A codename is the internal project name a lab assigns a model while it is still in development, before any marketing name exists. Codenames let teams ship, benchmark, and route a model through their stack without committing to a public brand — and because they appear in public-facing infrastructure, they surface weeks ahead of the official announcement. Reading them well turns rumour into an early, structured signal.
What is an AI model codename?
An AI model codename is the working label a lab attaches to a model in progress, distinct from its eventual public name. Labs use codenames for the same reason any large engineering org does: to reference a moving target internally before the product, the version number, and the positioning are locked. The codename is stable; the public name is a marketing decision made later, sometimes days before launch.
In Next AI Drop copy and on the live strip, codenames render bracketed and uppercase in monospace — ⟨GOLDFISH⟩, ⟨EMBER-ALPHA⟩, ⟨BEHEMOTH-2⟩ — so a coined internal name never gets confused with a shipped product name. The convention is defined in full as codename defined in the glossary, the page that owns that term across the site.
A codename is not a confirmation. It tells you a model exists and is moving through a pipeline; it does not tell you the final name, the ship date, or the price. That gap between "seen" and "shipped" is exactly what a forecast has to close.
Where do codenames leak from?
Codenames leak from the public-facing surfaces a model has to touch before it launches. A model cannot be tested, priced, or previewed without leaving traces in systems that developers and journalists can observe — and those traces carry the internal name long before the brand team picks a public one. None of this requires inside access; it is all on the open web.
The recurring public surfaces are:
- Public model-catalog listings. New or unannounced entries show up in provider and aggregator catalogs, sometimes under the raw codename or a thinly disguised slug.
- Preview and gated SKUs. Cloud platforms and developer consoles expose preview tiers, model IDs, and pricing-index rows for models that have no public page yet.
- Developer and exec chatter. Engineers, founders, and product leads reference codenames in posts, talks, and replies — sometimes as a wink, sometimes by accident.
- Build artifacts and changelogs. SDK release notes, config keys, and routing identifiers occasionally name a model before its launch post does.
The common thread is that each surface is publicly observable and independently checkable. We read what is in the open — official changelogs, public posts on X, public model-catalog listings — and treat a codename as a lead to verify, not a fact to repeat.
How do you map a codename to a real model?
You map a codename to a real model by triangulating four signals: provider, capability hints, naming lineage, and timing against the lab's release cadence. No single clue is conclusive — a slug in a catalog could be a test artifact — but where all four converge on the same provider and product slot, the mapping becomes credible enough to publish with a confidence stage.
- Provider. Which lab's infrastructure, account, or catalog surfaced the name? This usually fixes the vendor immediately.
- Capability hints. Context length, modality flags, parameter-class rumours, and benchmark leaks suggest where the model sits in a lineup.
- Naming lineage. Labs reuse internal naming families. A codename that rhymes with a previous one often signals the same product line, one tier up.
- Timing. Map the sighting against the provider's historical cadence. A codename that appears right when a flagship is overdue is a stronger lead than one with no deadline pressure.
Illustrative codename mappings
The mappings below are illustrative examples of the method, mirroring the live strip on the homepage — not live data, and not claims about any unannounced product. Each pairs a coined internal name with the public model it is most plausibly headed toward, then a credibility verdict in the same vocabulary the site uses everywhere: Confirmed, Rumor, or Market.
| Codename | Likely public model | Verdict |
|---|---|---|
| ⟨MYTHOS⟩ | Claude 5 | Rumor |
| ⟨GOLDFISH⟩ | Gemini 3.5 Pro | Rumor |
| ⟨EMBER-ALPHA⟩ | GPT-5.6 | Market |
| ⟨BEHEMOTH-2⟩ | Llama 5 | Rumor |
A verdict of Rumor means the mapping rests on chatter and lineage but lacks a confirming surface; Market means prediction-market activity on Polymarket or Kalshi corroborates the timing; Confirmed is reserved for an official source. The path from one to the next is its own subject — see how a codename becomes confirmed.
Why is a codename an early signal?
A codename is an early signal because its appearance moves a model from invisible to observable, which raises intel recency — one of the inputs to a release forecast. When a hot codename starts showing up in public catalog listings and developer chatter, it is concrete evidence that a model has reached a shippable stage, and that recency feeds directly into how ready a drop looks.
Concretely, intel recency is the 25% intel term in Drop Readiness, the single score Next AI Drop assigns each forecast: DR = .45 odds + .25 intel + .20 deadline + .10 volume, with stages of imminent, expected, or rumoured, refreshed hourly. A fresh, well-mapped codename lifts the intel input and can nudge a release from rumoured toward expected. The full computation lives on how codenames feed Drop Readiness, which we link to rather than restate here.
A "Market" verdict means prediction-market odds on Polymarket or Kalshi line up with a codename's timing. Those odds are a forecasting signal — the fastest public read on when something ships — not a tip, a bet, or financial advice. Next AI Drop synthesizes that signal; it does not take stakes.
How to watch codenames live
Watch codenames as they surface on the codename strip on the Next AI Drop homepage, where each tracked internal name is mapped to its likely public model and tagged with a confidence stage. The strip updates as new sightings come in, so a name that starts as rumoured can be re-tagged as corroborating evidence accumulates — the live view of the method this page describes.
Start with the live codename watch to see the current map, and check the public surfaces we read to understand what counts as evidence. When you spot a codename in the wild, run it through the four-signal method above before treating it as anything more than a lead.