Gemini 4 Carbon vs Gemini 3.1 Pro: Eight Months Apart, and the Index Will Not Let You Compare Them

Put Gemini 4 Carbon and Gemini 3.1 Pro side by side and the first thing you get is a gap that cannot be measured. Carbon's 59 on the Artificial Analysis Intelligence Index v5.0 is real, independently run, and current. Gemini 3.1 Pro Preview has no Intelligence Index score at all: Artificial Analysis lists it with a dash where Carbon has 59, and whatever index figure you have seen attached to it predates the v5.0 rebuild of 27 October. Any page that subtracts the two is doing arithmetic across two versions of a ruler. The comparison that survives is narrower, and it is more useful: one model is the best in the world and closed, the other is eight months into a preview and open at $2 per million input tokens.

What the two models actually are

Gemini 4 Carbon is Google DeepMind's frontier model announced on 28 October 2026, built on Gemini 4 Argon, aimed at cybersecurity and long-horizon engineering, priced at an introductory $2 input and $10 output per million tokens that later becomes $4 and $20, and available only to a subset of cyber defenders this week with no published model ID. Gemini 3.1 Pro is Google's previous Pro-tier flagship, shipped as Gemini 3.1 Pro Preview on 19 February 2026 and still carrying the preview suffix more than eight months later, with a 1,048,576-token context window, a 65,536-token maximum output, and multimodal input across text, image, speech and video.

• Context — Gemini 4 Carbon: 1M on the Artificial Analysis page, unstated in Google's own documentation. Gemini 3.1 Pro Preview: 1,048,576 tokens.

• Maximum output — Carbon: not published. 3.1 Pro Preview: 65,536 tokens, the number that gives away its age.

• Price, introductory — Carbon $2.00 in / $10.00 out per million. 3.1 Pro Preview $2.00 in / $12.00 out.

• Price, standard — Carbon $4.00 / $20.00 after an unstarted introductory period. 3.1 Pro Preview $2.00 / $12.00, stepping to $4.00 / $18.00 above 200,000 input tokens.

• Cache — Carbon: 95% off cached input. 3.1 Pro Preview: $0.20 per million cached, a 90% discount.

• Availability — Carbon: Fairwind Program only, no endpoint. 3.1 Pro Preview: live in the catalogue today.

The input price is the one line where the two are level, and it is level for a reason. Google has held the Pro tier's opening input rate at $2 since February while the output rate drifted down from $12 to $10 under Argon and Carbon's introductory pricing. The model that is eight months older is now the more expensive one on output — until Carbon's introductory period expires and it doubles.

The score you will see quoted is not a comparison

Carbon's 59 leads the v5.0 board outright: Claude Opus 5.5 at 58, GPT-6 Astra at 57, Gemini 4 Argon at 50. Gemini 3.1 Pro Preview appears on that board only as an unranked entry: Artificial Analysis lists it with a dash where Carbon has 59. Artificial Analysis rebuilt the index on 27 October 2026, and v5.0 is not a rescoring of the old test with new names — it is a different set of nine evaluations, so a number carried over from a v4-era run cannot be placed next to a v5-era one, even when both are printed with the same word "index" in front of them.

That leaves three honest ways to compare, and only three:

• Same test, same run — available for Coding Agent Index and the arena-style benchmarks, where both models appear in a single sweep. Not available for the Intelligence Index.

• Same task, measured cost — Carbon comes in at $7.76 per index task at standard pricing. Gemini 3.1 Pro's cost per task is not published on the current board in a form that can be placed beside it.

• Published capability claims with their vintage attached — 3.1 Pro's strengths were measured in the first half of 2026 against a different field, and the field has moved three times since.

Where the comparison does hold: the invoice

Per-token pricing is the one axis where both models have a published, current number, and it produces a clear answer for anything that writes a lot. At introductory rates Carbon is $10 per million output against 3.1 Pro's $12 — 17% cheaper on the exact side of the bill that dominates agent workloads. At standard rates Carbon is $20 against $12, which is 67% more expensive. The whole question therefore turns on a date nobody has been given: when the introductory period ends.

Long inputs bend it further. Gemini 3.1 Pro Preview doubles to $4 input and $18 output once a single request passes 200,000 input tokens, so a 500k-token document-analysis job pays the higher tier on both sides. Carbon has no context figure in Google's own documentation — the 1M on its Artificial Analysis page is a third-party number — and therefore no published long-input tier.

The pass-through rule is what makes this legible on our side. One API for 200+ models at 0% markup means the price you see for Gemini 3.1 Pro Preview is Google's own $2 / $12 with its tiering intact, not a reseller's version of it, and any change Google makes to that card lands the same day. That is the practical value of a pass-through catalogue when a launch like Carbon arrives with a price that is scheduled to double: you are never comparing a cached price to a live one.

Eight months of drift, and what 3.1 Pro still does

The preview label is not a technicality. Gemini 3.1 Pro Preview was released on 19 February 2026, and Google spent the following eight months shipping the Flash tier — 3.5, 3.6, 3.8, and a cybersecurity-tuned 3.8 Flash Cyber — while the Pro slot stayed put. Carbon is the first thing to land in that slot since, and it lands on top of Argon, which itself landed on 30 September and became generally available on 13 October.

Against that drift, 3.1 Pro still holds three cards. Its context window is published and enormous — 1,048,576 tokens, roughly 1,500 pages of dense text — where Carbon's is simply unknown. Its output speed is measured at 134.4 tokens per second and its time to first token at 19.95 seconds, both on Artificial Analysis runs against Google's own API. Carbon, on the same board, emits about 85 tokens per second and takes 103.54 seconds to first token — a two-minute think before typing, which is what long-horizon reasoning looks like when a harness measures it rather than the marketing.

Those three cards do not make 3.1 Pro the better model. They make it the better known model, which is a different competition.

The workloads where the answer flips

There are three reader profiles in this matchup, and the model choice is different for each.

If you are shipping anything this month, the decision is already made for you: Carbon has no address. Gemini 3.1 Pro Preview's 1M context and multimodal input make it the right pick for document-heavy work where the reference material is the whole point — contracts, codebases, long transcripts — and its $12 output rate is defensible at moderate volume.

If you are building a coding agent, neither Google model is the answer on the current board. Carbon is second on the Coding Agent Index at 59, three behind Claude Opus 5.5 at 62, and it is unreachable anyway. 3.1 Pro is not competitive on agentic coding at eight months old; the Flash tier's newer iterations and other vendors' models are where that workload currently sits.

If you are doing security work — the exact domain Carbon was built for — the honest position is that the benchmark you would want to consult, SEC-bench Pro at 88.2%, is a vendor-reported number with no independent reproduction, because the only testers with access are inside Google's Fairwind Program. A model that is benchmarked but untested by outsiders is not yet evidence of anything except Google's own confidence.

What would change the answer

Three events would make this a real head-to-head rather than a comparison of a live model against an announced one. First, Carbon getting a model ID and a context figure, which would settle the long-input tier and make the cost question answerable. Second, a v5.0 Intelligence Index row for Gemini 3.1 Pro, which would let the 59 sit next to a same-ruler number instead of a footnote. Third, the end of Carbon's introductory window, which is the only published event that changes what either model costs.

Until at least one of those lands, the fair summary is this: Gemini 4 Carbon wins the index by one point over Claude Opus 5.5 and nine over Gemini 4 Argon, but the competitor named in this headline has no score on that board, is not on the same ruler, and is not in the same room. Gemini 3.1 Pro is slower to be replaced than it looks, because the model replacing it cannot be called — and the model that a developer can call today is the one that has been sitting in preview since February.