Friday, October 23rd, 2026
Sign in to LinkedIn, then select Subscribe.
FOUNDATIONAL UPDATES
Anthropic previews Claude Duet, an agent that comes to you first
On October 21, Anthropic opened a waitlist research preview of Claude Duet, starting with US Pro and Max subscribers. It's built around Claude Opus 5.5 and runs on its own cloud computer, so it keeps working after you close your laptop. Each morning it sends a short brief from your connected apps, flags what needs attention, and asks before it takes a big step like sending or deleting something.
Why it matters for pharma: Agents are moving from answering questions to doing the work in the background. That's useful for trial teams buried in email and trackers, but you need to know what an agent can touch and who signs off before it acts. Get your approval rules on paper before these tools show up on people's phones.
Source: https://claude.com/resources/articles/claude-duet
OpenAI models are coming to Google Cloud
On October 21, OpenAI and Google Cloud announced that OpenAI's frontier models and Codex will be offered in Google Cloud's Model Garden. Customers can buy them through Google Cloud Marketplace and count it against existing Google Cloud commitments. Broader access is planned in the coming weeks, and Microsoft remains OpenAI's primary cloud partner.
Why it matters for pharma: Model choice used to depend on which cloud your IT team signed with. That barrier is getting smaller. Teams can pick the model that fits the task and keep data inside the access controls and governance they already validated, which makes qualification and vendor reviews simpler.
Source: https://openai.com/index/openai-models-google-cloud
CLINICAL TRIALS & AI
FDA says draft guidance on generative AI devices is coming "in a matter of months"
FDA's comment period on its generative AI medical device discussion paper closed October 19, with more than 500 submissions in. On October 20, Rick Abramson, who runs FDA's Digital Health Center of Excellence, said draft guidance will follow within months and called it a priority up to the White House. The paper covers risk assessment, premarket testing and postmarket monitoring, including foundation models and agentic AI.
Why it matters for pharma: This is device policy, but it will shape how FDA thinks about GenAI everywhere, including tools used in trials. If you use AI for patient-facing apps, eCOA or companion software, expect questions about benchmarking and ongoing monitoring. Start documenting how your tools are tested now, not after the guidance lands.
Vysioneer reports AI-assisted RECIST reads in a live, pharma-sponsored oncology trial
On October 20, Vysioneer said its VTrial platform was deployed prospectively in a pharma-sponsored oncology trial to support efficacy assessment, including AI-assisted RECIST 1.1. Its ESMO 2026 abstract reports that an independent AI-assisted central read produced confirmed objective response and disease control rates consistent with local investigator assessments. The company calls it, to its knowledge, the first reported prospective use of AI-assisted RECIST 1.1 as an efficacy endpoint in a sponsored trial.
Why it matters : Most AI-in-trials work is still retrospective. This one touched an endpoint while the study was running. If you're thinking about AI in imaging reads, the questions are practical. Who signs the read? How are disagreements with the investigator handled? And what goes in the charter before first patient in?
Boehringer Ingelheim and Microsoft will test digital patient twins, starting in lung health
Reported October 23, the two companies are exploring computer models of individual patients built from clinical data to forecast disease progression and outcomes. The work includes testing whether those models can support synthetic or hybrid control arms, which could mean fewer people assigned to control groups. It starts with feasibility work through prototyping and modelling.
Why it matters : Fewer patients on placebo is a goal everyone shares. But a model-based control arm lives or dies on whether regulators accept it for that specific use. Early design, data provenance and a clear context of use will matter more than the model.
Two recruitment bets: Oracle with AstraZeneca, Optum with Syneos Health
On October 20, Oracle announced a non-binding agreement with AstraZeneca to explore using AI and real-world data to help clinicians spot eligible trial patients earlier, with oncology, cardiovascular, renal and metabolic, respiratory and immunology, and rare disease named as possible starting areas. The same day, Syneos Health announced a partnership with Optum covering trial scenario simulation, patient and site access, and the data and governance setup to scale it. Syneos will be Optum's preferred partner as Optum builds its site management organization.
Why it matters : Recruitment is still where timelines slip. Both deals point the same way: find the patient inside the health system, not after a site opens. Watch how consent, data privacy and site burden get handled once these move past announcements.
Sources : https://www.oracle.com/news/announcement/oracle-and-astrazeneca-partner-to-advance-ai-enabled-healthcare-2026-10-20/ and https://www.globenewswire.com/news-release/2026/10/20/3384520/0/en/optum-and-syneos-health-partner-to-advance-an-adaptive-clinical-development-model.html
Insilico launches inClinico Agent to predict whether a trial will hit its primary endpoint
On October 22, Insilico Medicine launched inClinico Agent, built on a database of more than 100,000 trial records. Its efficacy model scored an AUROC of 0.828 on a held-out test set of 2,347 trials. On a prospective benchmark of 3,311 ongoing trials, general-purpose LLMs predicted success for 75% to 82% of them. Insilico's agent predicted about 50%, close to the 53% historical rate in its test set. It also drafts 13 regulatory document types, including INDs and Investigator's Brochures.
Why it matters : That gap is the story. General chatbots are optimistic about your trial. A tool that's calibrated to real success rates is more useful for portfolio calls, as long as you can see why it scored a study the way it did.
Originally published on LinkedIn. Read the original
