llmrun
Coding

LiveBench 2026 Long-Horizon Leaderboard

LiveBench 2026 Long-Horizon measures multi-step tasks that a model has to carry through many steps to finish. It belongs to the 2026 LiveBench release (first snapshot September 2, 2026), a redesigned LiveBench in which each of seven categories is scored by a single task and questions are refreshed to limit training-set contamination. It is a different test from the original LiveBench boards, so the scores are not comparable.

Source: livebench15 open models ranked+33 proprietaryData through Sep 2026

Open modelsAll models

All models ranked on LiveBench 2026 Long-Horizon

Proprietary / closed models are shown dimmed — you can't run them locally, but they show where the open field stands.

#ModelScoreVRAM @ Q4
1Claude Opus 5.5 · proprietary
85.0
API only
2GPT 6 Astra · proprietary
82.4
API only
3Claude Fable 5.1 · proprietary
80.8
API only
4Claude Sonnet 5.5 · proprietary
78.8
API only
5GPT 6.1 Sol · proprietary
73.0
API only
6Gemini 4 Argon · proprietary
69.9
API only
7GPT 6 Sol · proprietary
66.9
API only
8Claude Opus 5 · proprietary
62.1
API only
9Claude Haiku 5.5 · proprietary
56.6
API only
10Claude Fable 5 · proprietary
53.2
API only
11Muse Wonder 1.0 · proprietary
53.2
API only
12Grok 4.7 · proprietary
50.1
API only
13GPT 5.6 Sol · proprietary
48.5
API only
14GPT 5.6 Terra · proprietary
48.4
API only
15Claude Opus 4.8 · proprietary
48.3
API only
16GPT 6 Luna · proprietary
48.1
API only
17MiMo V2.6 Pro RL · 1020B
47.5
~613 GB
18GPT 5.5 · proprietary
44.6
API only
19DeepSeek V4.1 Flash · 763.2B
43.8
~458 GB
20Kimi K3 · 2779.9B
42.8
~1674 GB
21Qwen3.8 Max · proprietary
42.3
API only
22Grok 4.6 · proprietary
40.7
API only
23Muse Spark 1.3 · proprietary
40.4
API only
24Claude Sonnet 5 · proprietary
39.0
API only
25DeepSeek V4 Flash Vision Exp · 304.6B
39.0
~183 GB
26Qwen3.8 Flash · proprietary
37.1
API only
27GLM 5.3 · 753.3B
35.9
~456 GB
28GPT 5.6 Luna · proprietary
30.4
API only
29Gemini 3.8 Flash · proprietary
30.1
API only
30GLM 5.3 Flash · 321.3B
28.4
~195 GB
31Gemini 3.7 Flash · proprietary
28.3
API only
32Grok 4.5 · proprietary
27.9
API only
33Qwen3.8 27B · 27.8B
27.9
~17 GB
34Step 5 · proprietary
27.5
API only
35Smaug Flash · 304.2B
27.1
~609 GB
36Gemini 3.1 Pro · proprietary
23.9
API only
37Claude Sonnet 4.6 · proprietary
23.1
API only
38Muse Spark 1.2 · proprietary
21.3
API only
39Muse Spark Open 816B A42B · 822.0B
18.9
~543 GB
40Beam 1.0 · proprietary
17.3
API only
41NVIDIA Nemotron 3.5 Super VL 120B A12B BF16 · 124.5B
15.8
~82 GB
42Inkling · 952.4B
14.1
~572 GB
43Gemma 4 31B IT · 31.3B
13.3
~20 GB
44MiniMax M3 · 427.0B
12.4
~257 GB
45Mistral Large 4 · proprietary
11.8
API only
46NVIDIA Nemotron 3 Ultra 550B A55B BF16 · 560.5B
11.2
~370 GB
47Claude Haiku 4.5 · proprietary
11.1
API only
48NVIDIA Nemotron 3.5 Lightning 30B A3B BF16 · 31.6B
8.3
~19 GB

Score vs model size

Which models give the most quality for their size — the ones worth running locally.

100B1Tmodel size (log scale) →47.58.3Kimi K3 · 2.8T · 42.8GLM 5.3 · 753B · 35.9GLM 5.3 Flash · 321B · 28.4Smaug Flash · 304B · 27.1Muse Spark Open 816B A42B · 822B · 18.9NVIDIA Nemotron 3.5 Super VL 120B A12B BF16 · 124B · 15.8Inkling · 952B · 14.1Gemma 4 31B IT · 31B · 13.3MiniMax M3 · 427B · 12.4NVIDIA Nemotron 3 Ultra 550B A55B BF16 · 561B · 11.2NVIDIA Nemotron 3.5 Lightning 30B A3B BF16 · 32B · 8.3Qwen3.8 27B · 28B · 27.9Qwen3.8 27BDeepSeek V4 Flash Vision Exp · 305B · 39.0DeepSeek V4 Flash Vis…DeepSeek V4.1 Flash · 763B · 43.8DeepSeek V4.1 FlashMiMo V2.6 Pro RL · 1T · 47.5MiMo V2.6 Pro RL
Each dot is a model. Up = higher score, left = smaller (easier to run locally). The dashed line marks the efficiency frontier — the best score you can get at each size or smaller.
  • Qwen3.8 27B, 28B, score 27.9 — on the efficiency frontier (best score at its size or smaller).
  • DeepSeek V4 Flash Vision Exp, 305B, score 39.0 — on the efficiency frontier (best score at its size or smaller).
  • DeepSeek V4.1 Flash, 763B, score 43.8 — on the efficiency frontier (best score at its size or smaller).
  • MiMo V2.6 Pro RL, 1T, score 47.5 — on the efficiency frontier (best score at its size or smaller).

LiveBench 2026 Long-Horizon: frequently asked questions

What is the best open LLM on LiveBench 2026 Long-Horizon?
MiMo V2.6 Pro RL is the top open model on LiveBench 2026 Long-Horizon, scoring 47.5. Among all models tested — including proprietary ones — it ranks #17. The top model overall is Claude Opus 5.5 (Anthropic) at 85.0.
What's the best LiveBench 2026 Long-Horizon model you can run on a 24 GB GPU?
Qwen3.8 27B is the highest-scoring open model that fits in 24 GB at 4-bit quantization (~17 GB), scoring 27.9 on LiveBench 2026 Long-Horizon.
Can open models match proprietary models on LiveBench 2026 Long-Horizon?
Not quite on LiveBench 2026 Long-Horizon: the strongest proprietary model (Claude Opus 5.5) scores 85.0, ahead of the best open model (MiMo V2.6 Pro RL) at 47.5 — but you can run the open one yourself.

Scores aggregated from livebench. llmrun does not run this benchmark — see the source for methodology, or the about benchmarks for what it measures.