AI FOR SCIENCE · INTERVIEWS
OpenAI's life-sciences team says it is building a dedicated biochemistry model series starting from early discovery, and will open more capability only to verified professional users.
Gives the product shape (a model series starting with genomics and protein understanding, and a plugin with 50-plus templated workflows), admits GPT-5 was trained mostly on maths and code, which have verifiable answers, and explains why intent cannot be read from a prompt, so access must be tiered by who the user is.
The interview gives no model name, benchmark figure or release date; those came later, as the After the interview items show.
Wang says the new series is anchored on complex life-science research workflows, adding mechanistic understanding of genomics and proteins first, focused on early discovery. Product-wise, ChatGPT handles literature synthesis and Codex long agentic runs; for enterprise repeatability there is a plugin with more than 50 templated skills.
Jiao says models can already run open-source protein structure prediction tools in a research stack, read the output and adjust the input; what would make them stronger is becoming a biochemistry expert that uses the tools more intelligently. She notes people at the Baker Lab use Codex for simple jobs, such as a spreadsheet that cuts pipetting steps.
The collaboration began in July 2025, when GPT-5 had just finished training mostly on maths and computer science, because those have verifiable answers. Jiao recalls that when the first experiments GPT-5 designed came back, they had made a non-zero amount of protein, which surprised her. She is recalling the earliest round, not the collaboration's final level of evidence; the later OpenAI and Ginkgo report is under Related evidence below.
“we made a non zero amount of protein, that was actually quite surprising.”Joy JiaoWang says biology is the first dual-use risk that is also a capability risk, and the early steps toward something dangerous look much like legitimate work. Jiao says a prompt rarely reveals intent: 'help me clone a gene' could be green fluorescent protein or a toxin. So general traffic gets refusals or overviews, which irritates professional scientists; unlocking full capability needs users known to be professionals at regulated institutions.
“it's very difficult to figure out what a user's actual intentions are just from kind of reading a prompt.”Joy JiaoWang imagines many sub-agents orchestrating in parallel with researchers left to interpret; the team's slogan is to scale test-time compute to cure all disease. Jiao says evaluation uses existing experimental data and synthetic data with traps, but that the wet lab remains the final test; she also does not think scientists can be fully replaced.
The OpenAI and Ginkgo report of February 2026 (now on this timeline) says GPT-5 tested 36,000 reaction compositions in a cloud lab and cut the cost per gram of cell-free protein synthesis 40% below the previous best.
Jiao and Wang describe a dedicated life-sciences model series and plugin line focused on early discovery, and recall the first Ginkgo round as a non-zero amount of protein.
That report says human intervention was largely limited to preparing, loading and unloading reagents and consumables, with design, execution, analysis and new hypotheses handled by the LLM-driven lab.
Wang imagines the bottleneck shifting from people to compute, with many sub-agents orchestrating in parallel; Jiao says they do not think scientists can be fully replaced and a person stays in the loop.
OpenAI is building a dedicated biochemistry series beside its frontier models; Anthropic's lead says in another interview that general frontier models may suffice. Each is its own account, and no comparison under one benchmark exists yet.
Their argument is that a prompt does not reveal intent, so capability follows who the user is. That makes who verifies the user a gate in the product; the interview does not say who verifies or by what standard, which is our inference.
A biochemistry model series is being built, starting with genomics and protein understanding, focused on early discovery.
Company-reported
A plan described; no model name, benchmark figure or release date.
The life-sciences research plugin has more than 50 templated skills.
Company-reported
Wang's statement; not checked against a product page.
Models can already call open-source protein structure prediction tools in a research stack.
Current system
Jiao's description; no tool or case is named, and the Baker Lab example is a simple spreadsheet task.
The first Ginkgo experiments designed by GPT-5 made a non-zero amount of protein.
Company-reported
Jiao is recalling the earliest round (July 2025). The later OpenAI and Ginkgo report, with Jiao a corresponding author, says 580 plates and 36,000 reaction compositions over six months, a cost per gram 40% below the previous best ($422/g against $698/g) and titer up 27%, with human intervention largely limited to preparing and loading reagents and consumables. Non-zero is not the collaboration's level of evidence.
