winner on Toolathlon
direct benchmark result, not a broad vertical composite | source row dated 2026-04-25
scored on 2026-04-25 · stale source data (35d)
latest mapped results | top 20
| # | Model | Score | Evidence | Tested |
|---|---|---|---|---|
| 1 | DeepSeek: DeepSeek V4 Pro | 52.8 | model-only independent_benchmark | 2026-04-25 |
| 2 | DeepSeek: DeepSeek V4 Flash | 48.2 | model-only independent_benchmark | 2026-04-25 |
| 3 | Anthropic: Claude Sonnet 4.6 | 44.8 | model-only independent_benchmark | 2026-02-23 |
| 4 | Anthropic: Claude Opus 4.5 | 43.5 | model-only independent_benchmark | 2025-11-27 |
| 5 | Anthropic: Claude Sonnet 4.5 | 38.9 | model-only independent_benchmark | 2025-10-28 |
| 6 | Anthropic: Claude Haiku 4.5 | 26.2 | model-only independent_benchmark | 2025-10-28 |
| 7 | Google: Gemini 2.5 Pro | 10.5 | model-only independent_benchmark | 2025-10-28 |
| 8 | Google: Gemini 2.5 Flash | 3.7 | model-only independent_benchmark | 2025-10-28 |
what this result means
Toolathlon pass@1 on multi-tool agent tasks.
This benchmark contributes direct public evidence. Read its scope before generalizing the result.
A win here is a win on Toolathlon. Broad task pages require independent corroboration before naming a general winner.
source record
category: tool_use
metric: accuracy
matched models: 8
latest source date: 2026-04-25
direction: higher is better