Gemini 4 Carbon Is Number One on the Index and Unavailable: What Changed on 28 October

Google DeepMind published Gemini 4 Carbon: our next frontier model on 28 October 2026, and inside that post sat the number that has since been quoted everywhere: 59 on the Artificial Analysis Intelligence Index v5.0 — first place, one point clear of Claude Opus 5.5 at 58, and nine clear of Gemini 4 Argon, the model Carbon is built on and the model Google has been selling on its paid API since 13 October. Artificial Analysis is the third-party evaluator behind that figure, and it had the row published the same day. What did not arrive on 28 October, or on 29 or 30 October, is a way for anyone outside a named programme to call the model.

That gap — between a verified top ranking and a door with no handle — is the whole story of the week. Carbon is not a leak, not a rumour and not a preview that someone spotted on a leaderboard. It is an announced, benchmarked, priced model with a published rate card and no published model ID. Below is what actually changed, what Google withheld, and what the numbers mean once you stop reading the headline score and start reading the nine tests it is averaged from.

What the post said, and the four things it did not

The announcement names Koray Kavukcuoglu, Google DeepMind's SVP and chief AI architect, and positions Carbon as a cybersecurity and long-horizon engineering model built on Argon. The access wording is precise and worth reading twice: "Starting this week, we will make Carbon available to a subset of cyber defenders," then "gradually phase access to additional trusted partners through our Fairwind Program," and finally developers, businesses and consumers "as soon as possible" once safeguards are refined. Google also says it is engaged in the U.S. government's voluntary pre-release testing process. No date appears anywhere in that sequence.

Four things a model page normally carries are absent:

• Model ID — the Gemini API model list, checked on 30 October, contains no Carbon entry. Its only Gemini 4 string is gemini-4-argon.

• Context window — not stated in Google's post and not present on its pricing page. Artificial Analysis's model page lists 1M; that is a third-party figure, not a vendor specification, and it is the only one in circulation.

• Parameter count and architecture — not disclosed, and no technical report accompanies the launch.

• A model card — none published as of 30 October, which is why every secondary outlet is quoting the blog post rather than a specification.

Pricing, by contrast, is published in full, which makes the omission set look deliberate rather than unfinished: Google is willing to tell you what Carbon costs before it is willing to tell you what it is.

The 59 is an average of nine tests, and Carbon wins four of them

Artificial Analysis's own write-up is more useful than the rank. Carbon leads 4 of the 9 constituent evaluations — HLE-Diamond, Terminal-Bench-Science, AA-Omniscience and AutomationBench-AA — which means the first-place score comes from being very good across the board rather than from sweeping the set. The places it does not lead are the ones a developer hiring an agent would notice first.

• Coding agents — Carbon scores 59 on the Coding Agent Index with Antigravity CLI. Claude Opus 5.5 reaches 62 in Claude Code. Three points, on the exact workload where Google's marketing leans hardest.

• Terminal-Bench 4.0 — Carbon is second at 64%, behind Claude Opus 5.5.

• FrontierCode v1.1 — second at 51%, again behind Opus 5.5.

• AA-Briefcase — 1,609 Elo, up 126 from Gemini 4 Argon's 1,483. That is the biggest single jump in the launch, and it still leaves Carbon behind Grok 4.7 (1,638), Muse Wonder (1,659) and Kimi K3.1 Pro (1,678).

• AA-Omniscience v1.1 — 54, ahead of Opus 5.5's 48 and Argon's 46. Accuracy rose 10 points to 58%, and the hallucination rate rose from 4% to 11%. The second number is not in Google's post.

The honest reading of 59, then, is that Carbon is the strongest general-purpose model on the board and the second-strongest coding agent, with a worsened hallucination rate that Google did not volunteer. All three statements come from the same page.

Google's own table, labelled as Google's

The vendor numbers are the ones that will end up in procurement decks, so they deserve a precise attribution rather than a place in the same sentence as the independent score. Everything below is Google-reported and unreproduced outside Google:

• SEC-bench Pro — 88.2% on vulnerability discovery, the headline cybersecurity claim.

• ExploitGym — 54.8% on exploitation, against 42.0% for Gemini 4 Argon.

• CWE-bench — 71.0% on patching.

