Agents · Benchmark profile
A benchmark for multimodal agents performing tasks in real computer operating systems.
Data verified 15 July 2026 · Methodology 1.2.0
Benchmark score on OSWorld
Claude Opus 4.5 leads at 66.3%, followed by GPT-5.6 Sol (62.6%) and GPT-5.6 Terra (50.2%).
Visual analysis
Three views of the same sourced leaderboard: model placement, the score curve, and descriptive provider averages. These visuals never alter ranking calculations.
11 ranked models · higher is better
| Model | Rank | Provider | Score |
|---|---|---|---|
| Claude Opus 4.5 | #1 | Anthropic | 66.3% |
| GPT-5.6 Sol | #2 | OpenAI | 62.6% |
| GPT-5.6 Terra | #3 | OpenAI | 50.2% |
| GPT-5.6 Luna | #4 | OpenAI | 45.6% |
| Claude Opus 4.8 | #5 | Anthropic | 20.6% |
| Muse Spark 1.1 | #6 | Meta | 14.2% |
| Claude Opus 4.7 | #7 | Anthropic | 13.9% |
| GPT-5.5 | #8 | OpenAI | 13% |
| Claude Sonnet 4.6 | #9 | Anthropic | 8.3% |
| MiniMax M3 | #10 | MiniMax | 4.6% |
| Qwen3.7-Plus | #11 | Alibaba Cloud | 2.8% |
Shows how quickly scores separate down the field
| Model | Rank | Provider | Score |
|---|---|---|---|
| Claude Opus 4.5 | #1 | Anthropic | 66.3% |
| GPT-5.6 Sol | #2 | OpenAI | 62.6% |
| GPT-5.6 Terra | #3 | OpenAI | 50.2% |
| GPT-5.6 Luna | #4 | OpenAI | 45.6% |
| Claude Opus 4.8 | #5 | Anthropic | 20.6% |
| Muse Spark 1.1 | #6 | Meta | 14.2% |
| Claude Opus 4.7 | #7 | Anthropic | 13.9% |
| GPT-5.5 | #8 | OpenAI | 13% |
| Claude Sonnet 4.6 | #9 | Anthropic | 8.3% |
| MiniMax M3 | #10 | MiniMax | 4.6% |
| Qwen3.7-Plus | #11 | Alibaba Cloud | 2.8% |
Average within the visible cohort · descriptive only
| Provider | Models | Average score |
|---|---|---|
| OpenAI | 4 | 42.9% |
| Anthropic | 4 | 27.3% |
| Meta | 1 | 14.2% |
| MiniMax | 1 | 4.6% |
| Alibaba Cloud | 1 | 2.8% |
One best score per model · higher is better
| Rank | Model | Provider | License | Score |
|---|---|---|---|---|
| #1 | Claude Opus 4.5 claude-opus-4-5 | Anthropic | closed | 66.3% |
| #2 | GPT-5.6 Sol gpt-5.6-sol | OpenAI | closed | 62.6% |
| #3 | GPT-5.6 Terra gpt-5.6-terra | OpenAI | closed | 50.2% |
| #4 | GPT-5.6 Luna gpt-5.6-luna | OpenAI | closed | 45.6% |
| #5 | Claude Opus 4.8 claude-opus-4-8 | Anthropic | closed | 20.6% |
| #6 | Muse Spark 1.1 muse-spark-1-1 | Meta | closed | 14.2% |
| #7 | Claude Opus 4.7 claude-opus-4-7 | Anthropic | closed | 13.9% |
| #8 | GPT-5.5 gpt-5.5 | OpenAI | closed | 13% |
| #9 | Claude Sonnet 4.6 claude-sonnet-4-6 | Anthropic | closed | 8.3% |
| #10 | MiniMax M3 MiniMax-M3 | MiniMax | — | 4.6% |
| #11 | Qwen3.7-Plus qwen3.7-plus | Alibaba Cloud | — | 2.8% |
The top of this snapshot is led by Claude Opus 4.5 at 66.3%; third place is 16.1 points behind. The top-10 spread is 61.7 points.
About OSWorld
A benchmark for multimodal agents performing tasks in real computer operating systems. Results stay tied to the exact model variant and evaluation system. Multiple systems for the same model use the best published score on this page; overall Lumina scoring uses the median of ranking-eligible rows.
Open benchmark source ↗FAQ
A benchmark for multimodal agents performing tasks in real computer operating systems.
Claude Opus 4.5 by Anthropic currently leads with 66.3%.
11 models in the LuminaBench cohort have a qualifying score on this benchmark.
Yes. This family is ranking-weighted at 1.6% of its capability category. Claude Opus 4.5 is overall #30.
Related
