The Gemini After Gemini 4 Carbon: One Leak, Zero Specs, and a 28-Day Cadence

Just after 02:00 UTC on October 29 — the evening of October 28 in California, the day Google announced its new frontier model — the account @kimmonismus posted a sentence with nothing behind it but the sentence: “they have another unreleased Gemini model that’s even better than Carbon.” No codename, no benchmark, no date, no screenshot, no second source. The one checkable clause in the same post is the part everyone already knew: Gemini 4 Carbon had taken first place on the Artificial Analysis Intelligence Index v5.0 that day, at 59. And the model Carbon is built on, Gemini 4 Argon, was announced twenty-eight days earlier and reached general availability fifteen days ago.

So this is a what-we-know-so-far piece about a model that does not have a name, a score or a date — and the honest way to write it is to keep three kinds of evidence apart. What Google has published is a blog post, a price and an access plan. What independent measurement adds is a leaderboard position and a cost-per-task. What the post asserts is that something faster exists inside the lab. Only the first two can be checked, and only the third is news to anyone who has not been watching the cadence.

The claim, and what is actually in it

The post is four clauses long and only one of them is a claim about a future model. It says, in order: another unreleased Gemini exists; it is better than Carbon; Carbon is number one on the Artificial Analysis Index; and this looks like a significant step toward recursive self-improvement. The third clause is verifiable and true. The first two are unattributed. The fourth is an interpretation, and it is the part Google's own leadership has been most careful to hedge — more on that below.

What the post does not contain matters as much as what it does. There is no internal codename, which is unusual for this particular story: the Carbon leak that preceded the launch came with one, plus a platform name, plus a handful of employee quotes. There is no artifact of the kind that has carried every credible Gemini leak this year — a framework pull request, a model-registry entry, an arena alias, a Helm chart. There is no date, no size, no modality list, and no indication of whether the claim comes from a person inside Google, from a partner, or from inference over the launch itself. The replies under the post are jokes, including one that reads “Gemini [redacted] isn’t yet here though,” which is the correct amount of confidence to take from the thread.

It is also worth dating the “today” precisely, because it is the only timestamp in the story that can be pinned. The post went up at 02:06 UTC on October 29. In Mountain View that is 19:06 on October 28 — the same day Google published the Carbon announcement and the same day Artificial Analysis published its write-up putting Carbon first. The leaderboard claim is not a scoop; it is the launch, restated after dinner.

What Google published on October 28, and what it did not

The primary document is short and unusually specific about access. Koray Kavukcuoglu, SVP of Google DeepMind and chief AI architect, published “Gemini 4 Carbon: our next frontier model” on blog.google on October 28, 2026. It describes Carbon as “our strongest model with state-of-the-art cybersecurity and long horizon capabilities,” says it is built on Gemini 4 Argon, which “launched two weeks ago,” and states that Carbon “ranks #1 on the recently updated v5.0 of Artificial Analysis Intelligence Index at 59.”

The access paragraph is the one to read twice. Google says it will make Carbon available “to a subset of cyber defenders” starting that week, then “gradually phase access to additional trusted partners through our Fairwind Program,” and only then bring it “to developers, businesses, and consumers via our product surfaces as soon as possible.” That is a sequence with no dates in it beyond the first step. As of October 30 there is still no public model identifier, no endpoint, and no way to call Carbon through an API — which is the state the model was in when the post above was written, too.

What the post does not contain is a specification. No parameter count, no architecture description, no context window, no technical report, and no mention of any model newer than Carbon. Ars Technica’s reading, reported alongside the launch, is that Carbon may be a post-training upgrade of Argon rather than a new architecture; Google does not say either way. The absence of a successor is not proof there isn’t one — companies do not pre-announce the next checkpoint in the current one’s launch post — but it does mean the only first-party statement in this story is about Carbon.

What independent measurement adds

Artificial Analysis is the closest thing to a neutral referee in this story, with the usual caveat that it runs its own evaluations rather than replaying anyone’s vendor table. Its October 28 write-up puts Gemini 4 Carbon at 59 on Intelligence Index v5.0, first of the models it tracks, ahead of Claude Opus 5.5 at 58 and GPT-6 Astra at 57. It notes that Carbon leads four of the nine constituent evaluations — HLE-Diamond, Terminal-Bench-Science, AA-Omniscience and AutomationBench-AA — which is a narrower claim than “best model” and a more useful one.

