Hero title card reading Gemini 4 Neon vs Gemini 4 Pro, subtitle reading one of these models has no endpoint, footer line reading endpoint list per Google's Gemini API model index, prices per the Gemini API pricing page, checked 20 November 2026
Guides & Insights

Gemini 4 Neon vs Gemini 4 Pro: The Pro Slot Is Still Gemini 3.1

Author

Elias Hawthorne

Date Published

Latest models · 20View all models →
Benchmarks: Artificial Analysis · updated daily
Back to all posts

If you are choosing between G​emini 4 Neon and G​emini 4 Pro, the first thing to know is that one of those two models does not exist as a product. G​oogle's Gem​ini API model index lists four G​emini 4 identifiers — Carbon, Neon, Hydrogen and Argon — and no fifth. There is no gemini-4-pro endpoint, no rate card, no model card, and the per-model documentation path under that name returns nothing. What the Pro slot actually holds today is Gemini 3.1 Pro Preview, still labelled a preview, priced at $2.00 input and $12.00 output per million tokens. So the decision a reader is really making is between G​emini 4 Neon, the new high-frequency frontier model at $3.00 and $15.00, and the three-generation-old Pro model that G​oogle still sells. That is a more interesting comparison than the one the query asks for, and the answer is not the obvious one.

What "Pro" resolves to in this family

G​oogle's pricing page is the cleanest evidence, because a pricing page cannot list a model it cannot bill. Its G​emini 4 sections are Carbon, Argon, Neon and Hydrogen, in that order, and the only Pro entry anywhere on it is Gemini 3.1 Pro Preview — under its own heading, with its own rate card, and with the note that it can be tested in G​oogle AI Studio. The Gem​ini API's model index, which lists every identifier you can call, has no Pro entry in the G​emini 4 block at all: the four names there are Carbon, Neon, Hydrogen and Argon, each with its endpoint. The only Pro identifier the API documents is gemini-3.1-pro-preview, under a heading that still says Preview.

So the Pro tier is not missing so much as frozen. Every other tier moved into the G​emini 4 generation this autumn: Argon at the end of September, Carbon in October, Hydrogen on 12 November, Neon on 18 November. The Pro slot stayed on 3.1, in preview, at a price that has not changed. There is no published date for anything else arriving there.

Neon's spec sheet against the model the slot holds

Set side by side, the two models are not in the same weight class on paper, which makes the price relationship the surprising part.

• Price — G​emini 4 Neon $3.00 input and $15.00 output. Gemini 3.1 Pro Preview $2.00 and $12.00.

• Cached input — Neon $0.15 per million, a 95% discount. Gemini 3.1 Pro Preview $0.20 per million, a 90% discount.

• Context and output — Neon one unified 1,048,576-token window shared across input, thinking and output, 65,536 synchronous output with long decode to a million. Gemini 3.1 Pro Preview 1,048,576 input with a hard 65,536 output cap and no long-decode path.

• Thinking control — Neon defaults to high and supports low, medium, high and xhigh. Gemini 3.1 Pro Preview defaults to high and supports low, medium and high.

• Caching model — Neon explicit caching unsupported, implicit only. Gemini 3.1 Pro Preview supports explicit caching, including cachedContents with a storage price.

• Measured intelligence — Neon 59.6 at high and 61.4 at xhigh on Artificial Analysis Intelligence Index v5.0.1. Gemini 3.1 Pro Preview has no index score on that revision at all.

The last line is the one to sit with. The tracker that scores almost everything has an entry for Gemini 3.1 Pro Preview with a price, a speed and a context length, and no intelligence number — because the index v5.0.1 battery was never run on it. You cannot claim Neon beats it by nine points, or by any points, because the comparison does not exist. What you can say is that nobody has published a measurement that puts the older model ahead.

The one place the old Pro model clearly wins

Latency to first token. Artificial Analysis measures Gemini 3.1 Pro Preview at 15.02 seconds to first answer token and 108 output tokens per second, against G​emini 4 Neon's 56.51 seconds and 110 tokens per second. Generation rate is a dead heat; the wait before the first token is nearly four times shorter on the 3.1 Pro model.

That is not a trivia item. For an interactive product where a user watches a cursor, 56 seconds of silence is a different product from 15, and no amount of extra index points fixes it. It is also the one quality the G​emini 4 family as a whole has not improved: Hydrogen sits at 47.2 seconds, Argon at 46.0, and Carbon at 74.5. Every G​emini 4 model is slower to start than the preview G​oogle shipped three generations ago.

