GPT-Rosalind Discovery: A Second Rosalind Model Appears on OpenAI's Pricing Page With No Announcement

There are now two models in the Rosalind family, and the company has only ever announced one of them. Scroll to the Life sciences block on the vendor's own API pricing page — the section headed "Our latest Rosalind models" — and you will find GPT-Rosalind Discovery sitting directly beneath GPT-Rosalind Research. The API model ids are gpt-rosalind-discovery and gpt-rosalind-research. Both are listed at $5.00 per million input tokens, $0.50 per million cached input tokens, and $25.00 per million output tokens. The second line is new, it is not mentioned anywhere else on the vendor's developer site, and as of October 10, 2026 there is no announcement, no changelog entry, and no model documentation page for it. What follows is what the page actually says, what it does not say, and which readings of the name survive contact with the rest of the vendor's Rosalind material.

What the page says, exactly

The evidence here is thin but unusually clean, because none of it comes from a third party. Everything below was read directly off the vendor's developer documentation on October 10, 2026.

• The listing — the pricing page's Life sciences group carries two rows under the heading "Our latest Rosalind models. Prices per 1M tokens.": gpt-rosalind-research and gpt-rosalind-discovery

• The numbers — both rows read $5.00 input, $0.50 cached input, $25.00 output, per million tokens. Not close. Identical on all three columns

• The footnote — it names only the older model: "Billing for gpt-rosalind-research begins on October 5, 2026. Cache-write pricing does not apply to this model. Access is limited to approved internal research through the trusted-access program." It then adds that "All eligible organizations will continue to get access to the latest GPT-Rosalind models as they're released," plus a standard 10% uplift for regional processing endpoints

• What is missing — no announcement on the vendor's blog, no Rosalind entry in the changelog, and no documentation page. Both /api/docs/models/gpt-rosalind-discovery and /api/docs/models/gpt-rosalind return HTTP 404. The models catalogue still lists a single Rosalind entry, described as "Life sciences reasoning for approved organizations"

The 404 is the part that carries the most information. A pricing row is a billing-table entry, and billing tables are the sort of thing that gets seeded before a launch so that invoicing, quota, and rate-limit plumbing are already in place when the switch flips. A model documentation page is the thing that gets written when there is something to document. One of those exists and the other does not, which is a much better description of a model in staging than of a model in production.

Two SKUs, one price — and what price parity can tell you

The single most-quoted fact about this discovery is the one in the original signal: the new model "has the same pricing as gpt-rosalind-research." That is accurate, and it is worth being precise about what it does and does not imply, because the natural reading — "it's a repackaged Research" — is not the only one, and probably not the best one.

• Identical pricing is the norm for sibling SKUs — when a vendor splits one model into role-specific endpoints, the compute per token is usually the same and the difference lives in the system prompt, the tool set, or the safety and access envelope, none of which change the per-token rate

• Identical pricing is also what a placeholder looks like — a staging row is often created by copying the existing sibling and editing the id, which produces exactly this signature: same three numbers, same footnote, no docs

• Price parity rules out one thing — it is not a smaller, cheaper distilled variant. A cheap-discipline model in this family would have been priced below the $5.00/$25.00 line, the way the rest of the vendor's lineup steps down across tiers

• And it rules in another — whatever Discovery is, the vendor expects it to consume roughly the same token economics as Research, which points at comparable model size or comparable reasoning effort per call rather than a different class of system

Note the deliberate ambiguity: a billing table cannot distinguish between "a distinct capability tier priced coincidentally identically" and "a copy of an existing row that nobody has finished configuring." Both produce the same three numbers. Anyone who tells you the price proves one of these is guessing.

The name, and the one place it does not fit

"Discovery" is not a word the vendor has used for anything else in this family, so the most useful exercise is checking it against the Rosalind surfaces the company has actually documented. The Rosalind Workbench developer post from August 28, 2026 is the best available reference, and it is specific about the agent's modes: "The ChatGPT agent in Rosalind Workbench supports two modes: Explore mode … Research mode."

Two modes. Not three. Discovery is not one of them. That is a real datapoint against the simplest story — that this is just a renamed or newly-exposed Workbench mode — and it pushes the plausible readings toward the model layer rather than the product layer.

What survives, stated as speculation and labelled as such:

• A research-versus-development split — "Research" reads as hypothesis generation and literature-grounded reasoning; "Discovery" reads as target identification, screening, or hit-finding. In a life-sciences context that is a coherent division of labour, and it is the reading most people reached for first

• A partner-facing tier — the footnote's access language ("approved internal research through the trusted-access program") describes an approval-gated program. A second SKU could be a differently-scoped grant inside that same program rather than a new capability

• An evaluation or internal-only endpoint — some vendors expose models that are never meant to be called by customers, listed for internal accounting. A 404 docs page is consistent with this, though so is a premature row

• The boring one — a staging artifact that never ships under this name. This deserves to be on the list. Unannounced pricing rows precede announcements, but they also precede cancellations

None of these is confirmed, and no amount of staring at a pricing table will choose between them. The honest summary is that a name, a price, and a 404 are enough to establish that something exists and almost nothing else.

If you build with models, here is the practical read

There is no action item yet for anyone outside the trusted-access program, and it is worth saying plainly that OrcaRouter does not serve GPT-Rosalind Discovery or GPT-Rosalind Research — this week's /v1/models listing runs to 205 models and contains no Rosalind entry of any kind. The gating on this family is an approval process, not a routing decision, and no gateway can hand you access it does not have. Until the vendor publishes a docs page and opens the model up, the only route to either is the vendor's own API and the access program around it.

What is useful right now is the shape of the problem this leak creates for anyone who does get access. You will be handed an unverified model with a name that implies a job, no public benchmark, no model card, and a price that tells you nothing about which of two similarly-priced models to put on a production path. The standard way through that is to stop treating it as a bet and start treating it as a candidate: run it against the model you already use, on your own data, behind a router that can send a fraction of traffic to it and keep the rest where it is. That is ordinary routed-model work, and it is the layer where the specifics hold up — one API across 200-plus models at provider list price with 0% markup, so a vendor price change shows up on your invoice the same day it shows up on theirs; automatic failover when a provider degrades, which is what makes a two-week experiment on an unproven endpoint survivable; and a routing DSL that lets you compose several models into a single call, which is the natural way to ask whether Discovery and Research actually disagree on the same question.

One number worth carrying forward: the October 5, 2026 billing start for gpt-rosalind-research is the first hard date this program has ever had. Whatever Discovery turns out to be, it lands inside a family that has stopped being free.

What would change the read

This is a leak piece, so the closing judgement should be about falsification rather than prediction. Four things, in rough order of how much each would settle:

• A documentation page going live — a model card with a context window, a knowledge cutoff, and a supported-tools list turns "something is staged" into "something is specified"

• Its disappearance from the table — the cheapest and most likely resolution. Staging rows get pulled. If it is gone next week, the whole thing was plumbing

• A third price — if Discovery ever diverges from $5.00/$0.50/$25.00, the two models are demonstrably different systems and the parity reading was an artifact of a copied row

• A Workbench mode named after it — that would mean the name belongs to the product layer, and the two documented modes were simply incomplete

Until at least one of those happens, the correct posture is the unglamorous one: note that the line item exists, do not build a roadmap on it, and do not let a model id in a billing table change what you ship this month. The family is real, the gate is real, and the second entry has not been announced by the company that is going to sell it.