Anthropic: 26.8% on average, but 0 out of 90 for a target

Claude's autonomous protein design campaign boasts a validated binding rate of 26.8%, surpassing the industry standard of 10 to 15%. However, performance varies significantly across targets, with one objective where none of the 90 designs achieved binding. Estimates from published accounts suggest that a new target could often fall well below 26.8%.

Performance Varies Widely by Target

The results published by target range from 72 bindings out of 90 for TREM2 to 0 out of 90 for the maltose-binding protein. For the latter, 90 designs were synthesized, expressed, and measured without achieving a single binding. These discrepancies indicate that the overall rate of 26.8% is an average of very different performances across targets and cannot be applied to each specific case. According to a report from Anthropic, in silico confidence scores did not predict these failures, as MBP and BBF-14 received scores similar to those of the successful targets. Taken together, these elements suggest that a new target may experience much more variable and often less favorable outcomes than the average rate implies.

What the Campaign Actually Achieved and Validated

On August 25, 2026, Anthropic published the results of an autonomous protein design campaign. The campaign, which had a summary released on August 18, 2026, reports 354 bindings out of 1,320 designs, or 26.8%, compared to an industry standard of 10 to 15%. Two distinct contracted laboratories conducted independent validation of these results. The key figures, including the 0 out of 90 and the 26.8% rate, come from the same campaign, using the same prompt and model. A dozen open-source tools dedicated to protein design were coordinated by the autonomous agent, which achieved better success rates than human experts on the majority of targets. Nevertheless, the 26.8% rate corresponds to an overall measure, while practical applications generally focus on a single target.

What Estimates from Published Accounts Show

The 26.8% rate represents an average across sixteen targets, not the probability of success for a given target. A binomial test applied to the results by target rejects the idea of a shared unique rate with a p-value of approximately 1.8 × 10⁻⁶³. Taking into account the differences between targets, the estimate indicates that a new target has about a 38% chance of falling below the 10% threshold observed in the industry. Assuming a unique rate of 26.8%, achieving 0 successes out of 90 for the maltose-binding protein would have a probability of 6.3 × 10⁻¹³, or about one chance in 1.6 trillion. A beta-binomial adjustment to the eight available accounts places the 90% predictive interval between 0.1% and 84.7%, with a median of 18.6%. Based on this, a batch of 30 designs facing an unseen target has a 20.1% chance of producing no bindings, compared to 0.0086% if a shared rate were assumed, representing a factor of about 2,300.