Big Tech

OpenAI MedChemBench Reveals Where GPT-6 Fails at Drug Design

OpenAI's MedChemBench tests 11 GPT models on 99 drug problems. GPT-6 Astra scores 95% on structure reconstruction but falls short on property prediction.

Share:XLinkedIn

Key Takeaways

  • 99 problems drawn from 117 papers and patents spanning 761 molecular structures test four categories: structure reconstruction, scientific reasoning, compound selection, and numerical property prediction.
  • GPT-6.2 Astra scored 84.5% overall at the highest reasoning setting; GPT-6 Astra hit 95.4% on structure reconstruction but earned roughly half credit on compound selection and property prediction.
  • 11 GPT models were evaluated across 51 combinations; GPT-5.6 Sol scored 58.2%, a 25-point gap below GPT-6.2 Astra that shows how much reasoning capacity matters in chemistry tasks.
  • No AI-discovered drug has ever received FDA approval, and Isomorphic Labs, Recursion, and Chai Discovery each run specialized drug-design systems that were not evaluated on MedChemBench.
  • OpenAI explicitly cautions the results do not establish whether models can discover new drugs, and prior exposure to the 117 published source papers in the training corpus cannot be ruled out.

OpenAI published MedChemBench on November 20, testing eleven of its GPT models against 99 real drug discovery problems drawn from 117 published papers and patents. The scores split cleanly: GPT-6 Astra reconstructed molecular structures at 95.4% accuracy, then earned roughly half that when asked to select the right drug candidate or predict a compound's properties. The gap between those two numbers is not a quirk of the benchmark. It is a precise map of where large language models currently stand in the most consequential applied science of the decade.

What 99 Drug Discovery Problems From 117 Papers Actually Measure

MedChemBench tests medicinal chemistry reasoning across four scored categories: molecular structure reconstruction, scientific reasoning from experimental evidence, compound ranking and selection, and numerical property prediction. OpenAI built each of the 99 problems from a snapshot of a real drug research program, giving the model the context a working medicinal chemist would have had at a specific moment: the biological target, the research goals, the compounds tested so far, and selected experimental findings. The model then had to do what a chemist would do next, with no tools, no feedback loop, and no access to new lab results.

The benchmark spans 761 distinct molecular structures drawn from 117 published papers and patents, according to OpenAI's official MedChemBench announcement and the accompanying MedChemBench technical paper. All 11 GPT models were evaluated across 51 model-and-reasoning-effort combinations. Each task received a single attempt. Scores represent the percentage of available credit earned, with partial credit awarded for numerical predictions based on how close the model's answer came to the measured value. That partial-credit system matters because predicting a drug's potency is not a binary pass or fail: a model that estimates a binding constant at 12 nanomolar when the true value is 8 nanomolar is more useful than one that guesses at 500.

The benchmark measures four capabilities deliberately chosen to match the sequential decisions in an early drug program. Structure reconstruction tests whether a model can read a chemical diagram accurately enough to work with it. Scientific reasoning tests whether the model can connect experimental observations to the molecular structures that produced them. Compound selection tests whether the model can rank candidates against each other. And property prediction tests numerical estimates of how a compound will behave in a cell or in a body. The last two categories are where real programs spend most of their time and money, and they are where GPT-6's performance dropped the furthest.

Stay Ahead

Get daily AI signals before the market moves.

Join founders, investors, and operators reading TechFastForward.

GPT-6 Astra Scores 95% on Structure and Stalls Near 50% on Compound Selection

At the highest shared reasoning setting, GPT-6 Astra earned 95.4% of available credit on molecular structure reconstruction. OpenAI tested this category with an 18-atom ring problem: models had to assemble a compound while preserving a specific ring structure that defined the molecule's chemical identity. GPT-6 Astra produced the correct structure in all ten attempts at the highest reasoning setting. GPT-5.6 Sol returned a 15-atom ring. GPT-5.6 Luna, one model generation older, generated disconnected molecular fragments. The performance gap across just two generations is striking.

Overall, GPT-6.2 Astra scored 84.5% at the highest reasoning setting, followed by GPT-6 Astra at 83.3% and GPT-6.1 Sol at 82.0%. GPT-5.6 Sol dropped to 58.2%, a 25-point gap that shows how dramatically reasoning capacity affects chemistry performance as models scale down. The 11 GPT models in the evaluation span from GPT-5.5 through GPT-6.2 Astra, allowing a cross-generational comparison that individual model releases rarely permit, as Interesting Engineering reported.

Compound selection and property prediction tell a different story. On those two categories, GPT-6 Astra earned roughly half the available credit. A representative problem involves PRMT5 inhibitors: the model receives three related compounds and experimental data, then predicts which compound inhibits the PRMT5 protein most strongly in a cellular assay, which one crosses the cell membrane most readily, and which breaks down most slowly in the liver. The withheld experimental results show that compound 23 had stronger PRMT5 inhibition, while compound 24 absorbed better and metabolized more slowly. Getting that call right requires understanding that stronger inhibition and better pharmacokinetics rarely coincide in the same molecule, and that a drug program advancing the stronger inhibitor might select the candidate that fails in the clinic.

