agent Mode | Oct 19, 2026Published | 51active users | Source for DeepSeek V4 Pro (opens in a new tab) |
| 31 | Gemini 3.7 FlashListed as Gemini 3.7 Flash (High) | −0.01Confidence interval -0.02 to -0.0055,513 sessions | High effort · Arena Agent Mode | Oct 19, 2026Published | 4active users | Source for Gemini 3.7 Flash (opens in a new tab) |
| 32 | GLM-5.3 FlashListed as GLM-5.3-Flash | −0.02Confidence interval -0.02 to -0.0188,684 sessions | Arena Agent Mode | Oct 19, 2026Published | 170active users | Source for GLM-5.3 Flash (opens in a new tab) |
| 33 | −0.02Confidence interval -0.03 to -0.0014,526 sessions | Arena Agent Mode | Oct 19, 2026Published | 30active users | Source for Step 5 Preview (opens in a new tab) | |
| 34 | GPT-5.6 LunaListed as GPT-5.6 Luna (xHigh) | −0.02Confidence interval -0.03 to -0.0143,317 sessions | XHigh effort · Arena Agent Mode | Oct 19, 2026Published | 158active users | Source for GPT-5.6 Luna (opens in a new tab) |
| 35 | Qwen3.8 Flash NextListed as Qwen3.8-Flash-Next | −0.03Confidence interval -0.04 to -0.02101,114 sessions | Arena Agent Mode | Oct 19, 2026Published | 6active users | Source for Qwen3.8 Flash Next (opens in a new tab) |
| 36 | Qwen3.8 27BListed as Qwen3.8-27B | −0.04Confidence interval -0.05 to -0.0361,114 sessions | Arena Agent Mode | Oct 19, 2026Published | 6active users | Source for Qwen3.8 27B (opens in a new tab) |
| 37 | −0.04Confidence interval -0.05 to -0.0387,422 sessions | Arena Agent Mode | Oct 19, 2026Published | 5active users | Source for Gemini 3.1 Pro Preview (opens in a new tab) | |
| 38 | −0.04Confidence interval -0.06 to -0.0314,035 sessions | Arena Agent Mode | Oct 19, 2026Published | 4active users | Source for Mistral Large 4 (opens in a new tab) | |
| 39 | −0.05Confidence interval -0.07 to -0.0211,972 sessions | Arena Agent Mode | Oct 19, 2026Published | 1active users | Source for Beam 1.0 501B A23B (opens in a new tab) | |
| 40 | −0.06Confidence interval -0.08 to -0.0531,609 sessions | Arena Agent Mode | Oct 19, 2026Published | 18active users | Source for Hy3 (opens in a new tab) | |
| 41 | Qwen3.7 MaxListed as Qwen3.7-Max | −0.06Confidence interval -0.08 to -0.0542,377 sessions | Arena Agent Mode | Oct 19, 2026Published | 3active users | Source for Qwen3.7 Max (opens in a new tab) |
| 42 | −0.07Confidence interval -0.07 to -0.06146,559 sessions | Arena Agent Mode | Oct 19, 2026Published | No Tokscale usage in the last 30 days | Source for Muse Spark 1.1 (opens in a new tab) | |
| 43 | Qwen3.7 PlusListed as Qwen3.7-Plus | −0.07Confidence interval -0.08 to -0.0523,630 sessions | Arena Agent Mode | Oct 19, 2026Published | 3active users | Source for Qwen3.7 Plus (opens in a new tab) |
| 44 | Muse Spark 1.2Listed as Muse Spark 1.2 (xHigh) | −0.07Confidence interval -0.08 to -0.0652,015 sessions | XHigh effort · Arena Agent Mode | Oct 19, 2026Published | No Tokscale usage in the last 30 days | Source for Muse Spark 1.2 (opens in a new tab) |
| 45 | MiMo V2.5 ProListed as MiMo-V2.5-Pro | −0.08Confidence interval -0.09 to -0.0749,148 sessions | Arena Agent Mode | Oct 19, 2026Published | 3active users | Source for MiMo V2.5 Pro (opens in a new tab) |
| 46 | MiniMax M3Listed as MiniMax-M3 | −0.09Confidence interval -0.11 to -0.0843,277 sessions | Arena Agent Mode | Oct 19, 2026Published | 26active users | Source for MiniMax M3 (opens in a new tab) |
| 47 | Nemotron 3 UltraNot matched to a Tokscale model | −0.11Confidence interval -0.16 to -0.0613,043 sessions | Arena Agent Mode | Oct 19, 2026Published | Not matched to a Tokscale model | Source for Nemotron 3 Ultra (opens in a new tab) |
| 48 | Mistral Medium 3.5Not matched to a Tokscale model | −0.12Confidence interval -0.14 to -0.1013,095 sessions | Arena Agent Mode | Oct 19, 2026Published | Not matched to a Tokscale model | Source for Mistral Medium 3.5 (opens in a new tab) |
| 49 | −0.15Confidence interval -0.16 to -0.1343,692 sessions | Arena Agent Mode | Oct 19, 2026Published | No Tokscale usage in the last 30 days | Source for Inkling (opens in a new tab) | |
| 50 | Inkling SmallListed as Inkling-Small | −0.16Confidence interval -0.18 to -0.1414,197 sessions | Arena Agent Mode | Oct 19, 2026Published | No Tokscale usage in the last 30 days | Source for Inkling Small (opens in a new tab) |
Showing 50 of 51 models
Score by release date
Each dot is one model's best result on Agent Arena, placed at the model's release date. The dashed line is the best score among models released up to that date.
