Can AI do science yet?

The exam

An exam for AI, set by the lab itself

On November 6, OpenAI, working with enzyme-design company Ligo, introduced BioBasicsBench, a benchmark of 100 tasks that tests whether AI agents can interpret raw wet-lab evidence: gel images, colony plates, microscopy, sequencing reads, and instrument exports (OpenAI, 2026). Scientists wrote the tasks from their own experiments, each with an expert-written grading rubric. Every task passed review by a Ligo scientist, an external PhD-level wet-lab scientist, and a final internal check (OpenAI, 2026).

Most science benchmarks use the clean figures from published papers. This one tests the messy evidence labs actually produce, including failed experiments and misleading controls. At the highest reasoning setting, GPT-6 Astra earned 66.7% of the available rubric credit, against 43.8% for GPT-5.6 Sol, across 18,500 graded attempts (OpenAI, 2026; AI Frontier Post, 2026).

The worked examples show what better lab judgment looks like. In a cell-imaging task, Astra recognised that two dyes shared a detection window, so brightness alone could not prove cells were dead, and it called for better controls. In a cloning task, it read a handwritten “LB Kan50” label on a plate rim, noticed the plate held kanamycin while the experiment used ampicillin, and rejected the plate as a control. In another task, it recognised a series of microscope images as a depth scan, not cells vanishing (OpenAI, 2026).

OpenAI is equally candid about the failures. Astra sorted colony lanes into the right two groups, then picked the wrong one, which would have meant purifying the wrong plasmid. It miscounted colonies and overgeneralised from cell images that contradicted its conclusion. The score is partial rubric credit, not a pass rate, and there is no human baseline and no non-OpenAI model tested (OpenAI, 2026; AI Frontier Post, 2026). The dataset and rubrics were not publicly linked at launch, so outside labs cannot yet verify the scores or run their own models on the same tasks (AI Frontier Post, 2026).

The predictor

A molecule predictor that works from scant data

MIT researchers, working with BASF, have built CheMeleon, a machine-learning model that predicts how a molecule will behave even when few measurements exist for the property in question (Chandler, 2026). Drug and materials discovery is slow because testing each candidate molecule in the lab is expensive. A model that narrows the field from little data means fewer wasted syntheses and faster shortlists.

Instead of training on published experimental results, the team pre-trained CheMeleon on about one million molecules from PubChem, teaching it to predict roughly 1,600 easily calculated descriptors such as molecular weight and oxygen count (Chandler, 2026). The model then fine-tunes on a small amount of task-specific data. Across 58 test cases, it outperformed existing methods, including classical models built on expert knowledge (Zalte et al., 2026).

The model is open source through ChemProp, and pharmaceutical companies are already using it to choose which molecules to synthesise and test (Chandler, 2026). In a contest earlier this year to predict properties of molecules that had not previously been synthesised, nine of the top 20 entries had used CheMeleon (Chandler, 2026; Unite.AI, 2026).

The evidence here is the strongest of the week: a peer-reviewed paper in the Journal of Chemical Information and Modeling, published in the September 14 issue (Zalte et al., 2026). The news is MIT’s November account of that work and its reported industry adoption. Predictions still need experimental confirmation, and MIT does not quantify the lab savings.

The chemist

A chemist that designs drugs in an hour

On November 3, Google DeepMind added Cheminformatics Skills to Google Antigravity, turning the agent into a science workbench for drug discovery (Tsangalidou et al., 2026). Designing new medicines means curating data from charts, figures, and literature, training predictive models, and proposing candidate molecules. That workflow typically takes multiple days for a single iteration.

In a lung-cancer example, the agent searched literature and compound databases, ran matched molecular pair analysis, and proposed new rigidified macrocycles with synthesis routes in about an hour (Tsangalidou et al., 2026). The example builds on crizotinib and lorlatinib, two lung-cancer drugs, and the rigidification technique that made the second more potent. Independent coverage confirms the same hour-long workflow and the same drug example (Emergent, 2026).

The design matters as much as the speed. The language model handles workflow orchestration and scientific reasoning, while deterministic chemistry tools run the calculations. The skills are grounded in the open-source RDKit toolkit and connect to the ChEMBL and PubChem databases (Tsangalidou et al., 2026). DeepMind is offering 10,000 scientists its highest usage tier in the tool free of charge for six months (Tsangalidou et al., 2026).

This is a company demonstration, not an independent evaluation. DeepMind itself states the limits: sparse data, assay noise, and narrow measurement ranges limit predictive precision, and every proposed design must be synthesised and validated experimentally (Tsangalidou et al., 2026). The molecules are testable hypotheses, not finished drugs.

The receipt

A receipt that proves which model answered

On November 2, Anthropic released FlashProve, an open-source research prototype that attaches a cryptographic proof to an AI reply, showing it came from one exact set of model weights (Anthropic, 2026). The provider first publishes a short fingerprint, called a commitment, for those weights. It then produces a proof for each reply that anyone holding the commitment, the prompt, and the reply can check, without seeing the weights (Anthropic, 2026).