The same write-up is where the launch story gets its counterweight. Carbon is second, not first, on the Coding Agent Index at 59 against Claude Opus 5.5 with Claude Code at 62. It is second on Terminal-Bench 4.0 and second on FrontierCode v1.1. On AA-Briefcase, the knowledge-work Elo, it reaches 1609 and still sits behind Grok 4.7, Muse Wonder and Kimi K3.1 Pro. Its measured cost is $7.76 per Intelligence Index task, roughly 23% below Opus 5.5’s $10.04 — an efficiency win that the “number one” framing hides.

One methodological point has to travel with every number in this piece. Artificial Analysis rebuilt the Index to v5.0 on October 27, replacing 60% of the composite with private or held-out test sets and swapping several constituent evaluations. Scores from before that rebuild are not on the same scale, so Carbon’s 59 cannot be subtracted from an older Argon figure, and any article that shows you a nine-point gain across the rebuild is comparing two different rulers. The v5.0 numbers in this piece are all read from the post-rebuild run.

And the detail that matters most for this article: Artificial Analysis’s write-up mentions exactly two Gemini models, Carbon and Argon. There is no anonymous row above Carbon, no “Gemini preview” entry, nothing unnamed sitting at 60 or above. If a stronger model were already being measured somewhere public, that is the first place it would surface.

Why a newer checkpoint in the lab is plausible

The claim is not implausible, and the reason is arithmetic rather than inside knowledge. Gemini 4 Argon was announced on September 30 and reached general availability on October 13. Carbon was announced on October 28 — twenty-eight days after Argon’s announcement and fifteen after its GA. Google’s own post frames that gap as the point: Carbon “illustrates the pace at which we’re pushing the frontier forward.” Sundar Pichai described the releases, in his own post on X, as a cadence of periodic model releases rather than one-off generational launches.

Reporting from earlier in October fills in the machinery behind the cadence. Business Insider’s October 9 story, based on internal documents and employee messages, named a series of checkpoints — Argon, Barium, Carbon — and reported that “Barium-B” was the internal build that shipped publicly as Argon. Carbon was being tested on Google’s internal coding platform, Jetski, where one employee wrote that it “feels like Opus 5.5” for coding and another that “Carbon is really good.” A third internal note used the “Next” label the company keeps for Gemini updates: “New Gemini pro next model feels kinda nice.” Google declined to comment.

That reporting is the strongest structural support the October 29 post has, and it is worth being precise about what it supports. It establishes that Google was running several checkpoints in parallel and shipping them at short intervals, which is exactly the regime in which a newer, better checkpoint exists before it is announced. It does not establish that such a checkpoint exists now, or that anyone outside the company has seen it, or that it is better than Carbon rather than merely different. A fast cadence makes the claim likely; it does not make the claim sourced.

The names circulating are inventions, and the periodic table shows why

Search for the successor and you will find it named. You will see “Gemini 5,” usually attached to a ten-trillion-parameter claim, a 2027 date and TPU V8 hardware. You will see “Krypton,” and occasionally “Dysprosium,” derived from the element-themed codenames of the Gemini 4 line. None of these names has appeared in any Google document, blog post, API listing or executive statement. Google has never publicly used “Gemini 5”; the coverage that treats it as a known product is extrapolation from a naming pattern.

The pattern itself is thinner than it looks. Argon is element 18, Barium is 56, Carbon is 6. That is not a sequence, and it is not a countdown or a roadmap — it is three elements chosen for reasons nobody outside the company knows. The stronger counterexample is Barium-B, an internal name that became Argon, a different public name entirely. Codenames in this program have already been shown to decouple from product names, which means even a real leak of a codename would not tell you what the model will be called when it ships — and a guessed codename tells you nothing at all.

There is a precedent for how these guesses go wrong, and it is recent. On October 21 a widely shared argument held that the gemini-4-argon alias had quietly begun serving Carbon weights, inferred from arena behaviour rather than from any published change. It was never confirmed, and the launch seven days later described Carbon as a distinct model with its own name and its own announcement. Inference from behaviour produced a confident, wrong answer once already this month.

“RSI” is the weakest word in the post

The recursive-self-improvement half of the claim has a traceable origin, and it is a quip. When Business Insider’s October 9 story ran, Vedant Misra — a Google DeepMind researcher — replied on X with the line “Have you heard of recursive self improvement,” without a question mark. That is the seed the framing grew from, and it is a joke about the reporting, not a disclosure about the training loop.

