Gemini 4 Carbon vs Gemini 4 Pro: One Is Gated, the Other Was Never Shipped

The honest answer to "Gemini 4 Carbon vs Gemini 4 Pro" is that only one of those two names belongs to a model. Gemini 4 Carbon is real: announced by Google DeepMind on 28 October 2026, scored at 59 on the Artificial Analysis Intelligence Index v5.0, and handed this week to a subset of cyber defenders with no public endpoint. "Gemini 4 Pro" is a tier label rather than a product — there is no gemini-4-pro string in Google's API model list, no rate card and no model card behind it. What people mean when they type it is Gemini 4 Argon, the model Google announced on 30 September 2026 and made generally available on the paid API on 13 October. So the comparison this page can actually deliver is Carbon against Argon, with the Pro naming question answered first because it is the reason half the traffic arrives.

Where the second name comes from

Google's Gemini naming is positional, not sequential. Flash is the throughput tier, Pro is the general-capability tier, Ultra is the top tier when Google ships one, and Nano is the on-device tier. Over the 3.x generation the company moved Flash repeatedly — 3.5, 3.6, 3.7, 3.8 — while the Pro slot sat still from February's Gemini 3.1 Pro Preview onward. Anyone watching that pattern would reasonably expect the next flagship to arrive under the Pro name, and that expectation is what turned into search demand.

Reporting on the pre-launch leak gives the expectation a second leg. Internal material described the checkpoint behind Carbon as "Gemini pro next model," and that phrasing reads exactly like a model name to anyone outside the building. It is not one. The four checks that settle this are short:

• The Gemini API model list carries exactly one Gemini 4 entry, gemini-4-argon. There is no gemini-4-pro.

• Google's pricing page lists Argon's rates and no Carbon row. It lists no Pro row either.

• No model card named Gemini 4 Pro has been published at any point in 2026.

• Every benchmark number circulating under "Gemini 4 Pro" traces back to an Argon run, or to Carbon's announcement, or to neither.

That last point is the one that matters for anyone building on it. A name with no card behind it will absorb any number you attach to it, and the pages that compare "Carbon vs Gemini 4 Pro" on published figures are quietly comparing Carbon to Argon while implying a third model exists.

So the comparison is Gemini 4 Carbon against Gemini 4 Argon

Both are Google frontier models, fifteen days apart, sharing a price list and a rough architectural lineage — Carbon is explicitly built on Argon. They differ in exactly one dimension that a buyer cares about, and it is not performance.

• Announced — Carbon: 28 October 2026. Argon: 30 September 2026, generally available on the paid API 13 October.

• Intelligence Index v5.0 — Carbon: 59, first overall. Argon: 50, which is nine points back.

• Price — identical. Both carry a $2 input / $10 output introductory rate that becomes $4 / $20.

• Price clock — Argon's introductory rate ends 30 November 2026, with the standard rate applying from 1 December. Carbon's has no stated end date, because its clock starts when it gets an endpoint.

• Context — Argon's documentation lists a window above one million tokens shared between input and output. Carbon's announcement states no context window at all; the 1M on its Artificial Analysis page is that site's figure, not Google's.

• Access — Argon: callable on Google's paid API since 13 October. Carbon: Fairwind Program and a subset of cyber defenders, no model ID published.

The two numbers that actually separate them

Nine index points is the headline and it is not the most informative number in the launch. Three others are.

The first is 126: the Elo gain on AA-Briefcase, where Carbon reaches 1,609 from Argon's 1,483. That is the largest single improvement anywhere in the launch and it points at knowledge work rather than at coding, which is where Google's marketing points.

The second is 11%: Carbon's hallucination rate on AA-Omniscience v1.1, up from Argon's 4%, against an accuracy gain of ten points to 58%. Carbon is right more often and wrong more confidently, and the second half of that sentence is the one that matters in a security workflow where a fabricated finding is worse than no finding.

The third is $7.76: Carbon's standard-priced cost per Intelligence Index task, about 23% below Claude Opus 5.5 (max) and roughly 1.2x GPT-6 Astra (max). Both models run on the same rate card, so that number is a statement about how long Carbon thinks, not about what it charges.

There is also a reading of the launch that treats the nine-point gap as post-training rather than architecture — Ars Technica's assessment was that Carbon looks like "a slightly better version of Argon," possibly a refresh rather than a new model. Google has published nothing that settles it either way, and the absence of a technical report is exactly what keeps the question open.

Neither name is in our catalogue, and that is the useful part

Checked on 30 October, the OrcaRouter catalogue contains no Gemini 4 string of any kind — not Carbon, not Argon, and obviously not Pro. That is not a gap to paper over; it is the clearest available signal about what this generation currently is. Two announced models, one callable endpoint, and neither of them routed here yet. Any page that sells you "Gemini 4 Pro API access" today is selling a string that does not exist, and any page that sells Carbon access is selling a door Google has not opened.

What the catalogue does carry is the tier below: Gemini 3.1 Pro Preview at $2 input and $12 output per million tokens with a 1M-token context, Gemini 3.8 Flash and its siblings from $0.75 input, and the rest of the Google line from 2.5 through 3.8. On one key at 0% markup with provider list prices passed through, that is a working answer to the question "what do I call while Carbon is gated" — and the routing DSL is how you keep the answer flexible, composing a cheap model for triage with a stronger one for the final pass in a single call rather than standing up two integrations.

If you already call Argon in production

Two dates matter to anyone with Argon in a live path. The first is 1 December 2026, when the introductory rate ends and the bill doubles on both sides. The second is unconfirmed: on 21 October a third-party arena monitor flagged a capability change on the gemini-4-argon alias, and nobody has shown a before-and-after evaluation that would prove which weights are behind the name today. If that sighting was Carbon's post-training arriving under the Argon name, then the model you are calling today is not the model you benchmarked in October — and the nine-point index gap between the two names may end up describing one endpoint at two moments rather than two products.

That possibility is worth planning for rather than betting against. Automatic failover across providers on one key is the cheap insurance: if a primary deployment starts behaving differently, the call moves to the next provider in the chain without a redeploy, and you find out from a latency graph rather than from a customer.

So the matchup resolves into something simpler than the query suggests. Gemini 4 Carbon is the better model by nine points, and it is the one nobody can call. Gemini 4 Pro is not a model, and the thing it names — Gemini 4 Argon — is the one with a working endpoint, a published context window and a price increase scheduled for 1 December. If you came here to choose between two Gemini 4 models, there is only one to choose, and its real name is Argon.