- Used on Tokscale in the last 30 days
- No Tokscale usage in the last 30 days
- Best score so far
Claude Opus 5.5: +0.20 · released Sep 22, 2026 (date from models.dev) · Max effort · Arena Agent Mode · 709 active users on Tokscale in the last 30 days
47 models, released Feb 19, 2026 to Oct 13, 2026. Highest: Claude Opus 5.5, +0.20, released Sep 22, 2026 (date from models.dev). Use the left and right arrow keys to read each model in release order.−0.2−0.10.0+0.1+0.2+0.3
Release dates are from models.dev, because LMArena does not list them for this benchmark. 1 model has no known release date and is left out.
Show data
| Model | Released | Score | Tokscale users, 30 days |
|---|---|---|---|
| Gemini 4 Argon | Oct 13, 2026models.dev | +0.12High effort · Arena Agent Mode | 82 |
| Muse Wonder 1.0 | Oct 12, 2026models.dev | +0.04Max effort · Arena Agent Mode | 17 |
| Claude Haiku 5.5 | Oct 7, 2026models.dev | +0.01High effort · Arena Agent Mode | 477 |
| Mistral Large 4 | Oct 6, 2026models.dev | −0.04Arena Agent Mode | 4 |
| Beam 1.0 501B A23B | Oct 5, 2026models.dev | −0.05Arena Agent Mode | 1 |
| GPT-6.1 Sol | Sep 29, 2026models.dev | +0.14Max effort · Arena Agent Mode | 659 |
| Claude Sonnet 5.5 | Sep 28, 2026models.dev | +0.13Max effort · Arena Agent Mode | 555 |
| MiMo V2.6 Flash | Sep 22, 2026models.dev | +0.03Arena Agent Mode | 23 |
| GPT-6 Luna | Sep 22, 2026models.dev | +0.03Max effort · Arena Agent Mode | 432 |
| MiMo V2.6 Pro | Sep 22, 2026models.dev | +0.04Arena Agent Mode | 39 |
| GPT-6 Sol | Sep 22, 2026models.dev | +0.09Max effort · Arena Agent Mode | 245 |
| Claude Opus 5.5 | Sep 22, 2026models.dev | +0.20Max effort · Arena Agent Mode | 709 |
| Grok 4.7 | Sep 21, 2026models.dev | +0.03XHigh effort · Arena Agent Mode | 131 |
| Step 5 Preview | Sep 16, 2026models.dev | −0.02Arena Agent Mode | 30 |
| DeepSeek V4.1 Flash | Sep 10, 2026models.dev | +0.03Max effort · Arena Agent Mode | 225 |
| GPT-6 Astra | Sep 4, 2026models.dev | +0.15Max effort · Arena Agent Mode | 562 |
| Muse Spark 1.3 | Sep 2, 2026models.dev | +0.03Max effort · Arena Agent Mode | 21 |
| Gemini 3.8 Flash | Sep 2, 2026models.dev | +0.04High effort · Arena Agent Mode | 66 |
| Claude Fable 5.1 | Sep 1, 2026models.dev | +0.14Max effort · Arena Agent Mode | 361 |
| Hy4 Preview | Aug 28, 2026models.dev | +0.02Arena Agent Mode | 21 |
| GLM-5.3 Flash | Aug 26, 2026models.dev | −0.02Arena Agent Mode | 170 |
| Qwen3.8 27B | Aug 14, 2026models.dev | −0.04Arena Agent Mode | 6 |
| GLM-5.3 | Aug 14, 2026models.dev | +0.01Max effort · Arena Agent Mode | 138 |
| Gemini 3.7 Flash | Aug 13, 2026models.dev | −0.01High effort · Arena Agent Mode | 4 |
| Grok 4.6 | Aug 12, 2026models.dev | +0.01XHigh effort · Arena Agent Mode | 41 |
