GPT-Rosalind vs Qwen 4 Max: 3,000 Agents, One Bill, and a Model That Has Not Shipped

Thirty-two minutes, and more than 700 references. On 21 October 2026 O​penAI published a collaboration with PerturbAI and Scripps Research in which thousands of agents running on GPT-Rosalind worked through a whole-brain CRISPR atlas of millions of mouse cells, each agent taking one gene disruption, searching the literature around it, and writing a report — a job the authors estimate at months to a year done one perturbation at a time. The model itself is not new: O​penAI introduced GPT-Rosalind on 16 April 2026, widened it beyond a named partner list on 11 September, and began billing at $5.00 per million input tokens and $25.00 per million output on 5 October. What is new is the first public look at the model running as a fleet rather than as a call, and that changes what a comparison with Q​wen 4 Max should be about. Q​wen 4 Max was announced at A​libaba's Apsara conference on 22 September 2026 and still has no price, no context window and no model card. So the honest version of this matchup has three sides: a specialist whose real unit cost is per agent, an opponent that has not shipped, and the Q​wen model you can actually rent today — Qwen3.8-Max, at $2.00/$6.00 per million tokens. Announcements that never become a callable endpoint are recorded with their dates and sources in the AI Incident Archive, alongside the rest of the model and vendor incidents we follow.

What the 21 October run actually showed

The dataset is a Perturb-seq atlas: genes disrupted one guide RNA at a time across the mouse brain, with each cell's transcriptome read out afterwards. The output is a matrix of thousands of perturbations against millions of cells, far more than any lab can read closely — which is the bottleneck the paper attacks.

The method, which the authors call distributed biological reasoning, is not one large model call over a big table. Each agent receives a single gene disruption, examines the cellular response, retrieves the surrounding literature, and produces a structured report proposing a mechanism. Those reports are then assembled into what the paper calls a reasoning atlas: a graph linking genes, cell-type effects, papers, and candidate mechanisms. The runtime O​penAI reports for the whole sweep is 32 minutes.

The finding that matters most for anyone choosing a model is not the biology. It is the control the team ran on its own method, because that one has a number attached. They held the model and the compute fixed and gave a single agent the entire dataset instead of splitting it across thousands. As the workload grew, the single agent's literature coverage collapsed: at the largest workload it cited 11 references against more than 700 for the distributed agents. Blinded human experts preferred the distributed reports 75% of the time.

Read those together and the result is about decomposition, not about model IQ. A single context window degrades when you pour an atlas into it; a few thousand narrow agents with a retrieval tool each do not. The authors' own framing supports this — they present the DCTN1 cholesterol-transport hypothesis as a question for follow-up experiments, not a discovery — and the paper is a preprint posted on 16 October 2026, co-authored by O​penAI and PerturbAI employees, with no peer review yet. Treat the 32 minutes as a real runtime and the human comparison as an order of magnitude.

For a versus page, the consequence is specific. If the win came from the architecture, then the deciding property of the model underneath is not its benchmark score — it is what it costs to invoke three thousand times in half an hour, and whether you are allowed to.

The number that decides this matchup is cost per agent

Published comparisons of these two models tend to line up headline rates and stop. Headline rates are the wrong unit for the workload the 21 October paper just made concrete, because a distributed run multiplies the model call by the number of perturbations. Work it through with stated assumptions: an agent that reads a shared atlas preamble, pulls a handful of papers, and writes a report might consume 60,000 input tokens and 4,000 output tokens. At O​penAI's published rates that agent costs $0.40 — $0.30 of input and $0.10 of output. The same agent on Qwen3.8-Max at $2.00/$6.00 costs $0.14. Three thousand agents: about $1,200 against about $432.

Cache the shared prefix and the gap widens rather than closes, because the cached-input rates diverge more than the headline ones. Put 50,000 of those 60,000 input tokens in a cached prefix and the Rosalind agent drops to roughly $0.18 while the Q​wen agent drops to roughly $0.06 — a little over three times cheaper, because cached input is $0.50 per million on one side and $0.25 on the other. That is our arithmetic, not a vendor figure, and it is sensitive to how much of your prompt is genuinely shared; the direction is not.

