Title card reading Gemini 4 Neon and Cobalt under the banner Sighted in code, not announced, with a Google AI Studio bundle date of 18 November 2026, the subtitle two model paths, one carousel definition, no endpoint, three chips reading source: minified JavaScript, posted by @testingcatalog and status: no model ID, and two cards — Gemini 4 Neon with 24 billing rows in the SKU catalogue, and a pale-amber outlined Gemini 4 Cobalt with none.
Guides & Insights

Gemini 4 Neon and the New "Cobalt" Name: What Google AI Studio's Code Actually Reveals

Author

jinhao.song

Date Published

Latest models · 20View all models →
Benchmarks: Artificial Analysis · updated daily
Back to all posts

The most interesting line in the G​emini 4 leak that circulated on 18 November 2026 is not a price. It is a prompt about a hospital bill. A screenshot of G​oogle AI Studio's bundled JavaScript shows a user-interface object that names two models — G​emini 4 Neon and G​emini 4 Cobalt — and attaches example prompts written for people asking about medical charges, used cars and money-handling bugs. Neither model has shipped. Neither has a model ID you can call, a documentation entry, or a changelog row. G​emini 4 Cobalt, the name nobody had seen before, does not even have a billing line in G​oogle's own cloud price catalogue, while Gemini 4 Carbon, G​emini 4 Argon and G​emini 4 Hydrogen — the three G​emini 4 models G​oogle actually documents today — do.

That gap is the story. A price row tells you someone created a billable item. A prompt carousel tells you someone decided who the model is for.

What the leaked object actually contains

The image, posted by the account @testingcatalog at 18:33 UTC on 18 November, shows a minified JavaScript object with four fields that survive the minifier intact:

The screenshot posted by @testingcatalog on 18 November 2026, showing dark-themed minified JavaScript with a carousel object containing thinkingLevel 2, enableSearchAsATool false, enableCodeExecution false, an examples call, yib carousel, and an array of models/neon and models/cobalt.

• examples: _.PNa(Lvc, !0) — a call to a function that assembles an example list, with a truthy second argument. Minifiers rename identifiers, not string literals, so the shape of the call is readable even though the function name is not.

• yib: "carousel" — a layout hint. Whatever this object is, it renders as a carousel rather than a list or a modal.

• zz: ["models/neon", "models/cobalt"] — a two-element array of model paths. This is the line that names both models, and it uses the same models/ prefix G​oogle's API uses for every model identifier.

• Kk: a => Kvc.test(a.getName()) — a predicate that runs a regular expression against a model's name. In plain terms: show this carousel only when the current model matches a pattern.

Each card in that carousel carries its own prompt and generation settings. The visible ones read thinkingLevel: 2, enableSearchAsATool: !1 and enableCodeExecution: !1 — a mid-tier thinking budget with both G​oogle Search grounding and code execution switched off. The two card titles legible in the screenshot are "Spot what goes wrong when apps handle money" and "Simulate a phantom traffic jam in JavaScript."

The post itself lists four more prompts attributed to G​emini 4 Neon: "Review my code before I ship it," "Find the errors in my hospital bill," "Should my next car be an EV or a hybrid," and a fourth that is cut off mid-sentence — "Spot the red flags in my…"

Treat all of this as a screenshot of unverified provenance. We have not seen the source file, the build hash or the page it was pulled from, and nothing here has been confirmed by G​oogle. What can be said is that the fragment is internally consistent: a carousel component, a list of model paths and a name-matching predicate are exactly the pieces such a feature would need, and the prompts are written in the register G​oogle uses for consumer-facing suggestions rather than developer documentation.

The obvious comparison is the cloud SKU catalogue, which already lists G​emini 4 Neon billing entries across text, image, video and audio tiers, plus batch, flex, priority, caching and contributor variants. Those rows are evidence that someone created billable line items, and this blog has already walked through what they imply about price. There is no new price evidence in the 18 November screenshot and no reason to re-derive it.

The carousel is a different kind of artifact. A SKU row can be created defensively, months ahead, and sit unused; the catalogue has plenty of rows like that, and the surrounding multipliers are often a better guide to whether a row is real than the row itself. Interface copy is harder to create speculatively, because someone has to decide what a model is for in order to write the example that demonstrates it. Those decisions leak positioning in a way that prices do not.

And the positioning here is unusually consistent. Of the six prompts visible in the post and the screenshot, four are consumer or small-business questions — a hospital bill, a car purchase, a money-handling bug, a physics simulation — and the two code-related ones are phrased for people who write code occasionally rather than for teams running a CI pipeline. "Review my code before I ship it" is the least technical phrasing of a code review request available. That is not how you write examples for a frontier reasoning model. It is how you write examples for a model meant to be useful to people who will never read a benchmark table.

Pair that with search and code execution being off by default and a middling thinking budget, and the profile is coherent: a general-purpose G​emini 4 tier that answers from what it already knows, quickly, without tools. That is a description of a product decision, not a capability. It does not tell you how good the model is. Nothing in a leak ever does.

Cobalt: the name with a price list of zero

G​emini 4 Cobalt is the genuinely new information in the screenshot, and it is also the part easiest to over-read.