Why Reading a Molecule and Choosing One Are Different Cognitive Problems

Structure reconstruction is a skill that maps well onto pattern recognition. Organic chemistry uses a consistent notation system built on carbon backbones, bond angles, and functional groups. A model trained on a large corpus of published chemistry literature learns that notation the same way it learns any formal grammar, and 95% credit on reconstruction reflects genuine pattern-recognition strength applied to a consistent symbolic language. The 18-atom ring problem is hard for humans to get wrong with the full structure in front of them. Getting it wrong reveals a failure of representation, not of chemistry reasoning.

Compound selection is something categorically different. It requires reasoning across competing objectives simultaneously: potency against a target protein, absorption across a cell membrane, metabolic stability in the liver, distribution through tissue, and toxicity risk in multiple organ systems. In pharmaceutical development, this is called multi-parameter optimization, and industry data consistently shows that more than 90% of drug candidates entering clinical trials never reach approval. The majority of those failures stem from ADMET problems, absorption, distribution, metabolism, excretion, and toxicity, rather than from failure to block the intended target. The target-binding step is frequently the easiest part.

A model that scores 50% on selection is making the right call in roughly one of every two cases. That may exceed intuitive expectations for a general-purpose language model applied to chemistry. But clinical drug development is not a domain where a coin-flip success rate is acceptable, and the cost of a wrong selection decision compounds across a multi-year research program. A lead compound selected on incomplete criteria can consume three to five years and $50 million to $200 million in development costs before wet-lab experiments reveal the error. The gap between 95.4% on structure and 50% on selection is not an abstract benchmark artifact. It is an estimate of how much of the value in a drug program still depends on human chemical intuition and experimental iteration.

The Risk OpenAI Is Careful Not to State Outright

OpenAI's announcement text includes an explicit caution: the results "do not establish whether models can discover new drugs or improve laboratory decisions." The company further acknowledges that prior exposure to the published source material cannot be ruled out. MedChemBench draws from 117 papers and patents, most of them publicly available through PubMed and chemical databases. A model trained on a broad internet corpus has almost certainly read a subset of them. The benchmark cannot currently distinguish between a model that reasons from first principles about molecular properties and a model that pattern-matches from training data that includes the same published experimental outcomes.

The bear case for MedChemBench is more structural than OpenAI's own caveats suggest. The benchmark evaluates 11 GPT models and no others. No results appear for Isomorphic Labs' IsoDDE engine, the Recursion platform, Chai Discovery's Chai-3 antibody model, or any of the other purpose-built drug-design AI systems now operating at pharmaceutical scale. Critics would note that a benchmark constructed by one company and tested exclusively on that company's models produces a document that proves performance on a self-defined task, not a demonstration of chemistry reasoning in any objective sense. By that standard, MedChemBench is a product announcement with supporting methodology, not an independent scientific benchmark.

There is also a question of ecological validity. MedChemBench uses historical data: the model receives information that researchers already had at a specific point and predicts what they then found. Prospective drug design, the actual task, requires predicting outcomes for molecules that have never been synthesized or tested. Historical success rates on published research, which skews heavily toward positive results, are a systematically optimistic proxy for performance on real programs where most compounds fail before they appear in a journal.

What Isomorphic, Recursion, and Chai Are Building Instead

The drug-discovery AI field has divided into two distinct architectural bets. One bet is the general-purpose route that OpenAI is now entering formally: apply a frontier language model trained on broad data to chemistry problems and test how far that general capability extends. The other is the specialized route: build systems trained primarily on biological and chemical data, fed by proprietary wet-lab experiments, and designed from the ground up for the specific decisions in a drug program. MedChemBench measures the first approach. The second approach is not yet on any public benchmark.

Isomorphic Labs, the drug-design company that Google DeepMind spun out in 2021, represents the most resource-intensive version of the specialized approach. The company raised pharma partnerships in 2024 that total roughly $4 billion in headline milestone value, according to Scientific American: Eli Lilly at $45 million upfront against $1.7 billion in development milestones, and Novartis at $37.5 million upfront against more than $1.2 billion, with Johnson and Johnson joining the roster later. IsoDDE, the drug design engine the company released in February 2026, is proprietary and has not been benchmarked on MedChemBench or any public evaluation. The company was targeting entry into a first Phase I clinical trial with an AI-designed molecule in late 2026, a milestone that would be the most consequential data point in the field, well beyond any benchmark score.

Recursion took the acquisition route, closing its purchase of Exscientia in November 2024 and combining Recursion's high-throughput imaging biology platform with Exscientia's computer-aided drug design stack. That combination targets the full workflow from target identification through lead optimization, with experimental data feeding model refinements at each step. Recursion also co-released Boltz-2, an open-source structure and binding-affinity prediction model that provides an open alternative to Isomorphic's proprietary stack. Chai Discovery built in a narrower direction: its Chai-3 antibody design model was licensed to both Pfizer and Eli Lilly within twelve months of release, suggesting that specialized models for specific molecular types find paying customers faster than general-purpose tools. As Fortune reported in January 2026, big pharma and venture capital have committed billions to this space, but the question of which approach wins in the clinic has not been answered.

