Last Translation Benchmark: leaderboard
Translation cases with human-authored verification rules. This board reports the LTBv1-eval blind track: models translate without access to privileged rules or reference answers, and a verifier judges whether each translation passes. The source page identifies Gemini 3.1 Pro as its verifier. Scores retain the published API pass fraction as a percentage; the source table displays two decimals. The human-contributor reference is not an AI model.
Metric: Verified Translations (%). Source: last-translation-benchmark.vilda.net. Saturation forecast: Around October 2027. 50 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-6 Astra | 44.02 | #6 |
| 2 | Gemini 4 Argon | 40.94 | #8 |
| 3 | Gemini 3.8 Flash | 39.85 | #23 |
| 4 | Gemini 3.1 Pro (Preview) | 39.3 | #59 |
| 5 | Claude Fable 5.1 | 32.93 | #9 |
| 6 | GPT-5.6 Sol | 28.1 | #18 |
| 7 | Muse Spark 1.3 | 18.44 | #27 |
| 8 | GPT-5.6 Luna | 15.37 | #72 |
| 9 | DeepSeek V4.1 Flash | 13.83 | #33 |
| 10 | Qwen 3.8 Max (0902) | 12.51 | #28 |
| 11 | GPT-5.6 Terra | 12.18 | #40 |
| 12 | Gemini 3.5 Flash Lite | 11.86 | #158 |
| 13 | Kimi K3 | 11.75 | #30 |
| 14 | DeepSeek V4 Pro | 10.76 | #106 |
| 15 | Qwen 3.7 Plus | 10.21 | #93 |
No result here: #4 Claude Opus 5.5, #7 GPT-6.1 Sol, #10 Muse Wonder Full.
Interactive version: theaggregate.ai/benchmark?slug=last-translation-benchmark · How It Works · Data refreshed daily, snapshot 2026-10-23.