AI designs a new drug or cures a disease within a few years; a drug takes about a decade to market; autonomous labs in ten years.
Target
Jiao's and Wang's hopes; no roadmap.
Biology is the first dual-use risk that is also a capability risk, with high-level safeguards activated and incremental deployment.
Industry context
Wang's statement; no trigger criteria are given.
No published biology-model benchmark; an antibody-binding prediction eval is still being built.
Not yet closed
Wang says she is excited about that eval and does not say it is done.
Whether the life-sciences series has externally scored evaluations on genomics and protein tasks.
Would show up as A third-party or independent evaluation, not OpenAI's own benchmark.Partly answered In OpenAI's GPT-Rosalind-5.5 system card of 2026-06-03 the benchmarks are designed by OpenAI with industry experts; none is scored externally. The LifeSciBench preprint of 2026-08-18 (written by OpenAI with Tacit Labs, 750 tasks, graded by GPT-5.5) has GPT-Rosalind passing 36.1%, but the tasks, the top model and the grader all come from OpenAI. A bioRxiv preprint of 16 October from Scripps Research, OpenAI and PerturbAI puts GPT-Rosalind on a genomics task outside OpenAI's own benchmarks: a whole-brain Perturb-seq atlas with the source paper withheld, where it matched or refined 48 of the 55 findings derivable from the expression data, scored by the source study's own authors, with blinded experts as well. But both guests of this interview, Jiao and Wang, are authors and the data come from the same group, so it is still not an independent evaluation; the preprint is not peer reviewed, and there is no counterpart on protein tasks.Whether the wet-lab collaborations still under way publish results, stating which steps people did.
Would show up as A report from a partner or OpenAI stating experiments, yields and manual steps.Partly answered The Ginkgo collaboration has a report (2026-02-05, written by OpenAI and Ginkgo, stating experiments, yields and manual steps); results from other partners are not seen.Whether the access criteria and the verifier are published.
Would show up as A published OpenAI access policy stating who qualifies and how they are verified.Partly answered The GPT-Rosalind-5.5 system card (2026-06-03) publishes the review criteria for organisations: mission, use case, research and operational governance, and the ability to control access; non-government applicants also face business verification and compliance screening, with a separate government pathway. The Rosalind Academic Plan of 15 October opens the door to US university labs (on the timeline since 17 October, its terms read from OpenAI's announcement page as supplied by the editor): a seat does not unlock GPT-Rosalind, which needs a separate trusted-access review of the lab head's identity, employment and authority and of the research purpose, with a named institutional safety contact. The reviewer is still OpenAI itself, with no outside verifier; what counts as an eligible lab, how long review takes and how many labs will be admitted are not published.The gap between a dedicated biochemistry model and general models on one benchmark.
Would show up as An independent comparison on the same benchmark.Still open OpenAI's own system card compares them (its proprietary Labwork Bench: 63.2% against 55.8%); no independent comparison is seen.Later evidence that tests or limits what was said, and targets whose date has passed.
The day after the interview OpenAI announced GPT-Rosalind as the first model in its life-sciences series, a research preview through a trusted-access program in ChatGPT, Codex and the API, with Amgen, Moderna, the Allen Institute and Thermo Fisher named as partners. The unnamed series became a named product.Fierce Biotech ↗
OpenAI published the GPT-Rosalind-5.5 system card: incrementally trained from GPT-5.5 and offered as a research preview to trusted organisations. It states the access criteria (mission, use case, governance, ability to control access, plus business verification for non-government applicants); on its proprietary Labwork Bench the model scores 63.2% against 55.8% for GPT-5.5. The evaluations were designed with industry experts by OpenAI and are not externally scored.OpenAI, GPT-Rosalind-5.5 system card ↗
GPT-Rosalind left research preview and became available to eligible organisations worldwide through the trusted-access program, with API billing from 5 October, limited to approved internal research. Access by verified user, as the interview describes it, became the formal access model.AiCybr ↗
Read 2026.10.01 The full machine transcript from a third party (Wave), checked against the Acast episode page; names follow the episode page. The items under After the interview were read through press reports, since OpenAI's own pages returned 403 to both fetch paths. Open the interview ↗Acast episode page ↗