Mercury 3 vs Gemini 4 Pro: One of These Names Is Not a Model

There is no G​emini 4 Pro. There is a G​emini 4 Argon, which G​oogle made generally available on its paid API tier on October 13, 2026, and there is Mercury 3, which Inception Labs released on October 19, 2026 — and a great many people are searching for a comparison between Mercury 3 and a G​emini 4 Pro that G​oogle has never announced, never given an identifier, never priced and never published a context length for. This piece takes the query seriously anyway, because the question underneath it is real: how does a 1,068-token-per-second diffusion model from a small lab compare with G​oogle's current flagship on the things you can actually measure? The answer runs against G​emini 4 Argon, the model that exists, and the first finding is that the two are not in the same market at all. The documented AI failures and agent incidents behind releases like this one are collected in the OrcaRouter AI Incident Archive where each entry carries its own sources, severity, confidence grade and disputed flag.

How the missing name checks out

The absence is not an inference from silence. It survives five separate checks.

• G​oogle's G​emini API model documentation lists exactly one model under its G​emini 4 heading — gemini-4-argon — and the page was last updated on October 19, 2026, one day before this article. A second G​emini 4 card has not followed it.

• The DeepMind models page carries the Argon card and nothing else in the G​emini 4 line.

• G​oogle's own blog has no G​emini 4 Pro announcement, and the Argon release post describes Argon as the flagship of the G​emini 4 family, not as a member of a tiered line with a Pro above or below it.

• There is no G​emini 4 Pro model identifier anywhere in the API documentation, no rate card, no context length, no output ceiling and no licence term. A model with none of those things is not a model that has been released quietly; it is a name people have assumed will exist.

• It is not on our catalogue either. The G​oogle models OrcaRouter routes are the 3.x line — gemini-3.1-pro-preview, gemini-3.8-flash, gemini-3.6-flash, gemini-3.5-flash and the lite variants — plus nothing from the 4 line. We are not withholding a route; there is no route to withhold.

If a G​emini 4 Pro ships later this year, the structure of this comparison probably survives, because the price point it would occupy is already taken by Argon at the top of the line. But the numbers below belong to G​emini 4 Argon, and they are labelled as such.

What G​emini 4 Argon is, since it is the thing being compared

G​emini 4 Argon was announced on September 30, 2026 and reached general availability on the paid tier on October 13, 2026. It is multimodal on both ends of the call — text, images, video, audio and PDF in, text out — with a 1M-token unified context, a 65,536-token default output ceiling that extends to 1,048,576 tokens in long-decode mode. Pricing is an introductory $2.00 per million input and $10.00 per million output, rising to $4.00 and $20.00 on December 1, 2026, with cached input at $0.10 doubling to $0.20 on the same date. Caching is implicit only; there is no explicit cache tier to buy. It is not on G​oogle's free tier.

On Artificial Analysis's Intelligence Index it scores 52.6, at a compute cost of about $1.99 per index task. G​oogle's own published figures include 57.1% on Terminal-Bench 4.0, 61.8% on SciCode and 57.1% on Humanity's Last Exam — vendor-reported, and labelled as such, but at least stated on a named harness with named tasks.

So the model on the other side of this comparison is a multimodal flagship with an independent index score of 52.6 and a price that doubles in six weeks. Mercury 3 is a text-only diffusion model with no index score and a price that halves at launch. Almost every dimension of the comparison is a contrast between two different theories of what a language model is for.

The spec contrast, on the dimensions that decide deployments

• Output price — Mercury 3 at $0.35 per million at launch and $0.70 at list, vs G​emini 4 Argon at $10.00 now and $20.00 from December 1, 2026.

• Input price — $0.10 launch and $0.20 list, vs $2.00 now and $4.00 after December 1.

• Cached input — $0.01 launch and $0.02 list on Mercury 3, vs $0.10 now and $0.20 after December 1 on Argon, implicit caching only.

• Context — 1,000,000 tokens on both, though Argon's is unified across modalities and Mercury 3's holds text alone.

• Maximum output — 131,072 tokens on Mercury 3, vs 65,536 by default on Argon and up to 1,048,576 in long-decode mode.

• Input modalities — text only vs text, image, video, audio and PDF.

• Throughput — 1,068 tokens per second vendor-reported on Mercury 3, vs no comparable single-stream figure published for Argon.

• Independent index score — none for Mercury 3, vs 52.6 for Argon at about $1.99 of compute per index task.

• Price direction — Mercury 3's discount has no published end date; Argon's increase is scheduled for December 1, 2026 and applies to the whole request once the input crosses the long-context boundary.

The price line is the one that will decide most real deployments, and the December 1 date makes it urgent in a way the Mercury launch does not. A team that builds a document pipeline on Argon at $2.00/$10.00 and re-runs the cost model in January will find the input rate doubled and the output rate doubled with it. That is not a subtle risk; it is a calendar entry.

Where the price comparison breaks down completely

G​emini 4 Argon takes PDFs, screenshots, audio and video. Mercury 3 takes text. That single line eliminates the price comparison for a large share of the workloads where Argon is actually used — invoices, scanned forms, recorded calls, screen recordings, anything where the input arrives as pixels or waveform rather than characters.

