Gemini 4 Carbon vs Gemini 4 Argon: The 15-Day Upgrade You Can't Call

Two models, one rate card, nine index points apart — and only one of them has an endpoint. Gemini 4 Carbon was announced on 28 October 2026, fifteen days after Gemini 4 Argon became callable in the paid Gemini API, and Google's own post describes Carbon as the stronger of the two. Strength is not the same thing as availability, though: there is no gemini-4-carbon entry in Google's model list or its pricing page today, and the people running it are cyber defenders and trusted partners admitted through the company's Fairwind Program. The honest version of this matchup is a capability comparison between two models and an access comparison between a model and a programme.

What Google put on the record on 28 October

The announcement came from Koray Kavukcuoglu, Google DeepMind's senior vice president, in a post titled "our next frontier model". Carbon succeeds Argon — announced 30 September, generally available since 13 October — inside the same generation number. Google's stated focus is cybersecurity and long-horizon agent work, the kind of task where an agent keeps going through dead ends for days rather than one where a prompt gets a single answer.

Three details matter more than the marketing:

• 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 for the last three groups was given.

• 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 public model card, no endpoint, no identifier. The Gemini API model list, checked on 29 October, contains exactly one Gemini 4 entry — Gemini 4 Argon — and the pricing page contains none for Carbon. The context window Google published for Argon (a unified 1M tokens, with decodes up to 1M) has not been restated for Carbon; Artificial Analysis lists 1M tokens for it, which is a third-party listing rather than a Google specification.

The nine-point gap, and who measured it

Artificial Analysis rebuilt its Intelligence Index as v5 on 27 October, the day before Carbon shipped, replacing two benchmarks, upgrading three others and making most of the index private. That timing is the reason to be careful with any before-and-after chart you see this week: scores from before the rebuild are not on the same scale as scores after it. On the current, post-rebuild listing, both models have been scored and the gap is real and measured by the same instrument.

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

• Cost per index task at standard pricing — Carbon $7.76 against Argon $7.68. The newer model scores nine points higher 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 in the trackers, but the direction is the wrong one.

• Output volume — 123k tokens per index task, 29% below Claude Opus 5.5's 173k, which is part of why the cost per task stays low despite the same sticker price.

• Speed — 99.8 output tokens per second on Artificial Analysis's measurement, with time to first token at 97.48 seconds at high reasoning effort. That is the number to watch if you are putting the model behind an interactive interface rather than a long-running job.

Artificial Analysis also measures the thing the headline hides: on its Coding Agent Index, Carbon scores 59 with Antigravity CLI, three points behind Claude Code driving Claude Opus 5.5 at 62. Google's own web-development result points the other way — WebDev Arena at 1,768, fourth place, a gain of 89 Elo — which is a reminder that "number one overall" and "best at code" are different claims.

Google's own numbers are the loudest ones in the post

Google reported a substantial security set, and it is worth separating what is measured from what is claimed. On SEC-bench Pro, vulnerability discovery, Carbon scores 88.2% against Argon's 83.4%. ExploitGym, which validates exploitation, moves from 42.0% to 54.8%. CWE-bench, patching, goes from 68.0% to 71.0%. Internally, Google says Carbon generates vulnerability-free code more than twice as often as Argon, and that its Chrome team found it more than ten times better than "another lab's large frontier model" at producing high-quality patches — a comparison with no named opponent and no published method.

The safety numbers are more consequential than the capability ones, because they are the first place Google has published Carbon beating Argon on behaviour rather than only on scores. On Gray Swan's indirect prompt-injection benchmark, Carbon's attack success rate is 0.0% against Argon's 0.7%. On Terminal-Bench 4.0, cheating was attempted in 0.4% of runs and succeeded in none, where Argon's behaviour was worse. High-severity misalignment alerts in internal live coding fell by almost 80%, unsupported claims halved, overreach fell 26% and under-service 29%.

Two caveats belong next to those figures, not in a footnote. Google says the cybersecurity evaluations were run without guardrails, which is the right way to measure a model's raw ability and the wrong way to predict what a deployed system will do. And the largest multipliers — the 2x secure code and the 10x patching — are internal benchmarks and a team report, not independent results. The Register's read on launch day was that the security case is promising and largely vendor-reported; nothing published since has changed that.

