Two pre-release codenames leaked — gemdelta and mewfour — and one checkable claim doesn't hold

Two leaks landed today naming internal checkpoints rather than product names, which makes them more useful than the usual date guess — a codename is a thing you can actually watch a catalog for.

Gemini 3.5 Pro. The claim: Google has settled on a final checkpoint and deployed it, the internal id is "gemdelta", a model tag appeared in Google's backend yesterday, A/B testing is widespread, and it "has appeared on LMArena under Gemini-3.5-Pro now". The poster adds their own read from A/B use: a big upgrade over anything Google currently serves, but "still nothing crazy".

Astra. The claim, via a second account: OpenAI is preparing to launch imminently, targeting next week. A massive new pretrain — the largest since GPT-4.5 — with a release-candidate checkpoint internally called "mewfour", limited monitored deployments underway, and ten mathematical advances already produced.

Here is what we could check, and what we found.

The arena claim does not hold on the public board. We read arena.ai's text leaderboard directly. It carries gemini-3.1-pro-preview, gemini-3-pro, gemini-3.6-flash, gemini-3.5-flash in high, medium and lite variants, and a long tail of 2.x entries. There is no Gemini-3.5-Pro on it. That is not the same as calling the claim false: arenas run models anonymously in battle mode before they have the vote volume to be listed, so a model can genuinely be in rotation and invisible on the public board. But "has appeared on LMArena" reads as something a reader could go and see, and right now they cannot.

Worth noticing why that board is easy to misread: it is full of gemini-3.5-* entries that are Flash, not Pro. Someone scanning for "3.5" finds plenty. This is the same collision that made us tighten our own pattern two days ago, when it started matching a 3.5 Flash reseller listing and would have announced the wrong model's launch.

Neither codename exists in any catalog. We checked 399 OpenRouter ids and 6,149 models.dev ids: no gemdelta, no mewfour, no gemini-3.5-pro, no astra. Both models are genuinely pre-release, which is consistent with the leaks rather than against them — it just means nothing is servable yet, so nobody outside can benchmark either one.

The Astra details line up with things we already confirmed. "Ten mathematical advances" is the Lean-verified proof set OpenAI published on August 1, which we covered; "previewed to US officials" is the Washington demo from the same week. A leak that agrees with the confirmed record on its checkable parts earns more benefit of the doubt on its unconfirmable parts than one that doesn't. It is still a leak.

What we did about it. Both codenames are now armed on the board as their own entries — mewfour alongside the existing Astra hunt rather than instead of it, and gemdelta separately from the Gemini 3.5 Pro entry, because that one only fires on a non-variant 3.5 id and would match a row called "gemdelta" not at all. This is the Zinc/Magnesium lesson from two days ago: a staging catalog surfaces the codename first and the product name later, and by then the family pattern may already be marked live and therefore silent. Verified clean against every catalog before arming, so neither can fire on something that already shipped.

And the running tally, since this is now the sixth Gemini 3.5 Pro date: June, July 17, July 24, an August 12–14 leak, "tomorrow" as of yesterday, and today. Google has confirmed only that the model is in partner testing and will be broadly available when ready.