Can it solve difficult reasoning and knowledge tasks?
Artificial Analysis + HELM
Independent evidence
Benchmark
Observatory.
No benchmark is perfectly neutral, and no single score identifies the “best” model. This page combines independent and academic sources so capability, preference, coding, cost, and speed can be considered together.
SNAPSHOTSeptember 12, 2026Source links lead to live tables
Balanced scorecard
Frontier model snapshot.
Selected leading model configurations from the Artificial Analysis leaderboard. One highest-scoring configuration per model family is shown to reduce duplicate variants.
| Model | Creator | Intelligence index | Cost / task | Output tokens / sec |
|---|---|---|---|---|
| Claude Fable 5.1 | Anthropic | 53 | $7.63 | 65 |
| GPT-6 Astra | OpenAI | 53 | $3.26 | 55 |
| Claude Opus 5 | Anthropic | 51 | $5.86 | 51 |
| Muse Spark 1.3 | Meta | 48 | $1.60 | 206 |
| GLM-5.3 | Z AI | 45 | $2.01 | 62 |
| Grok 4.6 | SpaceXAI | 44 | $1.86 | 55 |
| Kimi K3 | Kimi | 44 | $2.00 | 38 |
| Gemini 3.8 Flash | 41 | $1.24 | 269 |
Higher is better: Intelligence index and output speed.
Lower is better: Cost per task.
Important: Reasoning settings affect quality, time, and cost.
Open the complete live leaderboardMultiple lenses
What should
we measure?
A benchmark is useful only when its task resembles the decision being made. These five questions prevent one attractive number from becoming the entire story.
Do people prefer its answers in blind comparisons?
Arena
Can the complete agent resolve real repository issues?
SWE-bench
What does a completed benchmark task cost?
Artificial Analysis
How fast does output start and complete?
Artificial Analysis
Source map
Four complementary views.
Each source answers a different question. The limitations are included because transparent uncertainty is part of a trustworthy comparison.
| Source | What it measures | Why it helps | What to remember |
|---|---|---|---|
| Artificial Analysis | Composite capability, API cost, output speed, and latency | Independent testing across proprietary and open-weight models | A composite index can hide strengths and weaknesses inside individual tasks |
| Stanford HELM | Capabilities, safety, and vision-language evaluation | Academic framework with broad scenarios and transparent evaluation | Coverage is broad, but the evaluated model set and refresh rate can differ by leaderboard |
| Arena | Blind human preference across text, code, vision, image, and other arenas | Measures what people prefer in real comparisons | Preference can reward style and presentation; it is not an objective task-success measure |
| SWE-bench | Resolution of real software issues from public repositories | Domain-specific, executable coding tasks with a human-filtered Verified set | Scores depend on the full agent scaffold—not only the underlying model |
Editorial method: Company-reported benchmark claims remain news. Only results from the linked third-party tables enter this observatory, and every snapshot keeps its date, source, metric definition, and stated limitation.