In 2026, Claude Opus 4.8 tops our intelligence index at 55.7, just ahead of OpenAI’s GPT-5.5 (54.8). Google’s Gemini 3.5 Flash (50.2) and open-weight GLM-5.2 (51.1) sit close behind, with Gemini 3.1 Pro (46.5) and DeepSeek V4 Pro (44.3) rounding out the frontier. The single “best” model depends on how you weigh intelligence against price and speed.
The Intelligence column is a composite 0-100 score based on the Artificial Analysis Intelligence Index v4.1, which blends nine demanding evaluations spanning reasoning, coding, agentic tool use, and scientific knowledge. Higher is smarter — but a two-point gap at the very top rarely changes real-world output quality, so treat clusters of models as roughly equivalent rather than reading tiny differences as decisive.
Read intelligence alongside price and speed. A frontier model like Opus 4.8 costs far more per token than DeepSeek V4 Flash (40.3) or GLM-5.2, which deliver 70-90% of the intelligence at a fraction of the cost. For high-volume or latency-sensitive work, a cheaper, faster model usually wins; for the hardest reasoning and agentic tasks, the top of the table earns its premium. Sort by the metric that matches your use case.
| # | Modell ↕ | Entwickler | Intelligence ↕ | Kontext ↕ | Eingabe $/1 Mio. ↕ | Ausgabe $/1 Mio. ↕ | Open weights |
|---|---|---|---|---|---|---|---|
| 1 | Claude Opus 4.8 | Anthropic | 55.7 | 1 Mio. | $5.00 | $25.00 | Nein |
| 2 | GPT-5.5 | OpenAI | 54.8 | 1,05 Mio. | $5.00 | $30.00 | Nein |
| 3 | GLM 5.2 | Zhipu AI | 51.1 | 1 Mio. | $1.40 | $4.40 | Ja |
| 4 | Gemini 3.5 Flash | 50.2 | 1 Mio. | $1.50 | $9.00 | Nein | |
| 5 | Claude Sonnet 4.6 | Anthropic | 47 | 1 Mio. | $3.00 | $15.00 | Nein |
| 6 | Gemini 3.1 Pro | 46.5 | 1,05 Mio. | $2.00 | $12.00 | Nein | |
| 7 | DeepSeek V4-Pro | DeepSeek | 44.3 | 1 Mio. | $0.44 | $0.87 | Ja |
| 8 | Kimi K2.7 Code | Moonshot AI | 42 | 256 K | $0.60 | $2.50 | Ja |
| 9 | DeepSeek V4-Flash | DeepSeek | 40.3 | 1 Mio. | $0.14 | $0.28 | Ja |
| 10 | Claude Haiku 4.5 | Anthropic | 37 | 200.000 | $1.00 | $5.00 | Nein |
| 11 | DeepSeek R1 | DeepSeek | 20.1 | 128K | $0.50 | $2.15 | Ja |
| 12 | Mistral Large 3 | Mistral AI | 15.9 | 256 K | $2.00 | $6.00 | Ja |
| 13 | Llama 4 Maverick | Meta | 14.3 | 1 Mio. | $0.15 | $0.60 | Ja |
| 14 | Qwen3 235B-A22B | Alibaba | 13 | 128K | $0.45 | $1.80 | Ja |
| 15 | Qwen3 32B | Alibaba | 12 | 128K | $0.08 | $0.28 | Ja |
| 16 | Llama 4 Scout | Meta | 10.0 | 10 Mio. | $0.10 | $0.30 | Ja |
| 17 | Gemma 3 27B | 7.4 | 128K | $0.08 | $0.16 | Ja | |
| 18 | Phi-4 | Microsoft | 4.9 | 16K | $0.07 | $0.14 | Ja |
| 19 | Claude Fable 5 | Anthropic | — | 1 Mio. | $10.00 | $50.00 | Nein |
| 20 | DeepSeek R1 Distill Llama 70B | DeepSeek | — | 128K | $0.80 | $0.80 | Ja |
| 21 | Gemini 2.5 Pro | Google (Google DeepMind) | — | 1M (1,048,576 tokens) | $1.25 | $10.00 | Nein |
| 22 | Gemma 3 12B | — | 128K | $0.05 | $0.15 | Ja | |
| 23 | Gemma 3 4B | — | 128K | $0.05 | $0.10 | Ja | |
| 24 | Grok 4 | xAI | — | 256,000 tokens | $3.00 | $15.00 | Nein |
| 25 | Llama 3.1 8B | Meta | — | 128K | $0.02 | $0.03 | Ja |
| 26 | Llama 3.3 70B | Meta | — | 128K | $0.10 | $0.32 | Ja |
| 27 | Mistral 7B | Mistral AI | — | 32K | $0.02 | $0.03 | Ja |
| 28 | Mistral NeMo 12B | Mistral AI | — | 128K | $0.02 | $0.04 | Ja |
| 29 | NVIDIA Nemotron 3 Nano Omni | NVIDIA | — | 256 K | — | — | Ja |
| 30 | Qwen3 14B | Alibaba | — | 128K | $0.12 | $0.24 | Ja |
| 31 | Qwen3 30B-A3B | Alibaba | — | 128K | $0.12 | $0.50 | Ja |
| 32 | Qwen3 8B | Alibaba | — | 128K | $0.04 | $0.14 | Ja |
Click any column heading to re-sort. Intelligence = a 0–100 composite of public reasoning/knowledge benchmarks; — means not yet scored. Prices are USD per 1M tokens. Updated July 2026.
Häufig gestellte Fragen (FAQ)
What is the best LLM right now?
As of mid-2026, Claude Opus 4.8 is the highest-scoring model on the Artificial Analysis Intelligence Index (55.7), narrowly ahead of GPT-5.5 (54.8). Both lead on reasoning, coding, and agentic tasks, so the practical choice between them usually comes down to price, speed, and ecosystem rather than a meaningful intelligence gap.
What is the smartest open-source model?
GLM-5.2 is currently the most intelligent open-weight model, scoring 51.1 on the intelligence index — ahead of DeepSeek V4 Pro (44.3) and DeepSeek V4 Flash (40.3). That puts the best open models within roughly four points of proprietary frontier models like GPT-5.5, while remaining free to self-host or run through low-cost APIs.
Which AI model is the best value?
DeepSeek V4 Flash and GLM-5.2 offer the strongest intelligence-per-dollar. DeepSeek V4 Flash scores 40.3 at a tiny fraction of Opus 4.8’s price, delivering roughly 70% of the top score for a small percentage of the cost. For budget-sensitive, high-volume workloads, these open models are the clear value leaders.
How is the intelligence score calculated?
Our scores mirror the Artificial Analysis Intelligence Index v4.1, a 0-100 composite of nine demanding evaluations — including Humanity’s Last Exam, GPQA Diamond, Terminal-Bench, SciCode, and agentic tool-use tasks. It measures reasoning, knowledge, coding, and agentic ability in one number, so a higher score means broadly more capable across hard, real-world tasks.
