Gemini 4 vs Gemini 3.1 Pro: A Model You Can Call Today Against One You Cannot

The short answer to Gemini 4 vs Gemini 3.1 Pro is not about capability, and it is not about price — the two are within a few cents of each other on input. It is about a badge. Gemini 3.1 Pro Preview has been generally reachable from G​oogle's API since 2026-02-19 and G​oogle has never marked it Stable; the endpoint still ends in -preview, seven months later. Gemini 4 Argon, announced 2026-09-30 and moved to general availability on G​oogle's paid tier on 2026-10-13, carries G​oogle's Stable badge and a published spec sheet — and it is not in our catalogue, while the preview model is.

That is the whole shape of this matchup. One of these two is a model you can put in a request through this platform today; the other is the current flagship of the same vendor's line, priced and documented, and reachable through G​oogle's own API and several third-party platforms. Notice also what the comparison is not: both names end in "Pro" or sit in the Pro slot, but only one of them is a Gemini 4 model. Everything in generation 4 is Argon.

The seven-month preview, read as a fact rather than a complaint

G​oogle announced Gemini 3.1 Pro Preview on 2026-02-19. In the seven months since, the company shipped Gemini 3.5 Flash, 3.6 Flash, 3.7 Flash and 3.8 Flash, and the model Gemini 3.1 Pro replaced — Gemini 3 Pro — is listed as shut down. The Pro slot moved zero times. G​oogle's own site still serves the Gemini 3.1 Pro page under a 3.1-generation title while the Flash slot has advanced four releases.

For a reader deciding what to build on, the preview badge is a statement about commitments rather than about quality. A preview release is a commitment to serve the model today and no commitment about tomorrow. It is not a prediction of shutdown — Gemini 3.1 Pro Preview has run unchanged and correctly priced for most of a year, which is more stability than many GA models manage. It is simply the absence of a guarantee, and it is why the same model that looks well-supported can still be the riskier long-lived dependency of the two.

The rate cards, which are closer than the gap in capabilities

Both figures below are the vendors' own published rates and our pass-through of them, per million tokens, as read on 2026-10-14:

• Input — Gemini 4 Argon $2.00 introductory, stepping to $4.00 on 2026-12-01, vs Gemini 3.1 Pro Preview $2.00 up to 200,000 input tokens, then $4.00. A tie at the base tier, and a tie again above 200K — one because it stepped up, the other because it started there.

• Output — Gemini 4 Argon $10.00 introductory, stepping to $20.00 on 2026-12-01, vs Gemini 3.1 Pro Preview $12.00 up to 200K tokens, then $18.00. Argon is cheaper today and dearer after December, and the two cross over on long prompts.

• Cached input — Gemini 4 Argon $0.10 vs Gemini 3.1 Pro Preview $0.20. Argon is cheaper on the line that dominates long agent runs, and it supports implicit caching only.

• Cache write — Gemini 3.1 Pro Preview publishes $0.375 per million. G​oogle publishes no separate cache-write rate for Argon because explicit caching is not offered.

• Context — 1M tokens on both, though the published figures differ slightly: 1,000,000 for Argon, 1,048,576 for Gemini 3.1 Pro Preview.

• Maximum output — Gemini 4 Argon 1,000,000 tokens vs Gemini 3.1 Pro Preview 65,536. This is the largest single difference between the two and it is not close, in either direction: a 65K decode ceiling is a real constraint for long-form generation, and a 1M ceiling makes Argon's worst case far more expensive on the same input.

• Modality — Gemini 3.1 Pro Preview accepts text, image, audio, video and PDF input; Gemini 4 Argon is published as text and image on the independent boards. The older model is the broader reader, which is worth knowing before assuming a generation number settles it.

If you are running both, the practical detail is that Gemini 3.1 Pro Preview is served here at list price with no markup added, so the vendor's own arithmetic is what lands on your invoice. Argon is not hosted by us as of 2026-10-14, and no amount of routing makes it so — the routing layer moves traffic between providers of a given model, and it cannot conjure one we do not carry.

What the independent boards say, and where they stop being useful

Artificial Analysis measures Argon at high effort at 52.56 on Intelligence Index v4.3.2, the same revision on which Claude Opus 5.5 reads 57.62 and GPT-6 Astra reads 52.67. Its reading for Gemini 3.1 Pro Preview is the harder number to state honestly, because the index was revised underneath the model and the figure moved with it. It launched near 57, and the current reading sits far below that — our own two pages that quote it carry different values taken at different revisions, one at 29.7 and one in the upper forties. When two of our own pages disagree about a number, the number is not the point; the disagreement is. Anyone quoting a single Gemini 3.1 Pro Preview index score without the revision date attached is quoting a measurement that has already been replaced.

What does survive the revision churn is cost per task on the evaluator's own harness: Argon at $1.99 per index task against Gemini 3.1 Pro Preview at $0.67. The older model is three times cheaper per task and it is not three times worse — it is answering a smaller, easier question that costs the harness less to run.

The rule

If what you need is a frontier Pro-tier model on a key you can call today, with a hard deprecation conversation you can have with your own team, and your work fits inside a 65,536-token decode with audio, video or PDF on the input side — Gemini 3.1 Pro Preview is the reasoned choice, and it is the half of this matchup we host. If you are standardising on G​oogle's current flagship, need the 1M-token decode ceiling, and can accept a rate that doubles on 2026-12-01 plus access that is not yet available through us, Gemini 4 Argon is the successor in every sense except the badge on it.

The one fact that most often decides this in practice is that only one of these two names can be pasted into a request today from the platform you are reading this on — and that fact is about our catalogue and G​oogle's release process, not about which model is stronger.

Where each of these is reachable

Gemini 3.1 Pro Preview is served through OrcaRouter at Google's list price with 0% markup on top, so the arithmetic on this page is the arithmetic you are billed, and the same key gets you automatic failover across providers of that model. Gemini 4 Argon is not carried there as of 2026-10-14; it is reached through G​oogle's own API and several third-party platforms.

The Pro slot has not moved in seven months, but the Flash rung has. Gemini 3.8 Flash It is on the same key, which is what makes the substitution test cheap to run.

A model choice is also a question of what has gone wrong before: the documented AI incidents and agent failures for models of this class are collected in the OrcaRouter AI Incident Archive , where each entry carries its own sources, severity, confidence grade and disputed flag.