Google’s leadership has since said more, and it points the other way. Sergey Brin said in early October that Google was “starting to get onto that RSI treadmill,” and that “our researchers and the model work hand in hand to make the next version of the model.” Sundar Pichai, on the Big Technology Podcast in the week before the Carbon launch, was more careful: “We are pretty far from the end state of RSI. The right way to describe it is, you have RSI loops along with AI researchers in various aspects of model development. I would say you almost have agents helping AI researchers, in some automated way, work together to build these models.” He expected that continuum to keep improving into next year.

Google’s own launch post uses the same cautious construction: “Our researchers and models are working hand in hand to advance capabilities faster.” That sentence is a description of tooling — agents assisting the people who build models — and it is the vendor statement closest to the RSI claim. It is not a statement that a model designed its successor, and nobody at Google has made that statement.

The distinction is worth a paragraph of its own because it is the difference between a pace story and a capability story. A Gladstone AI report cited in the same Business Insider piece found that most researchers surveyed at Anthropic and DeepMind expect RSI within three to twelve months. A survey of expectations is not a result. What is observable is the interval: 28 days from Argon’s announcement to Carbon’s, and a new model at the top of an independent index at the end of it. That is a fast cadence. Whether humans or models did the iterating is precisely the part no outside observer can check, and a single post asserting it does not close the gap.

What “better than Carbon” would have to beat

If the claim is real, it has a scoreboard to clear, and every line of it is published. This is the useful part of the story even while the model stays unnamed: it turns “better” from an adjective into a set of thresholds.

• Intelligence Index v5.0 — Gemini 4 Carbon 59, first overall; a successor would need 60 or more on the same post-October-27 rubric, not on the older one.

• The four evaluations Carbon already leads — HLE-Diamond, Terminal-Bench-Science, AA-Omniscience and AutomationBench-AA; a newer checkpoint that gains a point overall while losing one of these is a different model, not a strictly better one.

• Coding Agent Index — Carbon 59, second behind Claude Opus 5.5 at 62, so “better at coding” means passing 62, not passing Carbon.

• Google’s own cybersecurity figures, run without guardrails — SEC-bench Pro 88.2%, ExploitGym 54.8%, CWE-bench 71.0%; these are vendor-reported, unreproduced by any third party, and they are the numbers a successor would be marketed against.

• Price — $2 per million input and $10 per million output during the introductory period, moving to $4 and $20 afterwards, with cached input at 95% off; a successor priced above that is not automatically the better buy, and one priced below it would matter more than a point on the Index.

• Access — the same gate. Carbon is Fairwind-only today; a stronger model behind a tighter gate changes nothing a developer can act on.

What you can actually call today

Carbon is not callable. Argon is. Gemini 4 Argon has been generally available as gemini-4-argon since October 13 at the same introductory tier — $2 per million input tokens, $10 per million output — and it is the only Gemini 4 tier with a public endpoint. If the choice in front of you is “wait for the unnamed model” or “ship against something that answers today,” the second option has a model string, a price and a two-week track record; the first has a post.

Neither Carbon nor Argon is in the OrcaRouter catalogue as of October 30 — we checked, and there is no google/gemini-4-* entry to route to, so nothing here should be read as an offer to serve either one. What the catalogue does carry from Google is the 3.x line, including google/gemini-3.8-flash, which is the practical floor for work that needs a Google model today. You can see the models we route in the catalogue, and the reason that matters here is timing rather than preference: when Google does open Carbon — or whatever the next checkpoint is called — it will arrive as one more identifier behind a key you already hold, priced at list with 0% markup passed through, next to a failover chain that keeps the incumbent serving if the new tier turns out to be slower or rate-limited on the day. Testing an unannounced model is a routing decision; betting a production path on one is not.

What would turn this from a post into a fact

Five things are checkable, and four of them are free. A row above Carbon on the Artificial Analysis Intelligence Index, on the same v5.0 rubric. An anonymous Gemini entry on LMArena, which is how Carbon itself first produced public outputs before launch. A new identifier in Google’s Gemini API model list, which is the difference between a checkpoint and a product. A Fairwind expansion notice on Google’s blog, which is the mechanism by which any gated Gemini 4 model has reached developers so far. And a second, independent publication with its own sourcing rather than a paraphrase of the same post.

Until at least one of those appears, the honest summary is the one the evidence supports: Google shipped a frontier model on October 28 that is first on an independent index and second on the coding-agent one, gated it to cyber defenders, priced it at $2/$10 introductory, and described a cadence of periodic releases that makes a further checkpoint likely. A single post says that further checkpoint is better. The cadence is the part you can plan around; the comparison is the part you cannot, because the model it describes has no name, no score and no date — and in a market where the callable model is always the one that decides your bill, an unnamed successor is not a reason to delay anything.