| Muse Spark 1.2 | Aug 5, 2026models.dev | −0.07XHigh effort · Arena Agent Mode | No Tokscale usage in the last 30 days |
| Qwen3.8 Max | Aug 3, 2026models.dev | +0.02Arena Agent Mode | 16 |
| Inkling Small | Jul 30, 2026models.dev | −0.16Arena Agent Mode | No Tokscale usage in the last 30 days |
| Claude Opus 5 | Jul 24, 2026models.dev | +0.08Max effort · Arena Agent Mode | 163 |
| Kimi K3 | Jul 16, 2026models.dev | +0.02Max effort · Arena Agent Mode | 56 |
| Inkling | Jul 15, 2026models.dev | −0.15Arena Agent Mode | No Tokscale usage in the last 30 days |
| GPT-5.6 Luna | Jul 9, 2026models.dev | −0.02XHigh effort · Arena Agent Mode | 158 |
| GPT-5.6 Terra | Jul 9, 2026models.dev | +0.00XHigh effort · Arena Agent Mode | 119 |
| GPT-5.6 Sol | Jul 9, 2026models.dev | +0.07XHigh effort · Arena Agent Mode | 217 |
| Hy3 | Jul 6, 2026models.dev | −0.06Arena Agent Mode | 18 |
| Claude Sonnet 5 | Jun 29, 2026models.dev | +0.04High effort · Arena Agent Mode | 138 |
| GLM-5.2 | Jun 13, 2026models.dev | +0.02Max effort · Arena Agent Mode | 13 |
| Claude Fable 5 | Jun 7, 2026models.dev | +0.08High effort · Arena Agent Mode | 41 |
| Qwen3.7 Plus | Jun 2, 2026models.dev | −0.07Arena Agent Mode | 3 |
| MiniMax M3 | Jun 1, 2026models.dev | −0.09Arena Agent Mode | 26 |
| Qwen3.7 Max | May 21, 2026models.dev | −0.06Arena Agent Mode | 3 |
| DeepSeek V4 Pro | Apr 24, 2026models.dev | −0.00High effort · Arena Agent Mode | 51 |
| GPT-5.5 | Apr 23, 2026models.dev | +0.03XHigh effort · Arena Agent Mode | 30 |
| MiMo V2.5 Pro | Apr 22, 2026models.dev | −0.08Arena Agent Mode | 3 |
| Muse Spark 1.1 | Apr 8, 2026models.dev | −0.07Arena Agent Mode | No Tokscale usage in the last 30 days |
| GPT-5.4 | Mar 5, 2026models.dev | +0.01High effort · Arena Agent Mode | 1 |
| Gemini 3.1 Pro Preview | Feb 19, 2026models.dev | −0.04Arena Agent Mode | 5 |
Method and matching
How results are run
- Arena's Agent Mode harness. Reasoning effort is part of each entry's name.
- Scores are causal effect estimates, not win rates or ratings. Compare them only within this board.
Matching to Tokscale models
- 50 of 53 results are matched to 48 Tokscale models: 18 by exact id, 31 by normalized name, 1 by a manual rule.
- No match is marked as reviewed by hand yet.
- 3 results from 3 models are not matched. They stay in the table without a Tokscale link.
- Reasoning effort, thinking budget and agent are conditions of a result. They never change which model a result belongs to.
License and attribution
- CC BY 4.0 (opens in a new tab), retrieved Oct 21, 2026.
- LMArena leaderboard dataset (lmarena-ai/leaderboard-dataset on Hugging Face), CC BY 4.0
Sources and attribution
Each score keeps its own source, unit and run conditions. Tokscale does not combine benchmarks into one score, and a missing score means the model was not evaluated, not that it scored zero.