Set the two against each other on the dimensions that decide a fleet run:

• Input price — GPT-Rosalind $5.00 per 1M vs Qwen3.8-Max $2.00. Q​wen 4 Max: not published.

• Output price — GPT-Rosalind $25.00 per 1M vs Qwen3.8-Max $6.00. Q​wen 4 Max: not published.

• Cached input — GPT-Rosalind $0.50 per 1M, and O​penAI's pricing page states cache-write pricing does not apply to the model; Qwen3.8-Max $0.25. Q​wen 4 Max: not published.

• Cost per 3,000-agent sweep — about $1,200 on GPT-Rosalind vs about $432 on Qwen3.8-Max at the assumptions above. Q​wen 4 Max: not computable.

• Context — 1M tokens on both sides. Q​wen 4 Max: not published, and specifically not inherited from the Qwen3.8 window by virtue of the tier name continuing.

• Maximum output — GPT-Rosalind does not publish a per-response ceiling; the Qwen3.8-Max 0902 snapshot documents 131,072 tokens. • Latency — Qwen3.8-Max measures a p50 time-to-first-token of 6.32 seconds on OrcaRouter and 40.3 output tokens per second on the Artificial Analysis index. GPT-Rosalind publishes no latency figure, which for a fan-out run is the number you would want most.

• Independent score — GPT-Rosalind is not tracked on the Artificial Analysis Intelligence Index at all, because it is a gated domain model; Qwen3.8-Max records 45, ranked 33rd of 214 models, at $5.41 per completed index task and 40.3 output tokens per second.

• Access — GPT-Rosalind requires clearance through O​penAI's trusted-access programme; Qwen3.8-Max is a normal commercial API with an open sign-up.

Two of those lines do more work than the rest. The independent score is the only number in the table nobody involved in either model supplied, and it says the Q​wen flagship is a competent mid-frontier generalist rather than a frontier leader — 45 against 58 for the best model on that board. And the access line is not a footnote: it decides whether the fleet can exist at all.

Where Q​wen 4 Max actually stands

A​libaba used its Apsara conference on 22 September 2026 to name four tiers — Q​wen 4 Max as the flagship, Q​wen 4 Flash for throughput, Q​wen 4 Plus in the middle, and an open-weight Q​wen 4 27B — described the family as training on a new-generation architecture, and said it would arrive soon. It attached no date to any of them.

A month later there are three independent places a shipped Q​wen flagship would be visible, and it is absent from all three. A​libaba Cloud's own Model Studio documentation lists qwen3.8-max as the highest Max tier, with no Q​wen 4 entry. Artificial Analysis, which scores most of the commercial Chinese flagships within days of release, has no Q​wen 4 Max row. The OrcaRouter catalogue, which carries more than 200 models including every Q​wen 3.5 through 3.8 tier, has no qwen4 identifier. That is not proof of a delay. It is proof that as of 22 October there is nothing to compare.

The parameter figure that travelled with the announcement deserves one correction, because it keeps propagating into comparison tables. The 5-trillion-to-10-trillion number in circulation is a forward-looking statement about the generations after Q​wen 4 — the Q​wen 4.5 and Q​wen 5 timeframe. A​libaba has attached no parameter count to Q​wen 4 Max. A Q​wen 5-class size printed next to a Q​wen 4 model name is a borrowing error with a long half-life.

What the announcement does tell you is that A​libaba intends to keep contesting the flagship tier, which is a reason to stay inside the family rather than leave it. It tells you nothing you can plan a migration around, because "soon" is not a date and a conference slide is not a rate card.

What the specialist premium buys, and what it does not

GPT-Rosalind's case rests on domain evaluations O​penAI published with the model: leading BixBench performance, six of eleven LABBench2 wins against GPT-5.4, and per-token gains of 53.7% on GeneBench, 18.0% on MedChemBench and 19.6% on LabWorkBench. Every one of those is vendor-reported. None has been reproduced independently, and the model has no entry on the general intelligence boards — which is normal for a gated domain model and also means the numbers cannot be checked from outside.