• WebDev Arena — up 89 Elo to 1,768, fourth place. This one is a public arena and therefore checkable, which is why it is the vendor claim least in need of a caveat.

• Gray Swan indirect prompt injection — 0.0% attack success, against Argon's 0.7%. Terminal-Bench 4.0 cheating: 0.0% successful, 0.4% attempted.

• Internal engineering claims — more than 2x secure-code generation success against Argon; a Chrome team statement that Carbon produced "more than 10x" better patches than a rival lab's frontier model; severe misalignment alerts down almost 80%; unsupported assertions roughly halved; task overreach down 26% and under-service down 29%.

Google's table is not uniformly flattering, and the two places it loses are the ones worth remembering: FrontierSWE 2 went to GPT-6 Astra, and the OSWorld-2.1 offline subset went to Claude Opus 5.5. A launch post that publishes its own losses is more credible than one that publishes seventeen wins with no column for the rest — which is exactly why the loss lines are the first thing to check when a vendor claims a sweep.

The rate card, and the clock Google did not start

Carbon's introductory price is $2 per million input tokens and $10 per million output tokens, with cached input at 95% off. The post then states that "$4 per 1M input tokens and $20 per 1M output tokens will apply" after the introductory period expires — and gives no expiry date. Argon, by contrast, has a hard one: its introductory rate runs to 30 November 2026, and the standard rate steps in on 1 December. Carbon's clock is not started, and it will not start until the model has a public endpoint to sell.

At standard pricing, Artificial Analysis puts Carbon's cost per Intelligence Index task at $7.76 — about 23% below Claude Opus 5.5 (max) at $10.04, and roughly 1.2x GPT-6 Astra (max) at $6.25, on 123k output tokens per task. Those are per-task dollars, not per-token dollars, and the distinction matters: a model that is cheaper per token but thinks longer can still cost more per answer.

This is the part of the launch that will become relevant the moment Carbon gets an endpoint, and it is the part the pass-through rule on the full model catalogue is built for. OrcaRouter charges 0% markup and passes the provider's own list price through, so a vendor rate cut or a step-up is live on our side the same day rather than after a repricing cycle. Today that rule applies to the Google models we do route; Carbon is not one of them, and the pages that claim otherwise are selling something nobody can deliver yet.

What a developer can actually do this week

On 30 October, the OrcaRouter catalogue contains no string matching Gemini 4. The frontier Gemini models that are callable today sit one rung down and are priced accordingly: Gemini 3.1 Pro Preview at $2 input and $12 output per million tokens with a 1M-token context, and the Flash tier — Gemini 3.8 Flash, Gemini 3.6 Flash, Gemini 3.5 Flash — from $0.75 input. That is what "the top model is gated" translates to in practice: the work still has to ship, and it ships on the best model with a door.

If you are evaluating Carbon for a security workflow, the credible way to prepare without betting a production path on an unreleased model is to build the fallback first. Automatic failover across providers on one key means a primary model that starts returning errors or slows down gets replaced inside the same call, so the day Carbon opens, the switch is a model-string change rather than a re-architecture.

The precedent that sets the date to watch

Argon is the only usable template for what happens next. Google announced it on 30 September 2026, moved it to general availability on the paid API on 13 October, and opened its introductory price window to 30 November. That is thirteen days from announcement to a callable endpoint. Apply the same arithmetic to Carbon's 28 October announcement and the implied date is around 10 November — arithmetic on a single precedent, not a commitment, and Google has been explicit that the government pre-release testing has to finish first.

The other variable is less comfortable: the model that ships in two weeks may not be the model that scored 59. Carbon is built on Argon, and there is an unconfirmed observation from 21 October, when a third-party arena monitor flagged a capability change on the gemini-4-argon alias and some developers read it as Carbon's weights arriving early. If Google's plan is to fold Carbon's post-training into the Argon name rather than launch a second endpoint, then the thing to watch is not a new model ID at all — it is a version bump on the one you can already call.

The ranking will hold for a while; leaderboard firsts usually do until the next lab answers. What will not hold is the assumption underneath most of this week's coverage — that a number-one model is a model you can put in a product. On 30 October, Gemini 4 Carbon is the best model in the world by one published point and the only frontier model in this generation with no address. Gemini 4 Argon, four weeks older by announcement and nine points lower, has the address, the rate card with a real deadline, and the traffic.