G​oogle has been naming G​emini 4 variants after elements and simple substances: G​emini 4 Argon, G​emini 4 Carbon, G​emini 4 Hydrogen, and G​emini 4 Neon in the leak cycle. Cobalt fits that scheme exactly, which is a weak but real signal — a fabricated name would be unlikely to land on an element that extends the existing sequence without repeating one. The array also puts Neon and Cobalt together, in that order, inside a single carousel definition. If the two paths were unrelated placeholders, there is no obvious reason for them to share one component.

What Cobalt does not have, as of 19 November 2026, is anything else:

• G​oogle's AI Studio models documentation lists three G​emini 4 models — G​emini 4 Carbon, G​emini 4 Argon and G​emini 4 Hydrogen — with model codes gemini-4-carbon, gemini-4-argon and gemini-4-hydrogen. No Neon, no Cobalt.

• The G​emini API changelog has no entry mentioning either name.

• A search of the cloud SKU catalogue returns 24 distinct "gemini 4 neon" row names across text, image, video and audio, plus batch, flex, priority, caching and contributor variants. The same search for "gemini 4 cobalt" returns nothing at all.

• Hugging Face has no repository under either name from G​oogle's organisation.

• The consumer G​emini app lists G​emini 4 Argon and G​emini 4 Carbon. The Neon string that was reported there earlier in the week is no longer reproducible in the current build, which means the surface stopped showing it — not that the model was cancelled.

So the ordering is: Neon has a price ladder and no product page; Cobalt has a code reference and nothing whatsoever. If Cobalt is real, it is earlier in the pipeline than Neon. If it is a placeholder for a variant that was renamed before anyone outside G​oogle saw it, the array would look exactly like this. Both readings survive the evidence, and the honest answer is that we cannot yet distinguish them.

A two-column scoreboard headed Neon and Cobalt, the scoreboard, listing for Gemini 4 Neon: docs model ID none, changelog entry none, 24 SKU billing rows, carousel reference models/neon, four consumer-facing example prompts, callable today no; and for Gemini 4 Cobalt: no model ID, no changelog entry, zero SKU billing rows, carousel reference models/cobalt, no example prompts visible, callable today no.

What "referenced in the code" does and does not mean

It is worth being precise here, because "spotted in AI Studio" is routinely reported as though it meant "coming soon." The AI Studio front end bundles configuration for models that are gated by account, region or rollout flag. A name appearing in that bundle means the front end knows the name exists. It does not mean the back end will answer a request for it, and it certainly does not mean an announcement is scheduled.

The dates we can put on the record:

• 10 November 2026 — G​emini 4 Carbon reached general availability at an introductory $2 per million input and $10 per million output tokens through 30 November, moving to $4 and $20 afterwards.

• 12 November 2026 — G​emini 4 Hydrogen shipped at $1.50 and $7.50, adapted from Carbon and carrying the same 1,048,576-token window.

• 18 November 2026, 18:33 UTC — the AI Studio carousel screenshot was posted, naming G​emini 4 Neon and G​emini 4 Cobalt.

• 19 November 2026 — documentation, changelog, model catalogue and Hugging Face all still show three G​emini 4 models and two callable IDs.

Four days between a general-availability G​emini 4 launch and a carousel reference is fast, but not implausible: interface scaffolding for the next tier is exactly the sort of thing that gets written while the previous tier is still being announced.

What you can call today

If you want a G​emini 4 model in production this week, the honest answer is that there are two: G​emini 4 Carbon and G​emini 4 Argon, both reachable through a single OrcaRouter API key alongside 200-plus other models. We do not serve G​emini 4 Neon or G​emini 4 Cobalt, because G​oogle does not serve them to anyone yet — a routing layer cannot route a model that has no endpoint, and any platform claiming otherwise is describing a plan rather than a product.

The OrcaRouter model page for Google Gemini 4 Carbon, showing the model identifier google/gemini-4-carbon, input support for text, image, file, audio and video, $2.00 input and $10.00 output pricing per million tokens, a 1,048,576-token context window, 65K maximum output, and an OpenAI-compatible API.

What routing does buy you at this stage is that the eventual decision gets cheap. Both live G​emini 4 models sit behind the same key at provider list price with zero markup, so a change to a published G​emini 4 rate moves the number you pay the same day rather than at the next billing migration. Failover across providers means a model you want to trial during a leak cycle can sit behind a production path without that path depending on it: you can point traffic at G​emini 4 Carbon now and re-point it at whatever Cobalt turns out to be with a routing change rather than an integration. If you want to test Neon the day it appears, the work is a model ID, not a new contract.

That is the whole of the practical advice in this piece. Everything else is a forecast about a model that has not shipped.

What would settle it

Three things would move this from a leak to a release, in roughly this order of reliability:

• A model ID appearing in the documentation with a gemini-4- prefix. That signal has not produced a false positive in the G​emini 4 cycle so far, and it typically precedes general availability by days rather than weeks.

• A changelog entry naming the model. Announcements lag IDs, but a changelog row is written by the team that ships, so it settles intent rather than rumour.

• The SKU catalogue gaining "gemini 4 cobalt" rows. If that happens, Cobalt has cleared the same gate Neon passed, and the multipliers attached to those rows will say which tier it is aimed at.

Until one of those lands, the correct reading of the 18 November screenshot is narrow: two names exist in a user interface that has not been switched on, with example prompts written for people asking about hospital bills and used cars. G​oogle is thinking about who uses this model. Nothing in the screenshot tells you when you can, or whether the second name survives to launch. Both questions stay open, and the only thing a reader can act on today is the pair of models that already answer requests.