GPT-Rosalind vs Gemini 4 Argon: One Gate You Can Pass, One With No Door Yet

GPT-Rosalind and Gemini 4 Argon are both frontier models aimed at scientific work, both announced as gated releases, and both impossible to evaluate the way you would evaluate a normal model release. The difference is what the gate is made of. GPT-Rosalind has been purchasable since 5 October 2026 at $5.00 per million input tokens and $25.00 per million output, to any organisation that clears O​penAI's trusted-access review — a real door, with a real queue in front of it. Gemini 4 Argon, announced on 30 September 2026, has a score, an introductory rate and a three-stage rollout, and no model identifier you can send a request to. One is gated. The other is described. Gated rollouts and launch dates that slip after publication are logged with their dates and sources in the AI Incident Archive, next to the other model and vendor incidents we track.

The test that separates a model from a name

Four things have to be true before a model is something a team can actually plan around: there is an identifier you can call, there is a price you can be billed against, there is a documented context window, and there is a date by which general access arrives. Run both models through those four and the comparison stops being about capability.

• Callable identifier — GPT-Rosalind: yes, through the API for approved organisations. Gemini 4 Argon: no. G​oogle's API model list contains no gemini-4 or argon entry, and the newest identifiers the documentation recognises are in the 3.8 family.

• Billable price — GPT-Rosalind: $5.00 / $25.00 per million, published and in force since 5 October. Gemini 4 Argon: $2.00 / $10.00 per million quoted as an introductory rate for a rollout that has not reached paying customers.

• Documented context — GPT-Rosalind: 1M tokens. Gemini 4 Argon: G​oogle published a 1M-token output ceiling, up from a previous 64K limit, and did not publish the input side; Artificial Analysis records 1M for the configuration it measured.

• General-access date — GPT-Rosalind: open globally to eligible organisations since 11 September 2026. Gemini 4 Argon: no date on any rollout stage past the first.

Two of four for Argon, four of four for Rosalind. That is not a knock on Argon's engineering; it is a statement about what a procurement team can sign.

What Argon is actually rolling out to

The announcement, credited to Koray Kavukcuoglu, describes three stages with no durations attached. First, a named cohort of cyber defenders through the Fairwind Program. Second, paid API customers and G​oogle AI Ultra subscribers. Third, developers, enterprises and consumers. Only the first has any observable members, and the programme is explicitly a cyber-defender cohort — which is a curious first home for a model whose published strengths are software engineering, legal and finance work, and scientific reasoning.

Compare the sequence O​penAI ran. GPT-Rosalind began in April 2026 with a named set of US trusted-access partners — Amgen, Moderna, the Allen Institute, Thermo Fisher — moved to global eligibility for approved organisations in September, and only then published a price. The gated cohort came first, then the general gate, then the invoice. G​oogle has done the first step and stopped.

Two scores that are not the same kind of number

This is where most comparisons go wrong, so it is worth being blunt. The number attached to Gemini 4 Argon — 52.56 on the Artificial Analysis Intelligence Index, v4.3.2 revision — was produced by an independent evaluator on a general composite of ten evaluations. It is a real measurement of a real configuration, and it is the only independent number either model has. The numbers attached to GPT-Rosalind come from O​penAI's own life-sciences suite, and no outside lab has reproduced any of them.

• Independent composite — Gemini 4 Argon 52.56 on AA v4.3.2, measured at one thinking level. GPT-Rosalind: not tracked on that board at all, because specialist models do not appear on the general leaderboard.

• Domain benchmark — Gemini 4 Argon: G​oogle's table leads with Harvey's Legal Agent Benchmark at 19.6 against a competitor's 3.8, and Vals Finance Agent v2 at 65.4. GPT-Rosalind: MedChemBench 27.5% vs 25.1% for GPT-5.5, GeneBench 21.6% vs 20.4%, LabWorkBench 63.2% vs 55.8%, all vendor-run.

• What the domain numbers have in common — Neither set has been independently re-run, and both vendors chose the benchmarks. The two largest relative margins in G​oogle's table are also the two rows most likely to be re-tested by someone who does not work for G​oogle.

