In early October 2026, TestingCatalog reported a finding that first surfaced on X: a model ID called gpt-rosalind-discovery, alongside a related gpt-rosalind-research entry, appeared in OpenAI's official pricing page under a "Life sciences models" category, with no official announcement, blog post, or help article explaining what they were. The finding was framed as a model that quietly showed up with no visible background. A check by AI24 shows that picture is only partial: OpenAI had already publicly launched GPT-Rosalind about six months earlier, and has since expanded it twice through public announcements covered at the time by multiple outlets.
The original launch: April 2026
According to reports from that period, OpenAI announced GPT-Rosalind on April 16, 2026 as its first model built specifically for life-sciences research, such as drug discovery and genomics. The original announcement described the purpose clearly: to help pharmaceutical companies, research labs and academic institutions analyze large amounts of biological and chemical data faster than existing tools allowed.
Access was limited from the start to a trusted-access program for verified enterprise customers in the United States only. During the preview period, according to those same reports, usage did not count against those customers' existing API quota, meaning organizations already part of the program gained extra access without a separate charge on top of what they already paid for other OpenAI services.
The launch was covered at the time by several outlets tracking the AI industry, including coverage describing it as a clear step by OpenAI to expand beyond general-purpose tools like ChatGPT and into the life-sciences market, where other players also compete.
The June expansion: an upgrade, a new benchmark and a partnership with Novo Nordisk
About two months later, in June 2026, OpenAI released an upgraded version of the GPT-Rosalind family, combining the reasoning capabilities of its general-purpose GPT-5.5 model with enhanced expertise in medicinal chemistry, drug discovery and genomics. The upgrade was presented as the next phase of the original expansion plan, rather than as a standalone new product.
As part of the update, the company introduced a new benchmark called LifeSciBench, on which it said GPT-Rosalind outperformed its other models on life-sciences-related tasks. The update also added life-sciences plugins inside Codex, OpenAI's AI-powered coding environment, intended for biological research and next-generation DNA sequencing analysis.
Novo Nordisk was named as one of the companies exploring the model under the expanded preview program, which indicated that large pharmaceutical companies were already working with GPT-Rosalind in practice, rather than being merely theoretical customers of a closed program.
The current pricing: a family of three models
A direct check of OpenAI's official pricing page shows the "Life sciences models" category now contains three identifiers. gpt-rosalind-discovery and gpt-rosalind-research are both priced at five dollars per million input tokens, fifty cents per million cached input tokens and twenty five dollars per million output tokens, while gpt-6-rosalind, a newer and more expensive identifier, is priced at fifteen dollars per input, a dollar fifty for cached input and seventy five dollars for output.
The sizable price jump for gpt-6-rosalind, roughly triple the input and output rates of the other two models, hints at an additional performance tier, though OpenAI has not published a specification directly comparing the three versions.
Alongside the pricing, OpenAI now also publishes official help pages clarifying the difference between the Discovery and Research tracks: these are two separate access tracks, each with its own permissions and review process, and approval for one does not automatically grant access to the other.
A public price does not make access open to everyone: using any of the three models still requires advance approval through the trusted-access program, whether at the organization level or for the individual user within it.
Why the social-media story was framed as a mystery
The finding published in early October is technically accurate: gpt-rosalind-discovery indeed was not accompanied by a new official announcement at that exact moment, and that is what prompted TestingCatalog's original report. But describing a model that "just appeared" with no background at all fits a product that was never reported on before, not one whose original launch, stated purpose and expansion had already been publicly documented six months earlier.
Anyone who looked into GPT-Rosalind's history beyond the pricing page itself would have found the April announcement and the June update without much difficulty. The difference between "a model never explained" and "a model explained before but not fully checked ahead of the report" matters, even though both versions start from the same underlying finding: a row in a pricing page.
This episode shows how important it is to check a technical finding, such as a change on a pricing page, against a product's full public history, rather than against only what is missing from the official record at that exact moment.
What still remains unknown
OpenAI has not published a full technical comparison between gpt-rosalind-discovery, gpt-rosalind-research and gpt-6-rosalind, so the exact performance differences between the three models remain unknown to the public.
The actual scope of the trusted-access program, the number of organizations genuinely participating in it, and any timeline for a possible move from preview to wider availability have also not been published by the company.
Those details remain closed, unlike the model's existence, its stated purpose and its launch history, all of which have been publicly documented for six months and are not a mystery only now being uncovered.