On October 10, 2026, OpenAI added the gpt-rosalind-discovery model to its API pricing page, priced alongside the Research model at $5 per million tokens for input, $0.50 for cached input, and $25 for output. This marks the first new entry in the Rosalind product line in several months, sparking speculation that the model may be specifically designed for pharmaceutical research, given that drug discovery has now been formally incorporated into Rosalind’s official positioning. Rosalind was first launched in April 2026, covering biology and drug discovery; its capabilities were updated in June, applications opened in September, and pricing was made public in October. However, the relationship between Discovery and existing versions, access permissions, and exclusive evaluation results have not been disclosed. The real question lies in whether it can help scientists identify the right research directions earlier—not merely repackage existing answers in a different way.Author and source: BioTender
On October 10, 2026, TestingCatalog reported a new name appearing on OpenAI’s pricing page: gpt-rosalind-discovery. BioTender subsequently checked the official page and confirmed it is listed alongside Research, but the page does not explain the differences in capabilities between the two.
The billing for both models is identical: $5.00 input, $0.50 cached input, and $25.00 output per million tokens. However, identical pricing only indicates that the public rates are the same; it does not prove that training methods, tool access, or task performance are equivalent.
The three pricing tiers for Research and Discovery are identical, priced in USD per million tokens (source: screenshot from OpenAI's official API pricing page).
A new name, which first introduced the pharmaceutical hypothesis.
TestingCatalog then asked: Will it be specialized for drug discovery? The name raises speculation, but the report also acknowledges that its purpose remains unclear. As of this verification, no independent official announcement or capability description for this model has been found.
The scope of the "second model" needs to be clarified: the recently disclosed item is another API variant. However, it is currently unclear whether it corresponds to an independent model, a fine-tuned version, or another product arrangement. A new entry has appeared on the pricing page, but the positioning of the specialized pharmaceutical model has not been confirmed.
TestingCatalog reported a new entry on October 10, 2026, noting that it has not yet been officially announced (source: screenshot from TestingCatalog's report page)[2]
This line was laid as early as April.
On April 16, 2026, OpenAI first released GPT-Rosalind, explicitly targeting biology, drug discovery, and translational medicine. However, this product line did not stop at the initial release: on June 3, official updates to its capabilities were announced. Therefore, today’s leak should not be regarded as the second model release in history.
Subsequently, on September 11, the official updated its announcement, concluding the research preview and opening applications to eligible organizations worldwide; public pricing took effect on October 5. However, these milestones still do not clarify the specific relationship between Discovery and existing versions.
Rosalind's public timeline from launch, capability updates, to the exposure of new models (Source: BioTender, based on OpenAI announcements and TestingCatalog reports)
It aims to take over the handover work in the lab.
Rosalind can be understood as a research assistant capable of searching for information, using tools, and organizing analyses. The initial companion plugin connects to over 50 scientific tools and data sources, but its value still depends on whether these components can be integrated into a trustworthy research workflow. Official partners listed include Amgen, Moderna, and others.
OpenAI’s Workbench puts this idea into a workspace where scientific questions, specialized tools, and verifiable evidence can be advanced together. However, the workbench, tool plugins, and underlying model are products at different levels. Observing the features of the current interface does not directly imply that Discovery inherits all capabilities.
The research workspace concept of Rosalind Workbench is part of the publicly disclosed product background (source: OpenAI's official Workbench blog; not the Discovery interface)
The pharmaceutical hypothesis has its context. Drug discovery has been incorporated into Rosalind's official positioning, and the June update emphasized medicinal chemistry and genomics. However, from the series' application direction to the dedicated task of the new model, an official explanation is still missing.
One hypothesis is that Discovery emphasizes the screening of candidates and hypotheses, while Research takes on broader research tasks. However, this distinction is merely a speculative analysis at the editorial level. The names provide directional clues but cannot replace a task list or evidence of capabilities.
Official example showing the binding structure of the proteasome and bortezomib in the same workspace (source: OpenAI Workbench blog; demonstrates existing Workbench features, not proof of Discovery capabilities)
How far are we from discovering drugs by talking about molecules?
Modifying a molecule requires balancing efficacy, toxicity, and its fate in the body. Such trade-offs fall under multi-parameter optimization. However, good performance on individual predictions does not guarantee an optimal overall decision. The official MedChemBench has incorporated these challenges, along with structural understanding, structure-activity relationships, and retrosynthesis tasks.
Another approach is to first understand the disease mechanism and targets, then organize validation. The Help Center lists early discovery and hypothesis generation as the most valuable current stage, but also covers literature synthesis, omics interpretation, and experimental planning. The public scope is more specific than "automatically generating new drugs."
The official example places fluorescent protein sequence differences and their functional explanations in the same workspace (source: OpenAI Workbench blog; not a Discovery-specific feature).
The current evaluation provides a limited metric: the June update scored 27.5% on MedChemBench, compared to 25.1% for GPT-5.5, a difference of 2.4 percentage points. Improvements were also observed in genomics and wet-lab assistance evaluations, but these results are from previously reported versions by OpenAI and cannot be used to endorse the new model.
Numbers especially need to be grounded in their intended use. The above ratios reflect performance on respective evaluation tasks, but they are not indicators of drug development success rates, nor can scores from different tests be directly compared to assess difficulty. Rosalind’s progress is noteworthy, but Discovery still requires its own controlled evaluations.
June update results compared with GPT-5.5 on three official evaluations; different testing criteria were used, none of which are Discovery scores (source: BioTender, based on OpenAI-published data; data not independently replicated).
Prices have been marked, but the door is not open to everyone.
Rosalind provides services to eligible organizations through the Trusted Access program. However, knowing the unit price does not equate to obtaining access. The Help Center clearly states that the API is intended for approved internal research tools, processes, and applications, and is not available for customer-facing products or external commercial applications.
There’s another easily overlooked detail on the pricing page: the footnote regarding billing starting October 5 specifically names gpt-rosalind-research. However, the independent enablement setup and access terms for Discovery are not elaborated in this footnote, and you cannot simply apply all terms from its sibling model verbatim.
The price footer notes Research's billing date and outlines approved internal research and trusted access requirements (source: official OpenAI API pricing page screenshot).
Next, we need to wait for it to present what kind of evidence.
To determine which is better suited for pharmaceuticals, give both models the same question: How to integrate evidence to identify promising targets or candidates? However, when comparing, keep the data, tools, and budget fixed. Otherwise, differences may stem from external resources rather than the models themselves.
You must also verify whether the recommendations hold up under real-world testing, not just based on completeness of responses. We prioritize traceable evidence, failure conditions, and subsequent validation, but no such proprietary results have been disclosed by Discovery. A price list cannot fill these gaps.
Feedback relationship among evidence integration, candidate hypotheses, and experimental validation (source: BioTender conceptual illustration; does not represent disclosed capabilities of Discovery)
The next set of instructions must address at least three questions: What task is being optimized? In what ways does it outperform Research under the same conditions? Is there experimental support? However, before these answers emerge, Discovery remains primarily a confirmed model name and a set of product hypotheses.
The official page has placed the two names together, and the pharmaceutical direction is also supported by existing products. However, the purpose, version relationship, and actual benefits of the new model remain unclear. What’s truly worth waiting for is how Discovery will change the next research decision—will it help scientists find the right direction sooner, or simply reorganize existing answers in a different way?