The other genuine advantage is explicit caching. If your workload has a large, stable prefix — a codebase, a contract set, a document corpus — and you want to control its lifecycle and pay storage for it, Gemini 3.1 Pro Preview gives you a handle to create and expire, and Neon does not. Neon's implicit caching is automatic and cheaper per cached token when it hits, but you cannot pin a prefix and reason about its state.

Where the extra dollar buys something

Three things, and they are all Neon's own claims rather than measured facts, so label them accordingly. G​oogle says Neon completes repository-level debugging and multi-file refactoring in fewer steps with cleaner edits, that across long sessions with many tool calls it reduces redundant file reads and handles context compaction reliably, and that it minimises fabricated claims and task shortcuts. Those are vendor statements about its own model, with no independent replication, and the same page that makes them also positions Neon specifically for high-frequency complex tasks rather than for the hardest problems — where Carbon is still the recommendation.

What is measurable and does favour Neon is the shape of the work it was built for: one unified window where input, thinking and output share a budget, a long-decode path that keeps generating past 64K, and a thinking ladder that reaches a rung the 3.1 Pro model does not have. A 3.1-era agent that wants to write a 200K-token migration document has to stitch it together client-side. Neon does it in one job.

Screenshot of Google's Gemini API model documentation listing all Gemini 4 models in a two-column table of model name and endpoint, showing Gemini 4 Carbon against gemini-4-carbon, Gemini 4 Neon against gemini-4-neon, Gemini 4 Hydrogen against gemini-4-hydrogen and Gemini 4 Argon against gemini-4-argon, with no Pro rowA two-column comparison scoreboard for Gemini 4 Neon and Gemini 3.1 Pro Preview, with Neon showing 59.6 at high on the Intelligence Index, 3 dollars and 15 dollars per million tokens, 56.5 seconds to first token, 110 tokens per second, a unified 1M window with long decode and four thinking levels, and the Pro Preview showing no index score published on this revision, 2 dollars and 12 dollars, 15 seconds to first token, 108 tokens per second, a 65536 output cap and three thinking levels

How to decide, given one of them is not a G​emini 4

If your reason for wanting "G​emini 4 Pro" was the word Pro — a general-purpose workhorse one rung under the flagship — the model that occupies that role in the current line is G​emini 4 Neon, and the one that occupies it in capability terms is Gemini 4 Carbon. G​oogle's own choosing-a-model page routes hard architecture, debugging, science and security problems to Carbon; Neon is for the high-frequency version of the same work. Neither is a Pro refresh, because G​oogle has not shipped one.

If your reason was price, Gemini 3.1 Pro Preview is still the cheaper of the two at $2.00 and $12.00, and it remains callable. That is a legitimate reason to stay, and it will stay legitimate until G​oogle either retires the preview or prices a G​emini 4 Pro model — neither of which has a published date.

If your reason was that you need a G​emini 4 model with a Pro-shaped capability profile, the honest answer is that the family does not contain one. It contains a maximum-capability model, an efficient one, a cheap one and a fast one, and the name you were searching for is the tier G​oogle skipped.

Calling either of them from one place

Both paths are live on our side of the fence in one respect and not the other: gemini-3.1-pro-preview and Gemini 4 Carbon are both in the OrcaRouter catalogue, and G​emini 4 Neon is not yet, so on this page we are describing a model you would call on G​oogle's own API. What the router changes about this specific decision is that the fallback is free to configure. If you move interactive traffic onto Neon and discover that 56 seconds to first token is too long for your UI, the same request can fail over to the cheaper, faster 3.1 Pro Preview without a second integration, and the price on your bill is each provider's list price passed through with nothing added on top.

That is the practical version of the answer to "Neon or Pro". The two models are not competitors so much as different bets — one on capability, one on latency and price — and the decision is easier to reverse if you never hard-code the model string in the first place.

Screenshot of the OrcaRouter model page for google/gemini-3.1-pro-preview, showing the preview model card with its context window and per-million-token pricing in the OrcaRouter catalogue

What would settle it

Two events, both visible from outside. If Artificial Analysis runs its v5.0.1 battery on Gemini 3.1 Pro Preview, the matchup gets the number it is missing, and no one outside the tracker knows where it will land — this page will not guess. And if G​oogle ships a G​emini 4 identifier with Pro in its name, the query that brought you here starts having a second real answer. Until then, the only Pro model with an endpoint is the pre-Gemini-4 one, and the comparison is between a new model and a held-over one.

Compared in this article1

Detected from this article · Benchmarks: Artificial Analysis · updated daily