Gemini 4 Carbon: Announced, Gated, and Nine Points Ahead of the Model You Can Actually Call

Gemini 4 Carbon stopped being a rumour on 28 October 2026, when Google DeepMind published a post titled "Gemini 4 Carbon: Our next frontier model." What it did not stop being is unavailable. There is no gemini-4-carbon string in the Gemini API model list, no entry on its pricing page, no model card and no endpoint — the model list, checked on 29 October, carries exactly one Gemini 4 entry, and it is Gemini 4 Argon. So the honest description of Carbon today is a shipped announcement wrapped around a staged rollout: capability documented, access rationed, price unpublished.

The independent measurements that landed alongside it make this more than a press release. Artificial Analysis rebuilt its Intelligence Index as v5 on 27 October — one day before Carbon's announcement — and on that post-rebuild scale both models have been scored by the same instrument. Carbon 59, first of 52 listed models. Argon 50, thirteenth. Nine points apart, at almost identical cost per task. That is a real number about a model most developers cannot call, and the gap between those two facts is what this page is about.

What Google actually put on the record

The announcement is short on the things a buyer needs and specific about the things a lab wants to be judged on. Stripped to its claims:

• Positioning — Carbon is described as the stronger of the two Gemini 4 models, succeeding Argon inside the same generation number. Google's stated focus is cybersecurity and long-horizon agent work — tasks where an agent keeps going through dead ends for days rather than a prompt that gets one answer.

• Access is staged and undated — Carbon went first to a subset of cyber defenders, then widens through the Fairwind Program, then to developers, businesses and consumers "as soon as possible." No date was given for the last three groups.

• Regulatory path — The release sits inside the United States government's voluntary pre-release testing process. Google says it is participating and is still refining safeguards before broader access.

• No paperwork — No public model card, no endpoint, no identifier. The context window Google published for Argon — a unified 1M tokens with decodes up to 1M — has not been restated for Carbon; the 1M figure that circulates comes from Artificial Analysis's listing, which is a third-party record rather than a vendor specification.

That is the announcement. The distance between "we have built the better model" and "here is the price" is, in this case, the distance between a blog post and a product.

The nine-point gap, and the instrument that measured it

The timing of the index rebuild is the thing to hold onto, because it invalidates most before-and-after charts circulating this week: scores taken before 27 October are not on the same scale as scores taken after. On the current listing, both models are measured by the same instrument:

• Intelligence Index v5 — Carbon 59, first of 52 listed models; Gemini 4 Argon 50, thirteenth.

• Cost per index task — Carbon $7.76 against Argon $7.68. Nine points of capability for eight cents more per task.

• Accuracy — Carbon 58%, a ten-point gain over Argon; AA-Omniscience 54 against 46.

• Hallucination — Carbon 11%, up from Argon's 4%. Carbon is still second-lowest on that tracker, but the direction is the wrong one.

• Output volume — 123,000 tokens per index task, 29% below Claude Opus 5.5's 173,000 — a large part of why the cost per task stays low despite the identical-looking sticker price.

• Speed — 99.8 output tokens per second, with a 97.48-second time to first token at high reasoning effort.

Which number matters most depends on what you are building. A 97-second first token is irrelevant for a long-running agent job and disqualifying for an interactive interface. And "first overall" is not "best at code": on Artificial Analysis's Coding Agent Index, Carbon scores 59 with Antigravity CLI, three points behind Claude Code driving Claude Opus 5.5 at 62 — while Google's own WebDev Arena result points the other way, at 1,768 and fourth place, a gain of 89 Elo.

Google's own numbers, and how to read them

The post carries a security benchmark set, and every figure in it is vendor-reported rather than independently reproduced. Label them that way when you cite them: on SEC-bench Pro, vulnerability discovery, Carbon scores 88.2% against Argon's 83.4%, and the exploitation-validation track and long-horizon agent results follow the same pattern of a few points over the incumbent. None of it has an outside replication yet, and a benchmark chosen by the lab that built the model is evidence of intent as much as of capability.

The more useful reading of that section is what it says about who Carbon is for. A model whose marquee results are vulnerability discovery and long-horizon agent endurance, released first to cyber defenders through a named programme, is not being launched as a general-purpose chat upgrade. It is being launched as a specialist — and priced, when it is priced, accordingly.

What the October leak got right, and what it did not

Background worth keeping, because the earlier reporting is now partly vindicated. On 9 October, Business Insider reported — relayed by the leak-tracking outlet TestingCatalog — that Google staff were internally testing a Gemini 4 checkpoint called Carbon, that early feedback described it as feeling comparable to Claude Opus 5.5 for coding, that Barium-B was the checkpoint selected for the public Argon release, and that whether Carbon would become an Argon update or a release of its own was unconfirmed.

Two of those came true in a specific way: Carbon is a separate thing rather than an Argon update, and it is the stronger of the two. The part that has not been vindicated is the flattering one — "feels comparable to Claude Opus 5.5 for coding" was a staff impression, and the independent coding measurement puts Carbon three points behind Claude Opus 5.5 in a Claude Code harness. A feeling is not a score, and this week the score arrived.

The 21 October sighting also reads differently now. A post claimed that the weights behind the gemini-4-argon alias had quietly become Carbon, supported by a capability-change notification from a third-party arena bot. With both models now scored separately by the same tracker — 59 and 50 — that claim looks less likely rather than more: if Argon's alias had been serving Carbon since 21 October, the two rows would not be nine points apart on an instrument that rebuilds monthly.

What is still missing, and when it matters

Three absences define the next phase, and each has a different consequence.

• An identifier — Until gemini-4-carbon appears in the model list, nothing you build can name it. This is the line between an announced model and a callable one.

• A price — Nothing is published at any tier. Argon's card is the only anchor in the family: $2.00 per million input and $10.00 per million output under the introductory rate, cached input at 95% off, then $4.00 and $20.00. Whether Carbon inherits that or breaks it is the most consequential unknown, because a nine-point gain at workhorse rates would reprice a large part of the market, and a nine-point gain at frontier rates would not.

• A model card — Which checkpoint, at which effort setting, with which context length actually offered. The 1M figure circulating today is a third-party listing.

Also unresolved: whether Carbon becomes an Argon update after all. The announcement describes it as succeeding Argon inside the same generation, which leaves open the possibility that the two names collapse into one identifier later — the exact scenario the October reporting could not rule out.

What to run while access is rationed

Neither Gemini 4 Carbon nor Gemini 4 Argon is on OrcaRouter. Querying our catalogue for either returns nothing, and it will keep returning nothing until Google makes them callable to us. Any account selling "Gemini 4 Carbon access" today is selling a label; the only documented route is the Fairwind Program.

What is callable is the model both the staff impression and the coding index keep being measured against. Claude Opus 5.5 is live on OrcaRouter at Anthropic's list price — $4.00 per million input and $20.00 per million output, cache reads at $0.20 — passed through with no markup, on the same key as everything else. Gemini 3.8 Flash is there at $0.75 per million input and $3.75 per million output with cache reads at $0.075, which is what the non-frontier share of your traffic should cost while you wait.

The decision this week is not "should we move to Carbon." It is how much of a nine-point capability gap, at an unknown price, gated behind a security programme, is worth planning around. Waiting costs nothing: whichever identifier becomes callable is added to the same endpoint at Google's own price on the day it is listed, because there is no renegotiation between a vendor publishing a price and us passing it through. Planning around it costs a migration.

What we are watching, in order: the identifier appearing in the model list; a price sheet; a model card; and the second stage of the Fairwind rollout, which is where "as soon as possible" gets a date or quietly does not.