So you cannot compare 52.56 to 27.5, and you cannot compare 52.56 to anything Rosalind has published, because the two models have never sat the same exam. What you can say is narrower and more useful: Argon has been measured by someone with no stake in the result, on tasks that have nothing to do with biology, and it landed mid-frontier. Rosalind has been measured by its maker, on biology, and the gains are modest in absolute terms — a 27.5% score on a medicinal-chemistry benchmark means both models fail most of it.

If Argon ships at the introductory rate, the price argument inverts

GPT-Rosalind at $5.00 input and $25.00 output is the most expensive model in this comparison by a wide margin, and it is the one with a live invoice. Argon's quoted $2.00 / $10.00 would undercut it by 60% on input and 60% on output — if that rate survives to the API stage. That is a large if. Introductory pricing on a staged rollout is a signal about intent, not a commitment, and G​oogle has attached no date to the stage where anyone would pay it.

• Input — GPT-Rosalind $5.00 per 1M, billable now vs Gemini 4 Argon $2.00 quoted, not billable.

• Output — GPT-Rosalind $25.00 per 1M vs Gemini 4 Argon $10.00 quoted.

• Cost per index task — Gemini 4 Argon $1.99, per Artificial Analysis, which also records a cache discount that brings effective input near $1.99 per million. GPT-Rosalind has no equivalent figure because it has no index score.

• Cache handling — O​penAI charges $0.50 per million cached input on Rosalind and publishes no cache-write line. G​oogle has not published a cache rate for Argon at all.

There is a version of this comparison, arriving whenever stage two begins, in which Argon is a general model at less than half the specialist's price with an independent score attached, and the biology premium collapses. That version does not exist yet. Planning against it today means planning against a press release.

Where the biology argument actually lands

Argon's announcement does not claim biology work. It claims software engineering, enterprise knowledge work in legal and finance, and cybersecurity defence. The scientific framing that surrounds GPT-Rosalind — the wet-lab protocols, the genomics, the medicinal chemistry — has no counterpart in G​oogle's material for this model. So the question "which is better for a research lab" has an unglamorous answer: only one of them has been tuned for it and published numbers on it, and the other has not entered the category.

That could change with a model card. A context window, a parameter disclosure, a scientific evaluation row — any of those would flip the reading inside a day. Until then, the honest summary is that these two models do not compete. One is a specialist with a gate you can clear. The other is a generalist with a score and a queue you cannot yet join.

What you can call today, and what that costs you

Neither model is on OrcaRouter. GPT-Rosalind is not on any third-party router and runs through O​penAI's trusted-access programme directly; Gemini 4 Argon has no endpoint to route to. Saying otherwise would be the easiest lie in this article, so we are not saying it.

What is on our routes is the part of the work that surrounds the science. Gemini 3.8 Flash is there at $0.75 input and $3.75 output per million tokens with a 1,048,576-token context — the model Argon will be compared against by anyone who owns a credit card, and the closest thing to an Argon-shaped workload you can run this week. Claude Opus 5.5 is there at $4.00 / $20.00 for the writing, the code review and the long agent loops. Both sit behind one key and one bill alongside 200+ other models, at provider list price with 0% markup, so a vendor price cut is live on our side the same day it is published rather than whenever a contract renews.

For the biology half, the architecture is a direct key and a named route, not a guess. If a request needs the specialist, it should fail loudly when the specialist is unavailable rather than quietly landing on a general model — which is exactly the failure mode that a fallback chain configured for cost will produce if nobody has drawn the line explicitly.

The tell to watch

One event settles this comparison, and it is not a benchmark. It is a gemini-4 identifier appearing in G​oogle's API model list. The moment that happens, the introductory rate stops being a figure in a blog post and becomes something you can be billed against, the context question gets an answer, and a two-model comparison becomes possible for the first time.

Until then, treat Gemini 4 Argon as a name with a score attached and GPT-Rosalind as a specialist with a queue. If your work is molecular biology and you can pass a review, the specialist is available now and the premium is the price of being early. If your work is anything adjacent — pipelines, literature, code, the scaffolding around the science — the models you can call today are cheaper, independently scored, and not waiting on a rollout stage that has no date.