llms-full.txt

# ALL Bench Leaderboard 2026 — Full Reference

> Complete model data for AI systems. See llms.txt for summary.

## LLM Rankings (42 Models)

### Flagship Models
| Model | Provider | GPQA | AIME | HLE | ARC-AGI-2 | Metacog | SWE-V | IFEval | LCB | Price(In/Out) |
|-------|----------|------|------|-----|-----------|---------|-------|--------|-----|---------------|
| GPT-5.4 | OpenAI | 92.8 | 97 | 52.1 | 73.3 | — | — | — | — | $2.50/$15 |
| GPT-5.2 | OpenAI | 93.2 | 100 | 35.4 | 52.9 | 62.76 | 80.0 | 90.5 | 80.0 | $1.75/$14 |
| GPT-5.3 Codex | OpenAI | 91.5 | 95 | 36.0 | — | — | — | — | — | $7.50/$30 |
| Claude Opus 4.6 | Anthropic | 91.3 | 100 | 40.0 | 68.8 | 56.04 | 80.8 | 93.1 | 76.0 | $5/$25 |
| Claude Sonnet 4.6 | Anthropic | 89.9 | 83 | — | 60.4 | — | 79.6 | 89.5 | — | $3/$15 |
| Gemini 3.1 Pro | Google | 94.3 | 97 | 44.4 | 77.1 | — | 80.6 | 91.0 | 80.0 | $2/$12 |
| Gemini 3 Flash | Google | 90.4 | 84 | 33.7 | — | — | 78.0 | 88.3 | — | $0.50/$3 |
| Grok 4 Heavy | xAI | 92.0 | 97 | 38.5 | 67.5 | — | — | 90.0 | — | $3/$15 |
| Kimi K2.5 | Moonshot | 87.6 | 96.1 | 44.9 | 12.1 | 68.71 | — | — | 85.0 | $0.14/$0.28 |
| DeepSeek V3.2 | DeepSeek | 82.3 | 92.8 | 25.7 | — | 60.04 | — | 91.2 | 71.6 | $0.14/$0.28 |

### Open-Source Models
| Model | Provider | MMLU-Pro | GPQA | AIME | License | Price |
|-------|----------|---------|------|------|---------|-------|
| Qwen3.5-397B | Alibaba | 84.6 | 88.1 | 96 | Apache2 | Free |
| DeepSeek R1 | DeepSeek | 79.8 | 87.3 | 97 | MIT | Free |
| Llama 4 Scout | Meta | 74.3 | 79.8 | 73 | Llama | Free |
| Llama 4 Maverick | Meta | 80.5 | 85.8 | 81 | Llama | Free |
| GLM-5 | Zhipu AI | 78.6 | 86.3 | 84 | Free | Free |
| K-EXAONE | LG AI Research | 81.8 | 75.4 | 85.3 | Prop | Prop |

## VLM Rankings (11 Flagship)
| Model | MMMU | MMMU-Pro | MathVista | Type |
|-------|------|---------|-----------|------|
| Gemini 3 Flash | 87.6 | 80.0 | — | Closed |
| Gemini 3 Pro | 87.5 | 80.0 | — | Closed |
| GPT-5.2 | 86.7 | — | — | Closed |
| Claude Opus 4.6 | — | 85.1 | — | Closed |
| GPT-5 | 84.2 | — | — | Closed |
| Gemini 3.1 Pro | — | 82.0 | — | Closed |
| InternVL3.5-241B | 77.7 | — | — | Open |
| Grok 4 Heavy | 76.5 | — | — | Closed |
| InternVL3-78B | 72.2 | — | 79.6 | Open |
| Qwen2.5-VL-72B | 70.2 | — | 74.8 | Open |
| Kimi-VL-A3B | 64.0 | 46.3 | 80.1 | Open |

## Agent Rankings (10 Models)
| Model | OSWorld | BrowseComp | Terminal-Bench | GDPval-AA |
|-------|---------|------------|----------------|-----------|
| GPT-5.4 | 75.0 | 82.7 | — | 83 |
| Claude Opus 4.6 | 72.7 | 84.0 | 74.7 | 1606 |
| Claude Sonnet 4.6 | 72.5 | — | 53.0 | 1633 |
| Gemini 3.1 Pro | — | 85.9 | 78.4 | 1317 |
| GPT-5.3 Codex | — | — | 77.3 | — |

## Generative AI Models

### Image Generation (10 Models)
GPT Image 1.5 (OpenAI) · Imagen 4 (Google) · Flux 2 Pro (BFL) · Midjourney v7 · Flux 2 Dev · Ideogram 3.0 · DALL-E 3.5 · Nano Banana 2 · SD 3.5 · Seedream 4.5

### Video Generation (10 Models)
Sora 2 (OpenAI) · Veo 3.1 (Google) · Runway Gen-4.5 · Kling 3.0 · Seedance 2.0 · Wan 2.6 · Pika 2.5 · Luma Ray3 · LTX-2 · HaiLuo AI

### Music Generation (8 Models)
Suno v4.5 · Udio v2 · Gemini Music · MusicGen Large · Stable Audio 2.0 · JASCO · Riffusion v2 · Loudme

## Benchmark Methodology

Composite Score = Avg(confirmed benchmarks) × √(N/10) where N = number of benchmarks with confirmed data out of 10 core benchmarks.

Confidence system:
- Cross-verified (✓✓): 2+ independent sources confirm the score
- Single-source (✓): One official or third-party source
- Self-reported (~): Provider claim only, not independently verified
- Null (—): No data available, never estimated or imputed

Last verified: 2026-03-08