Getting a PDF into Mercury 3 means running OCR or a layout model first, and that step has its own cost, its own latency and its own error rate. A 40-page scanned contract does not become free because the model that reads the extracted text is cheap. On the workloads where Argon's multimodality is load-bearing, Mercury 3 is not a cheaper option — it is not an option, and the honest comparison stops there.

Where the modalities do line up — text in, text out, and a great deal of production traffic is exactly that — the price gap is roughly 14 times at Argon's current output rate and 28 times after December 1. That is large enough to change an architecture. A retrieval pipeline that currently sends twenty candidate passages to a flagship for re-ranking could send a hundred to Mercury 3 for the same money, and the extra context may well be worth more than the per-token quality difference.

The evidence asymmetry, and why it matters more here

G​emini 4 Argon has an independent measurement and a vendor benchmark table. Mercury 3 has neither. What Mercury 3 has is a throughput figure Inception measured on unspecified NVIDIA hardware, a claim of more than 40x the parameters of Mercury 1, and one structured quality result: 50% completion on 50 live Mind2Web navigation tasks inside Inception's own agent harness, which the company says is "significantly more than GPT-6.1-Sol" at the same token volume and step count.

That is a genuine result — live web navigation is hard to game — but it is not comparable to a 52.6 on a neutral index. The two numbers answer different questions: one says "this model can complete half of fifty real navigation tasks"; the other says "this model scores 52.6 on a fixed, published suite that every other model in the index has also taken". You can rank the second against alternatives. You can only sanity-check the first.

There is also a calibration issue hiding in the comparison. Mercury 3 defaults to the lowest of its four effort settings when the parameter is omitted, so any evaluation that does not pin effort is measuring the model at its least capable setting. Argon has no equivalent dial. A benchmark run of both, done carelessly, would make the gap look larger than it is in either direction — and Inception's own comparative claims do not state which setting produced them.

The two models are not on the same side of a price change

Put the two rate cards next to each other and the direction of travel is opposite. Mercury 3 launched with a 50% discount that has no end date, which means a builder cannot tell whether the rate in six months is $0.10/$0.35 or $0.20/$0.70. Argon launched at an introductory rate with a hard end date, which means a builder knows exactly what the rate is on December 2: double.

Certainty about a bad number is often worth more than uncertainty about a good one, and this is a case where the two models give you one of each. If you are budgeting a pipeline for a year, Argon's $4.00/$20.00 is the number to model against from December onward, and Mercury 3's $0.20/$0.70 list is the number to model against because the discount is not a rate card. Even at those two conservative numbers the gap is 20 times on output, which is still the largest single fact in this comparison.

What no rate card tells you is what a task costs. Argon's index run costs about $1.99 of compute per task; Mercury 3 has no equivalent measurement, and its token economy under diffusion decoding is not the same shape as an autoregressive model's. Anyone quoting a per-task cost for Mercury 3 right now is extrapolating from a per-token rate and a guess about output length.

How to hold both without committing to either

G​emini 4 Argon is not on the OrcaRouter catalogue. What we route from G​oogle is the 3.x line — gemini-3.1-pro-preview at $2.00/$12.00, gemini-3.8-flash and gemini-3.6-flash at $0.75/$3.75, gemini-3.5-flash at $1.50/$9.00 and the lite variants — and Mercury 3 is not on the catalogue either, for the simpler reason that we do not carry any Inception model. So for this particular matchup, our honest position is that we are not the route for either side of it.

Where a router is still useful is in the part of the architecture that surrounds the choice. The 3.x models we do carry can serve as the fallback for either: route the bulk of the text work to Gemini 3.1 Pro or Gemini 3.8 Flash at their list prices with no markup added on our side, and send only the calls that genuinely need the flagship or the diffusion model to their respective vendors. That keeps one endpoint, one key and one set of credentials for the majority of traffic, and confines the two special cases to two narrow code paths — which is a much smaller commitment than picking one of them as the default for everything.

The pass-through pricing matters for Argon specifically. When December 1 arrives and G​oogle doubles the rate, a routed model's price changes the day the vendor changes it, with no renegotiation and no stale rate card in your cost model. We do not route Argon today; if that changes, the increase lands on the same schedule Google publishes rather than on ours.

What to do with a comparison to a model that does not exist

The practical advice is dull and worth stating. If a procurement document, an internal proposal or a vendor deck in front of you cites G​emini 4 Pro, ask for the model identifier in the API. Not the marketing name, not the family name, not the tier — the string you would put in a request body. If the answer is gemini-4-argon, you are looking at Argon with a different label on it, and the rate card is $2.00/$10.00 now and $4.00/$20.00 from December 1. If there is no answer, the document is citing a name and not a product.

The same test applies to Mercury 3 in the other direction. Its identifier is real, its rate card is published, its context length is published, and its throughput figure is a vendor claim. Between a model with a full spec sheet and no independent score, and a model with an independent score whose name people keep getting wrong, the second is still the safer default — not because G​oogle is a safer vendor, but because a number you did not have to take on faith is worth more than a discount you cannot verify. Until Inception's model has an index page, the choice between these two is a choice between evidence and price, and most teams should buy the evidence and rent the price.