Six Public Signals That an AI Model Launch Is Hours Away

Someone on your team is planning around a model that has not shipped. A migration, a procurement decision, a piece of writing, a budget line. The evidence they are working from is a mix of things a stranger could check and things nobody can, and the difference is rarely marked. This reference marks it.

It lists six public signals that a model is coming or has just arrived. For each it says what the signal proves, what it does not, and the exact page or API field that returns the same answer to anyone who fetches it. Three launches from the last month are scored against the five checkable signals, and one catalog's floating aliases are tabulated in full. The worked examples are illustrations; the method is the subject.

The test is short. Can a reader who does not trust you open a public URL, read a named field or a dated sentence, and see what you saw? If yes, the claim is evidence and can be dated to the reading. If no, it is a report of evidence, and the honest description is the category it belongs to.

Two instruments do most of the work. The first is a router's public model list, which needs no key and carries fields the router documents. The second is a vendor's own pages: the launch post, the release notes, the model documentation, the deprecation table. Where a page carries no date, the Wayback Machine's CDX index gives the earliest capture that contains the sentence you are relying on. Everything below is built from those three things.

The table gives the six in one screen. The paragraphs after it give the worked example for each, with the primary it comes from.

Signal 1, the stealth test. Z.ai’s GLM-5.3-Flash documentation carries the sentence quoted above, and the Wayback Machine holds a capture of that page from August 26, 2026 that already contains it. That is a lab confirming, in its own words, that an unusually strong anonymous entry on a router can be a frontier model in disguise. It licenses nothing more. No vendor publishes a list of aliases it is running now, and no second vendor has documented the practice; we looked at xAI, Anthropic, GitHub and Meta and found no equivalent sentence.

Signal 2, the dated slug. Routers keep a permanent canonical identifier apart from the friendly one, and vendors often embed a snapshot date in it. On the day of writing, the OpenRouter entry for Grok 4.7 carried the canonical slug x-ai/grok-4.7-20260916, while the route was listed on September 21 and xAI’s announcement is dated September 21. Five days. For Claude Fable 5.1 the gap to the announcement was one day and for GLM-5.3-Flash zero. Measured against the catalog listing instead, Claude Haiku 4.5 ran 14 days and Meta’s Muse Spark 1.2 Contributor 16. One Nemotron entry embeds a date three years before its listing. The string is a label the vendor chose, and it should be cited as a string with a read date.

Signal 3, the alias flip. Catalogs and vendor APIs publish pointers that re-target when a new model ships. xAI’s own model documentation describes three forms: the bare name for the latest stable version, a -latest suffix for the newest version, and a dated name that never changes. On OpenRouter the pointer ~x-ai/grok-latest resolved to Grok 4.7 on the day of writing, though the alias entry itself dates from July 8. A flip proves the new model is live; nothing in the response records when the target last changed, so it cannot show the flip came before the announcement, and we do not claim it did.

Signal 4, the deprecation notice. GitHub’s changelog entry of September 18, 2026 retires six Copilot models on October 19 and names a suggested alternative for each: Gemini 3.8 Flash for Gemini 3.7 Flash, GPT-5.6 Sol for GPT-5.5 and GPT-5.4, GPT-5.6 Luna for the two mini models, and Grok 4.6 for Grok 4.5. Three days later xAI announced Grok 4.7. The replacement column tells you what the platform expected to be current when it wrote the notice, and it ages from that moment. The machine-readable form is OpenRouter’s expiration_date field, which it defines as the deprecation date for the model endpoint; 21 of 444 entries carried one on the day of writing.

Signal 5, route metadata. Once a route exists, its fields keep moving without a vendor post. On the day of writing the Grok 4.7 route showed a four-rung effort ladder with reasoning mandatory, a 500,000-token context, a 450,000-token maximum output, no router-side moderation, and a price override that doubles every rate above 200,000 prompt tokens. xAI’s own pricing page states the same threshold. Vercel’s AI Gateway documents a similar unauthenticated list with a released field that is separate from its own listing timestamp. A field change proves the configuration changed. It does not prove the weights changed, and a router can edit its own entry.

Signal 6, the non-evidence. A screenshot cannot be re-fetched. An unattributed insider cannot be re-asked. A price exists in exactly two places, the vendor’s pricing page and a route’s pricing object, and a number with neither has no provenance. A benchmark score without a named harness, version and effort setting is not comparable to anything; every primary we read that publishes a score also publishes its harness. Each of these becomes evidence the moment someone supplies the page or the field, and not before.

Three launches announced on or before the editorial date, each checked for the five checkable signals. "Not found" means we looked on the vendor's surfaces and did not find it, not that it does not exist. Days are counted from the signal to the vendor's dated announcement. The sixth signal is not scored, because scoring it would require citing the secondary material it describes.

The pattern across the three columns is that the signals arrive late or on the day, not early. The longest lead in the table is the 84 days between Anthropic retiring its Mythos Preview identifier and the Fable 5.1 and Mythos 5.1 launch, and that notice named a migration to Mythos 5, not to anything unreleased. The next longest is five days, and it is a slug label. Nothing here predicts a date. What the signals do is let a reader confirm, within hours and from public pages, that a launch has happened, which is the moment most planning actually needs.

For the Grok 4.7 launch specifically, our same-day post reads the vendor table and the route price. The monthly view is in our September release tracker.

One catalog, every alias, read once. OpenRouter carried 18 tilde-prefixed floating aliases on the day of writing, each with a current target and a date the alias entry itself was listed. The listing date is the alias's birthday, not its target's; an alias older than its target has flipped at least once. Each row is reproducible with a single unauthenticated request to the models endpoint, reading the alias entry's alias_target.slug field.

Two things to read off the table. First, all six aliases listed on September 11 and 14 are younger than their targets, whose own catalog entries run from July 9 to September 10 in the same response, which means they were created to point at models that already existed; an alias is not always a flip. Second, the Anthropic Opus alias points at Claude Opus 5 and was listed in April. That row is the machine-readable form of the negative finding in the next section.

The method is to state what exists, state what does not, and name the category of everything else. The current example is the one most people asked about this month.

The same discipline applies to a leaked valuation, a rumoured price change or a screenshot of a console. Our post on reading an AI valuation leak walks through the financial version. For the prices that do have a vendor page, with their check dates, our price index is the maintained record. And if your team needs a model decision made on evidence that exists rather than evidence that is expected, that is part of what we do under AI transformation.