Gemini 4 Carbon vs Gemini 4: One Is a Model, One Is a Generation

"Gemini 4" has never been a model you could call, and as of 29 October 2026 it is also no longer a name that points to a single thing. The generation now contains two named models: Gemini 4 Argon, which has had a public endpoint since 13 October, and Gemini 4 Carbon, announced on 28 October and running only for cyber defenders and trusted partners inside Google's Fairwind Program. So a comparison between Gemini 4 Carbon and Gemini 4 is not a matchup between two products. It is a question about which string belongs in a config file — and that question got harder, not easier, when Carbon shipped.

Three strings, three different objects

Most of the confusion comes from treating one generation name, one product name and one identifier as if they were interchangeable. They are not:

• Gemini 4 — the generation. It appears as a section heading on Google's model documentation page and as shorthand throughout coverage. It has no model ID, no endpoint, no price and no card of its own.

• Gemini 4 Argon — a model. Announced 30 September, stable in the paid Gemini API since 13 October under the identifier gemini-4-argon, with a 1M-token unified context, decodes up to 1M tokens, a published rate card and no free tier.

• Gemini 4 Carbon — also a model, announced 28 October. It scores higher than Argon on Google's and Artificial Analysis's headlines, costs the same at list price, and has no public identifier anywhere in Google's documentation.

Two more strings float around and resolve to nothing useful. Gemini 4.0 is a spelling of the generation, not a version — Google has never used it on a model page. Gemini 4 Pro is a label that appears in coverage and in search demand, but the model it is supposed to describe was released as Gemini 4 Argon; there is no gemini-4-pro in the API list either.

Where each string resolves, checked today

This is worth doing by hand rather than by argument. Google's Gemini API model page lists one Gemini 4 entry: Gemini 4 Argon, marked new and stable, endpoint gemini-4-argon. The pricing page lists the same single entry, with an introductory rate of $2 per million input and $10 per million output through 30 November 2026, moving to $4 and $20 on 1 December, plus batch, flex and priority tiers. Carbon appears on neither page. Its price exists only in the announcement, and it comes with no end date for the introductory window — the opposite of Argon's hard-dated step.

The model list is the part that answers the title question cleanly. "Gemini 4" does not resolve to an endpoint. "Gemini 4 Carbon" does not resolve to an endpoint yet. "Gemini 4 Argon" does, and has since 13 October. Everything else is naming convention.

Why Carbon kept the 4

Fifteen days between two models, and Google chose not to call the second Gemini 4.1. That decision is the whole reason this comparison exists. Carbon is positioned as a new flagship inside the same series, aimed at cybersecurity work and long-horizon agent runs rather than at being a new generation.

The launch coverage converges on the same read. Ars Technica called it essentially a slightly better version of Argon, possibly a post-training refresh. The Register framed it as Google's strongest model, priced at Argon's introductory level but not yet broadly available, and put the security gains in the vendor-reported column. The shape of the release supports that: same list price, same 1M context from the same family, same text output and the same text, image, video and speech inputs. What differs is capability mix and access.

That is also why the arena sighting we wrote about on 21 October matters here. An observer reported that the gemini-4-argon alias was serving different weights, said to be Carbon, and the evidence at the time was a capability-change card and a video — not a checkpoint comparison. The 28 October announcement neither confirms nor refutes that claim: Carbon exists as a named model, which was the disputed part, but nothing Google published says it was ever served behind the Argon identifier. The alias question is still open, and it is the more important of the two questions if you run Argon in production.

Is Carbon a different model, or Argon with more training?

The practical test is what changes for a caller. On the numbers that exist, Carbon is nine points higher on the current Artificial Analysis Intelligence Index v5 — 59 against 50, both scored after the index was rebuilt on 27 October — at effectively the same cost per task, $7.76 against $7.68. It also hallucinates more: 11% against Argon's 4%, which is the one axis where the newer model is worse and the one that shows up in production rather than in a leaderboard screenshot.

On security benchmarks Google reports large gains: SEC-bench Pro 88.2% against 83.4%, ExploitGym 54.8% against 42.0%, CWE-bench 71.0% against 68.0%, all described as measured without guardrails. Those are the claims that make Carbon interesting to defenders and the least independently verified part of the package. Artificial Analysis, which did measure the general index, also puts Carbon behind Claude Code with Claude Opus 5.5 on its Coding Agent Index — 59 against 62 — so "newer and higher-ranked" still does not mean "best at code".

Whether that adds up to a different model or a better-trained one is a question about weights, and Google has published no model card. The honest answer today is that the behaviour differs measurably and the architecture is unstated. Anyone who tells you which it is, with confidence, is guessing.

What to call today

For most readers the answer is not Carbon and not "Gemini 4" — it is Argon, because it is the only one of the three strings that returns a completion. It is available in the paid Gemini API, at a price that steps up on 1 December. If you are writing integration code, use the identifier rather than the display name, and pin the version you evaluated rather than an alias that can be repointed.

If you are inside a cyber-defence organisation, Carbon is the one to apply for through the Fairwind Program, with the caveat that its introductory price has no published end date and its model card does not exist yet — two things to get in writing before you build a budget around it.

If you need a Google model in production through a single API today, neither Gemini 4 model is in OrcaRouter's catalogue, and we do not claim otherwise. What we do route is Gemini 3.8 Flash, at $0.75 per million input and $3.75 per million output with cache reads at $0.075, alongside 200-plus other models on one key with 0% markup on provider list price. That matters for the next six weeks in a specific way: when Carbon does get an identifier, the pass-through rule means the price Google publishes is the price you get, and automatic failover means trying an unproven flagship does not have to be a bet on one provider's uptime. The routing DSL does the rest — you can compose the new model into one call beside the one you trust, rather than migrating a product to it.

The question to ask before anyone sells you "Gemini 4"

Because the generation name has no identifier, it is the easiest thing in this family to sell loosely. Three questions sort a real offer from a label:

• Which model ID will I be calling? A real answer is gemini-4-argon or a Carbon identifier that appears in Google's documentation. "Gemini 4" is not an answer.

• Which rate card, and until when? Argon's introductory pricing ends 30 November 2026. Carbon's has no published end date. A quote that does not name the card and the date is a quote for a different product than the one you will be billed for.

• Is it gated? Carbon's access runs through the Fairwind Program, so a vendor offering it to the general public is either reselling an entitlement it may not have or describing a different model.

The rule of thumb is short enough to remember: Gemini 4 is where the family lives, Gemini 4 Argon is what you can call, and Gemini 4 Carbon is what Google says is better but has not yet let anyone outside its programme use. Until a gemini-4-carbon endpoint appears, any comparison involving "Gemini 4" is really a comparison with Argon, plus a note about a name that never had a model behind it.