Benchmark
MCP Atlas
Agentic benchmarks measure how well models carry out tasks that take several steps. Scores are percentages. Higher is better.
- Models and systems measured
- 46
- Scores recorded
- 95
- Sources
- 2
- IQ weight
- 1.56
Score distribution
- 71.6%
- Mean
- 75.3%
- Median
- 11.6%
- Std dev
- 57.7%
- 10th percentile
- 83.7%
- 90th percentile
- 24.6%
- Min
This benchmark has a weight of 1.56 for chat models, measured over 45 models. The weight is how much it counts toward llminfo IQ. We measure it from the data, not from the raw scores.
Leaderboard
| # | Model or system | Score |
|---|---|---|
| 1 | Muse Spark Open· LLM Stats data· Reported by the model's developer. The source relaying it did not verify it. claim | 89.2% |
| 2 | Muse Spark 1.1· LLM Stats data· Reported by the model's developer. The source relaying it did not verify it. claim | 88.1% |
| 3 | Muse Spark 1.1Meta Model API · xhigh · Scale AI MCP Atlas· Reported by Meta, Meta Superintelligence Labs· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it. | 88.1% |
| 4 | Inkling 2 Small· LLM Stats data· Reported by the model's developer. The source relaying it did not verify it. claim | 85% |
| 5 | ERNIE 5.5 Flash Preview· LLM Stats data· Reported by the model's developer. The source relaying it did not verify it. claim | 84.6% |
| 6 | Kimi K3· LLM Stats data· Reported by the model's developer. The source relaying it did not verify it. claim | 84.2% |
| 7 | Kimi K3500-task public subset · 100-turn limit · Gemini 3.1 Pro judge· Reported by Moonshot AI· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it. | 84.2% |
| 8 | Doubao Seed 2.1 Pro· LLM Stats data· Reported by the model's developer. The source relaying it did not verify it. claim | 83.8% |
| 9 | Hy4 preview· LLM Stats data· Reported by the model's developer. The source relaying it did not verify it. claim | 83.7% |
| 10 | Hy4 previewpublic subset· Reported by Tencent Hy· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it. | 83.7% |
| 11 | Gemini 3.5 Flash· LLM Stats data· Reported by the model's developer. The source relaying it did not verify it. claim | 83.6% |
| 12 | Gemini 3.5 FlashGemini 3.5 Flash Model Card · May 2026· Reported by Google DeepMind· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it. | 83.6% |
| 13 | Gemini 3.5 FlashOpus 4.8 System Card capability summary · adaptive thinking · max effort · 5 trials unless noted· Reported by Anthropic· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it. | 83.6% |
| 14 | LongCat 2.5· LLM Stats data· Reported by the model's developer. The source relaying it did not verify it. claim | 83.1% |
| 15 | Inkling 2 Mini· LLM Stats data· Reported by the model's developer. The source relaying it did not verify it. claim | 82.7% |
| 16 | Claude Opus 4.8· LLM Stats data· Reported by the model's developer. The source relaying it did not verify it. claim | 82.2% |
| 17 | Claude Opus 4.8max · Scale AI MCP Atlas· Reported by Meta, Meta Superintelligence Labs· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it. | 82.2% |
| 18 | Claude Opus 4.8Opus 4.8 System Card capability summary · adaptive thinking · max effort · 5 trials unless noted· Reported by Anthropic· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it. | 82.2% |
| 19 | Beam 1.0 501B A23B· LLM Stats data· Reported by the model's developer. The source relaying it did not verify it. claim | 80.8% |
| 20 | Doubao Seed 2.1 Turbo· LLM Stats data· Reported by the model's developer. The source relaying it did not verify it. claim | 80.3% |
| 21 | Claude Opus 4.7extended config · 256 turns / 100 tools · high effort· Reported by Anthropic· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it. | 79.7% |
| 22 | Inkling Small· LLM Stats data· Reported by the model's developer. The source relaying it did not verify it. claim | 79.6% |
| 23 | Claude Opus 4.7extended config · 256 turns / 100 tools · max effort· Reported by Anthropic· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it. | 79.5% |
| 24 | Claude Opus 4.7Gemini 3.5 Flash Model Card · May 2026· Reported by Google DeepMind· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it. | 79.1% |
| 25 | Claude Opus 4.7Opus 4.8 System Card capability summary · adaptive thinking · max effort · 5 trials unless noted· Reported by Anthropic· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it. | 79.1% |
| 26 | Hy3 (Deep Infra)· LLM Stats data· Reported by the model's developer. The source relaying it did not verify it. claim | 79.1% |
