Google Says Gemini 4 Argon Designed Five Working Enzymes

Google says its new Gemini 4 Argon model designed five new enzymes in hours and that all five showed catalytic activity when tested in a lab. The claim appears in a Google statement reported by HPCwire on October 21, 2026, and is presented alongside several other scientific-use anecdotes [1]. No preprint, journal article, or independent replication appears in the sources available to us.

What the source actually says

The HPCwire piece relays Google’s own summary: “In another experiment, Google used Argon to design five new enzymes in hours. All five showed catalytic activity when tested in the lab.” The article attributes the statement to Google and places it in a list that also includes Argon helping double AlphaQubit inference throughput, cutting quantum spacetime cost 100×, finding new stellarator mathematics with a University of Maryland professor, and spotting a chromosomal rearrangement at Gamow Labs [1].

None of the four sources available to us—Google’s safety report [3], the model card [4], a Neoteo summary of the safety report [2], or the HPCwire story itself [1]—contain:

• The enzyme sequences, structures, or reaction types
• Experimental conditions, controls, or yield measurements
• A link to a deposited dataset or publication
• Confirmation that the work has been peer reviewed or reproduced outside Google

Google’s own safety assessment finds no significant R&D acceleration

The same model’s Frontier Safety Report, published by Google DeepMind, evaluates whether Argon meets thresholds for “ML R&D acceleration level 1” (substantially accelerating AI progress beyond historical rates) and “ML R&D automation level 1” (fully automating a Google research team). The report states Argon does not meet either threshold [3][4]. An internal benchmark index rose 0.091 points per month before April 2026 and 0.118 after; the difference was not statistically significant, and Google attributed any uptick to “additional investment and training compute” rather than AI-driven acceleration [2].

Human steering remained dominant: fully delegated pull requests were 7.5% in April and 11.9% in September; ~88% still required repeated human input [2]. The enzyme anecdote is not referenced in the safety report’s formal evaluations.

Pattern of unverified scientific claims

The enzyme claim is one of several “real-world impact” examples Google lists without providing verifiable artifacts. Others include:

• Quantum error correction: “nearly double the inference throughput” and “cut the spacetime cost … by more than 100 times” — no benchmark code or independent run cited [1].
• Fusion: “new mathematical solutions” for stellarator simulations — no preprint or collaboration paper linked [1].
• Genetics: “spotted a chromosomal rearrangement … that conventional software had missed for years” — no clinical report or submission mentioned [1].

All are framed as accomplished results, but the sources show only Google’s characterizations.

External commentary is anecdotal, not evidentiary

HPCwire quotes two early users:

• Gopal Srinivasan (Deloitte): “completed successfully 100% of the time in our evals” on long-running tasks [1].
• Alex Cabrera (Axiom Bio): “Gemini 4 was at the frontier across most tasks, including our hardest: ranking the clinical risk between two drugs” [1].

Neither provides data, evaluation sets, or comparison baselines that a reader can inspect.

What remains unknown

• Whether the five enzymes are novel folds or variants of known scaffolds
• What “catalytic activity” means quantitatively (turnover, specificity, stability)
• Whether the designs were filtered by human experts before lab testing
• Whether the lab tests were blinded or included negative controls
• Timeline for any public disclosure, deposition (e.g., PDB, GenBank), or peer-reviewed publication

Bottom line

Google claims Argon designed five enzymes that worked in its lab. The claim is not supported by published evidence in the available sources, and Google’s own formal safety evaluation concludes the model has not accelerated its AI R&D pipeline in a statistically significant way. Until methods, data, and independent replication appear, the enzyme result remains an unverified corporate anecdote.