Every benchmark leaderboard on tensor.news carried a column headed Measured. For 93% of dated score rows, the date beside a score was the day the model was released, not the day the benchmark was run. The heading has been renamed to Released, and the data dictionary and the score feed now describe the field the same way.
The underlying limitation did not change. The record still cannot date a single evaluation. What changed is that it no longer claims to.
The record's score rows carry one date field, named measured_at in the data and now rendered
under the heading Released on every model and benchmark page. The same value is also
written as the model's release_date fact. In 2,680 of the 2,881 dated rows — 93.0% — the two
are byte-identical. Not similar. The same string.
The clearest demonstration is a model that was measured many times over. GPT-5 carries 30 benchmark scores in the record: MMLU, GPQA diamond, SWE-bench verified, ARC-AGI, HellaSwag and two dozen more. Every one of the 30 rows reads 2025-08-07, the date GPT-5 shipped. The same holds for Claude Opus 4.5, whose 30 rows all read 2025-11-24, and Gemini 3.1 Pro, whose 28 rows all read 2026-02-19.
That is not what a measurement date looks like. A benchmark harness run takes hours or minutes, and different benchmarks are run at different times. A model evaluated on 30 benchmarks over the following year would show 30 different dates. This record shows one.
| Model | Score rows | Distinct dates | That date is |
|---|---|---|---|
| GPT-5 | 30 | 1 (2025-08-07) | its release date |
| Claude Opus 4.5 | 30 | 1 (2025-11-24) | its release date |
| GPT-5.5 | 29 | 1 (2026-04-23) | its release date |
| Gemini 3.1 Pro | 28 | 1 (2026-02-19) | its release date |
| GPT-4o mini | 20 | 1 (2024-07-18) | its release date |
The upstream export confirms the semantics independently. Across 2,841 rows there are 557
distinct model-version pairs, and 554 of those 557 carry exactly one date between them.
The date is a property of the model version, not of the evaluation. The three exceptions split
by version exactly as a release date would: Qwen3-30B-A3B and Qwen3-32B carry 2025-04-28
and 2025-04-29, one day apart, and Yi-6B carries 2023-11-02 while Yi-6B-Chat carries
2023-11-22.
Four surfaces changed. The column header on the benchmark leaderboard and the model page
reads Released instead of Measured, in the table header and in the sort control
alike. The /data dictionary now defines measured_at as the model's release date and says
plainly that it is not the date the benchmark was run. The score feed reads "released" where
it used to read "measured". The /about and /methodology provenance paragraphs now say
"when the model was released" instead of "when it was measured".
Nothing else moved. No integrity grade changes, because the contamination window compares a model's release date to a benchmark's release date — the semantics were already correct for that purpose. GPQA diamond's window ratio re-derives exactly: 176 of 181 rows dated after the benchmark's 2023-11-20 release, a ratio of 0.972, matching the published figure. No claim status changes. The claim ledger tracks provenance separately and correctly: 559 of the 2,993 claims in the ledger are independently reproduced, each with a named evaluator and source.
The field name measured_at stays. It is an internal identifier.
The record can tell you when a model was released and what it scored. It cannot tell you when the scoring happened. A row that arrived in this week's refresh and a row that has sat unchanged since 2023 carry the same kind of date, and both carry their model's release date. Sixty-five rows have no date at all. The record's newest dated row, 2026-10-29, belongs to Mistral Large 4 — that is when the model was released, not when it was last scored.
Fixing that needs a genuine evaluation date in the upstream export, which is a column Epoch AI does not currently publish. Until it exists, the honest label is the one the site now carries.