Anthropic demonstrated it on Llama 4 Maverick, a 400-billion-parameter mixture-of-experts model. A single graphics card generated proofs at about 110 tokens per second, with each proof running about 19 megabytes per 1,024-token request, and verification taking seconds on an ordinary processor (Anthropic, 2026; Brandom, 2026).

The point is not the cryptography. It is what the cryptography makes possible. A reply alone cannot show which model produced it, so a cheaper model could be swapped in silently. A checkable receipt gives evaluators and paying customers a way to catch that swap. It also gives regulators a way to check that a lab is serving the model it claims to serve (Anthropic, 2026).

FlashProve is a research prototype, not a deployed service. Proving takes about 70 to 95 times the graphics-card time of ordinary inference, and there are no plans to deploy it at Anthropic or any other lab (Anthropic, 2026; Brandom, 2026). The code, a technical report, and formal proofs are public, explicitly so others can improve on them.

The limit

What the week does not prove

Put the four developments together and a pattern appears. Each one moves the test for AI in science closer to the work itself: from published figures to raw lab evidence, from data-hungry models to low-data prediction, from days of data wrangling to an hour, and from trusting a reply to proving which model wrote it.

But each record is partial, and the partiality is the point. BioBasicsBench is self-graded: OpenAI tested only its own models, with no human baseline and no public dataset at launch. CheMeleon’s paper is from September, and its industry adoption is reported by MIT, not independently audited. Antigravity’s hour is a company demonstration of its own tool. FlashProve is a prototype whose cost makes deployment impractical today.

None of that makes the week a press-release parade. The honest signal is the direction of the checks. The labs are no longer asking whether a model can answer science questions. They are asking whether it can read a gel, predict a molecule, design a drug candidate, and prove it did the work. The next step is independent replication: public datasets, other labs’ models on the same bench, synthesised molecules from agent designs, and proofs anyone can check at a usable cost.

So the week’s story is narrower than a breakthrough and larger than a demo. AI did real lab-adjacent work this week, in public, with its failures published alongside its scores. The question now is whether the independent checks arrive as fast as the capabilities.

What changed / what remains open

What changed. In one week, the test for AI in science moved from published figures to raw lab evidence, from data-hungry models to low-data prediction, from days of drug-design wrangling to an hour, and from trusting a reply to proving which model wrote it. The capability and the check arrived in the same week, from four different labs.

What remains open. Whether OpenAI publishes the BioBasicsBench dataset so other labs can run it, whether CheMeleon’s predictions convert into synthesised and tested compounds, whether Antigravity-designed molecules survive the lab, and whether provable inference becomes cheap enough to deploy. Each of those questions has a concrete next step, which is itself the week’s most useful legacy: the claims now have tests attached.

Worth watching

Worth watching

• BioBasicsBench, opened up. Whether OpenAI publishes the dataset, rubrics, and scoring code, and whether other labs run their models on the same 100 tasks.
• CheMeleon in the lab. Whether the pharmaceutical companies already using it publish results from model-prioritised molecules that were synthesised and tested.
• Antigravity’s 10,000-scientist offer. Whether user results from the free tier appear in published drug-discovery work over the next six months.
• FlashProve’s cost curve. Whether proof generation gets fast enough for real deployment, and whether any lab adopts provable inference for its served models.

References

1. AI Frontier Post. (2026, November 6). OpenAI’s BioBasicsBench tests AI on messy wet-lab evidence: GPT-6 Astra leads at 66.7%, but still misreads the bench. https://aifrontierpost.com/articles/openai-biobasicsbench-wet-lab-benchmark/
2. Anthropic. (2026, November 2). Progress on provable inference for large language models. https://www.anthropic.com/research/provable-inference
3. Brandom, R. (2026, November 3). Anthropic is working on a system to prove AI models’ identity. TechCrunch. https://techcrunch.com/2026/11/03/anthropic-is-working-on-a-system-to-prove-ai-models-identity/
4. Chandler, D. L. (2026, November 6). How to predict the behavior of molecules from scant data. MIT News. https://news.mit.edu/2026/how-to-predict-molecule-behavior-with-scant-data-1106
5. Emergent. (2026, November 3). Google Antigravity accelerates drug discovery workflows. https://emergent.sh/news/google-antigravity-accelerates-drug-discovery
6. OpenAI. (2026, November 6). Measuring how AI agents handle everyday wet-lab science. https://openai.com/index/measuring-how-ai-agents-handle-wet-lab-science
7. Tsangalidou, Z., Zhang, R., Singh, S., & LeVine, M. V. (2026, November 3). Agentic cheminformatics with Google Antigravity. Google DeepMind. https://deepmind.google/blog/agentic-cheminformatics-with-google-antigravity/
8. Unite.AI. (2026, November 6). MIT details CheMeleon model for low-data molecular prediction. https://www.unite.ai/mit-details-chemeleon-model-for-low-data-molecular-prediction/
9. Zalte, A. S., Burns, J., Abreu, C., Green, W. H., et al. (2026). Deep learning foundation models for low-data molecular property prediction. Journal of Chemical Information and Modeling, 66(17). https://pubs.acs.org/jcisd8/article-abstract/66/17/10692/5250516/Deep-Learning-Foundation-Models-for-Low-Data