| 27 | Hy3 (Deep Infra)Hy3 official Benchmark Appendix · highest reasoning tier unless noted · public · Scale April 2026 methodology · 100 calls/task · 500 tasks· Reported by Tencent Hunyuan· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it. | 79.1% |
| 28 | Gemini 3.1 Pro PreviewGemini 3.5 Flash Model Card · May 2026· Reported by Google DeepMind· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it. | 78.2% |
| 29 | Gemini 3.1 Pro Previewhigh · Scale AI MCP Atlas· Reported by Meta, Meta Superintelligence Labs· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it. | 78.2% |
| 30 | Gemini 3.1 Pro PreviewOpus 4.8 System Card capability summary · adaptive thinking · max effort · 5 trials unless noted· Reported by Anthropic· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it. | 78.2% |
| 31 | Claude Opus 4.7· LLM Stats data· Reported by the model's developer. The source relaying it did not verify it. claim | 77.3% |
| 32 | Claude Opus 4.7refreshed April 2026 harness · leaderboard config · adaptive thinking · max effort· Reported by Anthropic· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it. | 77.3% |
| 33 | GLM 5.2· LLM Stats data· Reported by the model's developer. The source relaying it did not verify it. claim | 76.8% |
| 34 | GLM 5.2public 500-task subset · think mode · 10-minute timeout· Reported by Z.ai· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it. | 76.8% |
| 35 | Qwen3.7 Max· LLM Stats data· Reported by the model's developer. The source relaying it did not verify it. claim | 76.4% |
| 36 | Qwen3.7 MaxQwen3.7-Max release · public set · Gemini 2.5 Pro judge· Reported by Qwen· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it. | 76.4% |
| 37 | Inkling FP4· LLM Stats data· Reported by the model's developer. The source relaying it did not verify it. claim | 76% |
| 38 | Kimi K2.7 Code· LLM Stats data· Reported by the model's developer. The source relaying it did not verify it. claim | 76% |
| 39 | Kimi K2.7 CodeKimi K2.7 Code official table · Kimi Code CLI · thinking + preserve_thinking · temp 1.0 · top_p .95 · 262,144 context · 100 tool-call budget · 32K max tokens/step · avg@3· Reported by Moonshot AI· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it. | 76% |
| 40 | Ling 3.0 Flash SanteLing-3.0-flash-Sante official evaluation image · highest reasoning tier · temperature 1.0 / top_p 0.95 unless otherwise stated · exact API snapshot and run count not published · public benchmark label · task count and harness not stated in card· Reported by Ant inclusionAI· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it. | 75.6% |
| 41 | Muse Glimmer 30B· LLM Stats data· Reported by the model's developer. The source relaying it did not verify it. claim | 75.5% |
| 42 | Muse Glimmer 30Bhigh reasoning· Reported by Meta· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it. | 75.5% |
| 43 | GPT 5.5· LLM Stats data· Reported by the model's developer. The source relaying it did not verify it. claim | 75.3% |
| 44 | GPT 5.5Gemini 3.5 Flash Model Card · May 2026· Reported by Google DeepMind· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it. | 75.3% |
| 45 | GPT 5.5Opus 4.8 System Card capability summary · adaptive thinking · max effort · 5 trials unless noted· Reported by Anthropic· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it. | 75.3% |
| 46 | GPT 5.5Scale AI · April 2026 update · xhigh· Reported by OpenAI· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it. | 75.3% |
| 47 | GPT 5.5xhigh · Scale AI MCP Atlas· Reported by Meta, Meta Superintelligence Labs· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it. | 75.3% |
| 48 | MiniMax M3· LLM Stats data· Reported by the model's developer. The source relaying it did not verify it. claim | 74.2% |
| 49 | MiniMax M3official model-page overview figure· Reported by MiniMax· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it. | 74.2% |
| 50 | MiniMax M3public set · Gemini 2.5 Pro grader· Reported by MiniMax Research· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it. | 74.2% |
Showing 1–50 of 94 results.
Not ranked
The source marks these results as descriptive, not capability scores, so we list them by name instead of ranking them.
| Model or system | Score |
|---|---|
| Muse Spark (version unspecified)Meta Model API · highest effort · version unspecified · Scale AI MCP Atlas· Reported by Meta, Meta Superintelligence Labs· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it. | 82.2% |
Data credits: Data by LLM Stats · BenchBoard by SYuan03 (github.com/SYuan03/benchboard). MIT covers its code and data organization; no results licence is inferred. Numeric relays retain the original stating-source links.