Benchmark evidence · last saved source review 12 Oct 2026
No verified Union Alpha benchmark score
The available figures are provider-published Pareto claims. Neither former Union Alpha alias is pinned to a Pareto version, configuration and date. No inspected immutable raw per-task result package or independent reproduction establishes a Union Alpha benchmark.
| Provider material and date | Published claim | What it does not establish |
|---|---|---|
| Pareto 26.9 model card, recorded in R5 on Sep 24 | Headline slate: DeepSWE 74; Terminal-Bench 4.0 51; MMMU-Pro 78; HLE no-tools 49; ArXivMath 88. Preliminary DeepSWE point: Sep 20 run, n=30, 70.0%, $0.29/task. Provider card · R5-S06 | Not a former-alias or 26.10 result. Score method, denominator and cost need raw-run verification. |
| Separate 26.10 Preview provider chart, observed Oct 9; run dated Oct 1 | Preliminary DeepSWE point: 69.9%, $0.24/task, shown against the 26.9 Sep 20 70.0%, $0.29/task point. Provider card · O-S06 | Not an independent result, immutable raw package, Union Alpha score or evidence that 26.10 replaced older paid Pareto. |
pareto-evals method repository | Configurable evaluation code and documented local output. | Method code is not a fixed-route result. On Oct 12 the visible root had no results/ directory; complete direct tree, tags and releases inventories were uninspectable. Global artifact absence was not proved. U-S01–03 |
Do not rank Union Alpha from these figures. A qualifying comparison needs a fixed route/version/date, task set and sample count, harness/settings, prompts, retries, controls, raw outputs and cost method.
Read attribution scope, the R5 report and the Oct 9 26.10 clarification.