Nex-N2.5 Pro vs Gemini 3.1 Pro: A One-Day-Old Open Model vs a Seven-Month Preview

The strangest symmetry in this matchup is that both models are officially unfinished, and each side proves it in a way the other cannot. Gemini 3.1 Pro Preview, Google's closed multimodal flagship, has been in preview since February 19, 2026 — seven months, an eternity for a frontier model — and its maker has still not declared general availability. Nex-N2.5 Pro, the 397-billion-parameter computer-use model from the Nex-AGI open-model alliance, is one day old as of this writing and already answering traffic from free hosted endpoints — yet its Apache-2.0 weights are marked "coming soon" on Hugging Face, which makes it a different kind of unfinished. One is a finished product in every way except the vendor's willingness to say so. The other is an unfinished product in every way except the vendor's willingness to ship it.

Those two flavors of incompleteness change how every number below should be read. Gemini 3.1 Pro is Google's current flagship reasoning model — multimodal across text, image, audio, video and files, with a 1M-token context, a 65K output ceiling, and a price of $2.00 per million input tokens and $12.00 per million output that steps up once a request passes 200K input tokens. Nex-N2.5 Pro is a sparse mixture-of-experts model of 397B total with roughly 17B active, text-and-image in, text out, with a 262,144-token context and a vision loop built for operating computers and browsers. They occupy the same price-adjacent neighborhood of "premium agentic model," disagree about almost everything else, and neither has a vendor who will call it done.

The spec sheet: two multimodal 1M-class models

Put the envelopes side by side:

• Status — Nex-N2.5 Pro: announced September 8, 2026; hosted endpoints live, weights pending. Gemini 3.1 Pro: preview since February 19, 2026, still not GA.

• Scale — Nex-N2.5 Pro: 397B total / ~17B active MoE, post-trained from Qwen3.5 lineage. Gemini 3.1 Pro: parameter count undisclosed.

• Modality — Nex-N2.5 Pro: text and image in, text out; vision is the computer-use feedback channel. Gemini 3.1 Pro: text, image, audio, video and file input, text output.

• Context / output — Nex-N2.5 Pro: 262,144-token context, documented serving length. Gemini 3.1 Pro: 1M-token context, 65K output ceiling.

• Reasoning control — Nex-N2.5 Pro: none / medium (adaptive, default) / high. Gemini 3.1 Pro: adjustable thinking budget; no public "max" tier drama.

• Weights — Nex-N2.5 Pro: Apache-2.0, coming soon. Gemini 3.1 Pro: closed.

• Price per 1M tokens — Nex-N2.5 Pro: free hosted tier, no list price published. Gemini 3.1 Pro: $2.00 / $12.00, stepping to $4.00 / $18.00 above 200K input tokens, $0.20 cached input.

The two spec sheets agree on very little beyond "long context, multimodal, expensive to build." The differences that matter are the 200K input surcharge on the Gemini side, the free-but-unverifiable economics on the Nex side, and the fact that only Nex-N2.5 Pro is architected around reading a screen.

Two different ways to be unfinished

Google: seven months of preview as a product decision

Gemini 3.1 Pro's preview status is not cosmetic, and it is increasingly the most important fact about the model. Google's Pre-GA terms allow production use while reserving the right to change behavior or deprecate the endpoint — a real risk for a team that puts a flagship on a critical path. The model's trajectory has done nothing to retire that risk: it launched in February as the highest-scoring model on the market, 57 on the Artificial Analysis Intelligence Index at the time, and has since been re-measured on a stricter index scale to roughly 48, while its intended successor, Gemini 3.5 Pro, slipped from a May announcement into repeated delays and reports of possible cancellation. In the meantime Google shipped Gemini 3.7 Flash in August at $0.75 / $3.75 and Gemini 3.8 Flash in early September, and the cheaper Flash line has climbed past the Pro flagship on the independent index. Adopting Gemini 3.1 Pro today means adopting a flagship whose vendor keeps it in preview while its cheaper siblings outrank it and its successor is in limbo.

Nex-AGI: a day of "coming soon" as a delivery state

Nex-N2.5 Pro's incompleteness runs in the opposite direction: everything is shipped except the thing that would make it real. The hosted endpoints are live and free, the OpenAI-compatible calls work, and the model card is filled with benchmark figures — OSWorld-2 at 56.4, Terminal-Bench 2.1 at 82.7, SWE-Bench Pro at 61.2, all measured through the alliance's own NexCUA harness, which it says it will open-source but has not. But the weights are not downloadable, no release date is attached to the banner, and no independent laboratory has run the model, because none can. Where Gemini 3.1 Pro is measurable-but-uncommitted, Nex-N2.5 Pro is committed-but-unmeasurable. Both states should give a production team pause, for different reasons.

Where the independent evidence lives

The independent record in this matchup belongs almost entirely to one side. Gemini 3.1 Pro has six months of public testing behind it: independent labs have benchmarked it, production teams have hammered it, and it carries a current Artificial Analysis Intelligence Index reading near 48 — modest for a flagship, and the reason Google's cheaper Flash models have overtaken it, but an actual measurement produced by someone other than the vendor. Nex-N2.5 Pro has no such entry. Its only scores are Nex-AGI's own, on a harness the public has not seen. The comparison that matters is therefore not "48 vs nothing" — it is "a model whose weaknesses are documented" versus "a model whose strengths are asserted." For a workload where a wrong answer is expensive, that asymmetry is often decisive regardless of the numbers.

What the long-context bill looks like

Both models advertise long context, but they charge for it in ways that reveal their intent. Gemini 3.1 Pro's 1M-token window carries a $2.00 / $12.00 rate that steps to $4.00 / $18.00 once a request exceeds 200K input tokens — a roughly 60% input premium and 50% output premium applied precisely to the workloads the 1M window exists for. Feed it a 500K-token repository context on every turn of an agent session and that premium lands on every single call. Nex-N2.5 Pro's 262K context is more modest, but its hosted tier is free and Nex-AGI has published no paid rate at all, so there is no long-context premium to model — there is also no price signal for what the model is worth. Google is telling you exactly what its window costs, and the price is high; Nex-AGI is telling you nothing, which is not the same as cheap.

What you can route today

Here the two sides separate cleanly for a working team. Gemini 3.1 Pro is on OrcaRouter at Google's list price — the same $2.00 / $12.00 with the over-200K step passed through unchanged — and its preview status is exactly the argument for routing around it rather than betting a critical path on it: automatic failover means a preview-model hiccup or deprecation redirects to a fallback without a code change, and the routing DSL lets you send the requests that need Google's multimodality to Gemini while cheaper models absorb the routine traffic. Nex-N2.5 Pro is not on OrcaRouter — no nex-agi model is listed in our directory — because its only live builds are the free endpoints Nex-AGI links from its card. The day a provider serves it at a real list price, it will appear here at that price with no markup, and this comparison becomes runnable rather than theoretical.

Who should touch which

The honest answer is shaped by which flavor of risk you can absorb. If you need multimodal reasoning at scale today and can tolerate a vendor that keeps its flagship in preview while its roadmap shifts, Gemini 3.1 Pro is the measurable choice — six months of independent testing is a real asset, and the Flash siblings' rise is a reason to route rather than a reason to avoid the family. If you are building computer-use agents on open weights and are willing to wait for something verifiable, Nex-N2.5 Pro is the more interesting long-term bet — but it is not yet a bet you can place, because you cannot run it, reproduce its scores, or sign a contract that depends on it. The teams that will be happiest with either decision are the ones that treat both models as what they are: a seven-month preview that has been tested by everyone, and a one-day-old preview that has been tested by no one.