A further challenger appeared in November 2026: Ptarmigan-1, a structure-free screening model described by The Scientist, which claims to screen more than one billion drug candidates per day against the dark proteome, the roughly 80% of human proteins for which no experimental crystal structure exists. If that throughput claim holds under independent testing, it represents a different category of tool entirely: not reasoning about which known compound to advance, but generating and filtering novel candidates at a scale no human team could match. Across all of these platforms, one fact stands unchanged: no AI-discovered drug has ever received approval from the US Food and Drug Administration, as Genetic Engineering and Biotechnology News has tracked continuously.

Three Signals Worth Watching Before Trusting a Benchmark Score in a Drug Program

The first signal to watch is whether Isomorphic Labs enters a Phase I clinical trial with an AI-designed molecule before the end of 2026. That was the company's stated target as of early 2026, and no announcement has appeared as of this writing. If the trial begins, it marks the first time a structure-prediction AI directly influences a human dosing study, regardless of outcome. It would also supply the first head-to-head data point between a general-purpose-LLM approach and a specialized drug-design system: MedChemBench scores exist for GPT-6, but no compound designed by GPT-6 is in a trial. If Isomorphic's molecule enters the clinic while OpenAI publishes benchmarks, the competitive framing shifts from scores to clinical outcomes.

The second signal is whether any external laboratory publishes MedChemBench results for a non-OpenAI system. Isomorphic, Recursion, Chai, or academic chemistry groups could run the 99 problems through their own models and release the numbers. If a specialized drug-design system scores above GPT-6's compound-selection results, it reframes the OpenAI benchmark as a gap report rather than a capability demonstration. If specialized systems score similarly or below, it validates the view that general-purpose frontier models are closing the advantage that purpose-built chemistry AI once held clearly. That comparison does not exist yet.

The third signal is regulatory. The FDA has not issued formal guidance on AI-assisted new drug applications. As AI tools begin influencing compound selection decisions in real programs, the agency will need to define what level of traceability and experimental validation it requires before an AI-driven recommendation can support an IND filing. Watch the FDA's public docket for any draft guidance on AI in drug development by mid-2027. Any such guidance will almost certainly treat benchmark performance on historical data as necessary but insufficient, and it will specify what prospective validation looks like. That document, whenever it comes, is what will actually determine how much weight MedChemBench-style scores carry in a regulated development program.

The gap between 95% on molecular structure and 50% on compound selection is not a benchmark artifact: it is a precise measurement of how far from clinical utility a language model currently stands.


Key Takeaways

  • 99 problems, 761 molecular structures, 117 papers: MedChemBench evaluates four categories: structure reconstruction, scientific reasoning, compound selection, and property prediction, with partial credit for numerical closeness.
  • GPT-6.2 Astra scored 84.5% overall; GPT-6 Astra hit 95.4% on structure reconstruction: but both models earned roughly half the available credit on compound selection and numerical property prediction, the categories most directly tied to clinical decisions.
  • 11 GPT models were evaluated across 51 combinations: GPT-5.6 Sol scored 58.2%, a 25-point drop from GPT-6.2 Astra that shows how much reasoning capacity matters in chemistry tasks.
  • No AI-discovered drug has ever received FDA approval: Isomorphic Labs, Recursion, and Chai Discovery are each running specialized drug-design systems that were not evaluated in MedChemBench.
  • OpenAI explicitly states the results do not prove models can discover drugs: and prior exposure to the 117 published source papers in the benchmark's training corpus cannot be ruled out.

Questions Worth Asking

  1. If GPT-6 Astra scores roughly 50% on compound selection, what would Isomorphic's IsoDDE or Recursion's platform score on the same 99 problems, and why has no one published that comparison yet?
  2. MedChemBench uses historical data from published research. How much of GPT-6's 84.5% overall score reflects genuine molecular reasoning versus pattern-matching on training data it has already encountered in its corpus?
  3. With no AI-discovered drug ever approved by the FDA, what prospective validation would actually change how a pharma company uses a benchmark result like this in a real development decision?

Sources

  1. OpenAI's official MedChemBench announcement openai.com
  2. MedChemBench technical paper cdn.openai.com
  3. Interesting Engineering reported interestingengineering.com
  4. Scientific American scientificamerican.com
  5. Fortune reported fortune.com
  6. The Scientist the-scientist.com
  7. Genetic Engineering and Biotechnology News genengnews.com

Read Next

Nothing Sets December 8 Reveal for Screenless AI Device

UN Defers Killer Robot Binding Rules for Two More Years

Meta Raises Quest Price $100 While XREAL Hits 95 Grams

Baseten bison-1 Tops Decision Index at 76 With Open Weights

Newsletter

Enjoyed this analysis? Join the list for the email edition.

Daily AI signals. No noise. Built for founders, investors, and operators.

Share:XLinkedIn
</> Embed this article

Copy the iframe code below to embed on your site:

<iframe src="https://techfastforward.com/embed/openai-medchembench-reveals-where-gpt-6-fails-at-drug-design" width="480" height="260" frameborder="0" style="border-radius:16px;max-width:100%;" loading="lazy"></iframe>