Figures circulating in launch coverage put Carbon ahead on Vals Index v2.2 (70.4% to 69.2%), SWE-sweep (9.6% to 4.3%), FrontierSWE 2 (61.2% to 55.0%), Agents' Last Exam v1.1 (77.4% to 69.0%), OSWorld-2.1 (81.5% to 74.7%) and long-context GraphWalks (96.2% to 84.2%). Those are Google-reported unless a benchmark owner says otherwise, and the pattern is consistent: Carbon is a broad improvement, not a uniform one, and it still trails Opus 5.5 and GPT-6 Astra on several software-engineering and computer-use tests.

Same sticker price, different clock

Both models carry the same published rate card: $2 per million input and $10 per million output introductory, with cached input at a 95% discount, moving to $4 and $20 afterwards. The asymmetry is in the calendar. Argon's introductory window has a date on it — Google's pricing page says the higher figures start on 1 December 2026 — and its batch, flex and priority tiers are published alongside. Carbon's discount has no end date at all; the announcement says only "after the introductory period".

Artificial Analysis prices both at the standard $4/$20, which is why its cost-per-task figures are the fair way to compare them and the introductory rate is not. A budget built on $2/$10 for either model has a cliff in it, and for Argon the cliff now has a date.

OrcaRouter routes neither of these models today: Gemini 4 Carbon has no public identifier to route, and Gemini 4 Argon is not in our catalogue either. The pass-through rule is still the relevant one, because it decides what happens when that changes — we charge 0% markup on provider list price, so the card Google publishes is the card a customer sees, including a mid-contract step from $2/$10 to $4/$20. What we do route from the same family is Gemini 3.8 Flash, at $0.75 per million input and $3.75 per million output with cache reads at $0.075, which is a different class of model but is the Google endpoint you can put in a production path this week.

Who should pick which

If you are a defender inside the Fairwind perimeter, the choice is not yours to make and Carbon is available to you now — the interesting question is whether the security gains survive your own harness, since the published ones were measured without guardrails and partly on Google's internal suites.

If you are building something that has to ship this quarter, Argon is the answer by default. It has an endpoint, a published price, and a date on the discount, which is more than Carbon has. Its weak spots are known: Terminal-Bench 4.0 in the high fifties on Google's launch figures, against Opus 5.5 at 66.4%, and a hallucination rate that Carbon has since overtaken in the wrong direction.

If you are scoping a security product, the decision is less about which model wins and more about which one you can test. Carbon is unavailable, so any vendor quoting you "Gemini 4 Carbon" performance today is quoting a label. Argon plus your own evaluation is the honest baseline against which Carbon's claims can later be checked.

The fifteen-day cadence is the actual story

Argon is four weeks old. Carbon arrived inside the same generation number rather than as Gemini 4.1 or Gemini 5, which tells you how Google intends to ship from here: variants, not versions. Ars Technica's reading on launch day was that Carbon is essentially a slightly better Argon — possibly a post-training refresh — and the release notes support that shape: identical price, identical context, identical input modalities, different capability mix.

For anyone integrating, that has a concrete consequence. A model name is no longer a stable description of behaviour. gemini-4-argon described one thing on 13 October and may describe another after a silent refresh; the arena sighting we covered on 21 October, in which an alias appeared to serve different weights, is neither confirmed nor refuted by the launch. The defence is the unglamorous one: pin the identifier you were actually evaluated on, keep a dated eval set, and treat any alias as something that can move under you.

What to watch next

• The first gemini-4-carbon entry in Google's model list — that, not the announcement, is the moment Carbon becomes a product.

• 1 December, when Argon's introductory discount expires and the effective comparison changes for every buyer still running it.

• A paired, third-party Carbon-versus-Argon evaluation on the security benchmarks, run with guardrails, since none has been published yet.

Until the first of those lands, the ranking is Carbon's and the decision is Argon's. That is an unusual thing to write about a model that has just been called the best in the world, but it is what the endpoints say.