Four September launches, revisited with the evidence now available

Claude Opus 5.5, GPT-6 Sol, Grok 4.7 and MiMo-V2.6-Pro carry catalogue release timestamps on September 21 and 22. A draft written around their arrival described a single-board launch race. The retained September 30 file now supports a more useful question: what evidence is available for those products today, and how should it affect a decision?

We do not publish the old launch-day ranks as verified history. The draft did not preserve a digest-identified September 23 input, and an all-version list today is not an archive of that day's ordering. Without the earlier file, subtracting an old draft score from a new score would give a number whose baseline is not independently checked.

The four rows at the retained cutoff

Catalogue release timestamps describe the catalogue's declared release field. They do not establish announcement time or the first instant a public endpoint was callable. Nor does a source row's publication date establish when ModelCap first ingested it. Preserve those time concepts separately when auditing a launch.

Which evidence is available now

Opus has Arena Overall, the AA Intelligence Index, Arena Coding, LMArena Agent and ARC-AGI-2. MiMo Pro has Arena Overall, AA and Arena Coding. Grok has those first three plus LMArena Agent. GPT-6 Sol has six observations, including both ARC tracks, but sits in the all-version scope while GPT-6.1 Sol occupies a current row.

The retained observations alone therefore do not support describing these four as single-general-board rows at publication. Broader coverage does not guarantee a higher point estimate: it can add disagreement. Grok's admitted normalized Arena Overall score is 69.4 and AA is 84.1. MiMo Pro's are 82.8 and 84.0, much closer. Source configurations remain disclosed; these are comparisons of admitted product evidence, not proof that every source ran an identical endpoint.

Understand the support arithmetic

One full-confidence AA result contributes 80 × 0.25 = 20 support points. A full-confidence Arena Overall result can add 80 × 0.75 = 60. Specialist observations add their own smaller contributions, scaled by confidence. This explains why a row with a high score can initially show 20% support and later broaden substantially.

At this cutoff, Opus has 65.4% support, MiMo Pro 70.8%, Grok 73.2% and archived Sol 77.2%. These figures describe current admitted coverage. They are not percentages of successful tasks and are not probabilities that a rank is correct.

The current leader, Sonnet 5.5, has only the AA observation and 20% support. Its leading position and wide 72.4–100 interval answer different questions: the score orders the point estimate; the interval exposes limited information. The broader-evidence price frontier excludes it under the measured 60% filter.

Succession is not score inflation

GPT-6 Sol's all-version rank is 40 at 77.7. GPT-6.1 Sol's current row is third at 88.2, with AA and both ARC tracks. That is what this cutoff says; it does not retrospectively establish GPT-6 Sol's launch placement or a task-level improvement between the products. Current-family representation and all-version ordering are separate catalogue scopes.

When comparing a current successor with an archived model, use the same snapshot, preserve both scopes, and inspect the configurations and shared benchmark rows. A newer version number cannot substitute for independent evidence. A single-board successor can also retain less coverage than its predecessor while occupying the current product seat.

What to retain for the next launch

• Save the model and evidence exports with generatedAt and input hashes at the time you cite a placement.
• Preserve the original estimate and interval separately from later score revisions and outcome records.
• Record the source publication time and the earliest retained receipt showing public availability; do not infer either from catalogue timestamps.
• Revisit the product when additional general or specialist evidence is admitted, explaining which inputs changed before interpreting the rank movement.

This follow-up replaces an unverifiable historical baseline with a checked current evidence map. It can help shortlist these products today. A true launch-to-follow-up score comparison still requires the missing preserved launch snapshot.