What is checkable is the tooling. The Life Sciences Research Plugin connects the model to more than 50 scientific tools and data sources as composable skills inside Codex, and the 21 October paper is a demonstration of that stack rather than of raw reasoning: thousands of agents each retrieving literature and writing a structured report is a tool-integration result. Qwen3.8-Max has no equivalent biology stack. It has strong general tool calling and structured outputs, and every scientific database connection would be yours to build.

That is the real shape of the premium. You are not paying 2.5× for a smarter model; you are paying it for domain-tuned reasoning over molecular evidence plus a pre-wired tool layer, and the 21 October run is the first public evidence that the combination produces something a biologist will read. The uncomfortable counterpoint is in the same paper: because the control experiment held the model fixed and changed only the decomposition, the result would largely survive on a cheaper model. The study does not isolate how much of the output is GPT-Rosalind and how much is the architecture — and the authors do not claim it does. If you are budgeting a distributed sweep, that ambiguity is worth about $770 per three thousand agents, which is the difference between the two bills above.

The gate is part of the specification

GPT-Rosalind is not generally purchasable. O​penAI's pricing page states that access is limited to approved internal research through the trusted-access programme, and since 11 September that programme has been open globally to eligible organisations rather than to a fixed partner list. Eligibility is assessed on the proposed use, governance and controlled access — which is a reasonable policy for a model that reasons over pathogen and protein data, and a hard constraint on the fleet pattern the company just demonstrated.

A 32-minute run across thousands of agents assumes you can burst to thousands of concurrent requests on demand. On a gated key, the constraint is not tokens per minute; it is the review, the account structure and whatever concurrency ceiling the programme grants. Any team planning a distributed sweep should price that in before comparing rates, because a cheaper model you can call freely beats a better one you can call on someone else's schedule.

This is where a routing layer earns its place, and where it does not. GPT-Rosalind is not on OrcaRouter — access runs through O​penAI's trusted-access application, and no router changes that. What a router does change is everything around the specialist call. A run like the one published on 21 October is mostly not biology: it is orchestration, report parsing, deduplication across thousands of outputs, and the code that turns an atlas into something queryable. Those steps do not need a $25-per-million-output domain model, and putting them there is how a 32-minute experiment becomes an expensive one. Qwen3.8-Max is on OrcaRouter at Alibaba's $2.00/$6.00 list price with 0% markup, behind the same OpenAI-compatible endpoint as more than 200 other models, so the bulk leg and the specialist leg sit on one key with automatic failover between providers and the routing policy written as configuration rather than as a model name baked into the application.

When Q​wen 4 Max ships, that is also the cheapest way to find out whether it matters: a routing rule that sends a slice of traffic to the new tier, with failover back to the model you already trust, is a trial that cannot take production down. Until then the catalogue simply has no entry for it, which is what an announced-and-unreleased model should look like.

What would flip this comparison

Four published numbers would turn Q​wen 4 Max from a name into an opponent, and they are worth naming so the real thing is recognisable when it appears.

• The rate card. A Q​wen flagship that undercuts $2.00/$6.00 changes the arithmetic in this article by a factor rather than a percentage. One that matches it changes nothing, because the buyable Q​wen model already sits there.

• Long-context behaviour, stated as a policy. Whether the 1M window carries a surcharge band, and at what threshold. A stepped rate above some input size would matter more to an atlas run than the headline number, because every agent in that pattern crosses the band.

• An agentic and tool-use result on a board nobody at A​libaba runs. The gap this article describes is not general reasoning — Qwen3.8-Max is competent there — it is scientific tool use. A published, independently administered result on that axis would be the first real test of the specialist premium.

• Whether the Max tier ships weights. Qwen3.8-Max did not, but the 27B tier of the same generation did, and a Q​wen 4 Max with downloadable weights would be the one thing GPT-Rosalind structurally cannot offer a regulated lab that needs to audit what happens to proprietary sequence data.

The decision this page actually supports does not wait for any of them. Put the specialist where being wrong is expensive — target prioritisation, mechanism proposals, the report a biologist will read — and put the fleet on the model that costs $0.14 an agent and answers without an application form. The 21 October result is a good argument for running distributed biology at atlas scale. It is a weaker argument for running all of it on the most expensive model in the room, and the paper's own control experiment is the reason to say so.