The record, as data
Every score this site publishes, with the part that makes it checkable: the metric it is in, who measured it, whether anyone reproduced it independently, and how trustworthy the benchmark itself is. 3,011 rows, refreshed daily from the same sources the pages read.
| column | meaning |
|---|---|
| model | Canonical model name. |
| model_slug | The model's page on this site. |
| developer | The lab that publishes the model. |
| benchmark | Canonical benchmark name. |
| score | The measured value, in the benchmark's own metric. |
| metric | The metric the number is in. |
| evidence | independently_reproduced, self_reported, contradicted, or unverified — the claim-ledger status of the score. |
| evaluator | Who ran the measurement, when the source names them. |
| measured_at | The model's release date — the date the source dates the model, not the date the benchmark was run. |
| source | The source id the measurement came from. |
| benchmark_integrity_grade | The integrity grade of the benchmark the score belongs to. |
A score is only useful with its evidence attached, so the evidence column is never empty and never optimistic: a vendor's own number is self_reported, and a number nobody has run yet is unverified.
| model | benchmark | score | evidence |
|---|---|---|---|
| DeepSeek-V3.2 | ARC-AGI-2 | 4.03 | unverified |
| DeepSeek-V3.2 | ProofBench | 8 | third_party |
| DeepSeek-V3.2 | GPQA diamond | 81.14 | independently_reproduced |
| DeepSeek-V3.2 | OTIS Mock AIME 2024-2025 | 87.81 | independently_reproduced |
| DeepSeek-V3.2 | DTBench | 76 | third_party |
| DeepSeek-V3.2 | LMCA | 33.91 | third_party |
| DeepSeek-V3.2 | FrontierMath-2025-02-28-Private | 38.77 | independently_reproduced |
| DeepSeek-V3.2 | Aider polyglot | 74.2 | third_party |
Use the file freely for research, reference and reporting. Cite it as tensor.news verified model scores and link back to this page. Redistribution, and any commercial use, needs a licence — the measurements inside the file were produced by other people and their terms are not ours to waive. Ask for one at hello@tensor.news.
Every number traces to a named source, and the rules for reading them are on the methodology page. This is the same record the pages are rendered from — the file is not a separate, weaker copy.