Claude's drug work: mostly preclinical.

Anthropic published a progress report on Oct. 27 about using Claude in drug development with six outside organizations. Its own description of the stage is the headline: so far the work is "mostly preclinical", which Anthropic explains as, for the most part, not yet at the point where drugs are tested in humans.

The examples run along the pipeline. With Generate Biomedicines, Claude designed shortened genetic sequences for CAR T-cell therapy; Anthropic says one design was about 70% shorter than the standard version and, in lab tests, let immune cells make roughly 10 times as many receptors. In work with Divergence Labs on food allergies, Claude generated roughly 35,000 antibody designs in about a month. In a separate project on two antibody-fragment leads, BigHat's lab made and tested about 1,010 variants in about 10 weeks.

The result closest to ordinary office work involved paperwork. With Enveda, Claude reviewed 1,061 pages of documentation, caught all 11 issues that human reviewers had found and flagged seven they had missed. Anthropic says the method reduced human review time by 89%, from over 25 hours to less than three.

Read as evidence, this is a set of partner case studies reported by the model's maker. Every figure above is Anthropic's account of work with its partners, and this issue found no independent measurement of them. The lab numbers also depended on people: BigHat's scientists made and tested the variants Claude proposed.

The document-review result is the one most teams could recognize, but the 89% figure describes one partner's test, and this issue could not find how review time was measured. Anthropic also says it has already prioritized specific improvements to its models based on what it has learned. The next question worth tracking is whether any of these programs reaches human trials. The post gives one conditional date: if one researcher's drug candidates continue to hold up, Anthropic expects a drug to be ready to start human trials as soon as late 2028.

Not tested: this is a note from GitHub's changelog, not a workflow this issue ran. Since Oct. 27, in public preview, a GitHub App installed on an organization can hold the read-only Copilot Requests permission, and a service can pass a short-lived installation access token to the Copilot SDK. An agent can then make requests for the organization without relying on any one person's signed-in credentials.

GitHub says the usage is attributed and billed to the organization that owns the installation, and its model policies, budgets and spending controls apply. Organization owners can find app usage in AI usage CSV reports under an app_client_id field, and suspending or uninstalling the app stops its access. Which services get the permission, and under what budget, remains a person's decision.