Zephyr AI
Zephyr AI sells oncology software and real world data to biopharma and diagnostics companies through its Zephyr Platform. The AIM Suite holds the models. AIM-Bx and AIM-io predict a patient's response to small molecule drugs and immune checkpoint inhibitors from clinical and molecular data, AIM-CDx adapts those models into companion diagnostics co developed with partners on their own assays, and AIM-Ex reconstructs gene expression from DNA based laboratory tests with high stated concordance to RNA sequencing. Nexus turns record, claims, genomic and pathology data into cohorts, synthetic control arms for trials and health economics work, and intervention specific risk models.
Biopharma partners fine tune the models for patient stratification, trial optimization and indication expansion and validate them in real world cohorts, while diagnostics partners add research use expression signatures to existing DNA or methylation panels without new assays. The data covers more than 90 million patient lives, 6 million in oncology and more than 1 million with next generation sequencing, plus more than 400,000 lifetime consented patients from the ORIEN Total Cancer Care study at leading cancer centers, reached through Aster Insights. At AACR 2026, Zephyr reported its response model separating outcome groups with a hazard ratio of 0.23 in held out data. Guardant Health is a research partner, and work starts with pilots that scale into development programs.
Capability Axes
An AI Health Index grade measures what a buyer can verify from public sources on the date shown. It is not a rating of how good the product is. A vendor can build an excellent system and grade low on an axis because it publishes nothing an outsider can check. How grades read
Models and data both carry the offer. AIM-Bx, AIM-io, AIM-CDx and AIM-Ex exist only as model output: response predictions for small molecules and checkpoint inhibitors, companion diagnostic models adapted to a partner's assay, and gene expression reconstructed from DNA panels. Nexus and the ORIEN cohorts are different.
Linked records for more than 90 million patients and 400,000 lifetime consented cancer patients support cohort construction, synthetic controls and outcomes analysis, the real world data business Aster Insights ran before the models arrived. Remove the models and the data and cohort tools still sell, while the AIM Suite does not exist without them.
Outputs are research analysis for scientists and program teams. Partners compare AIM predictions and Nexus cohorts with trial data, clinical histories or other cohorts before using them in enrollment strategy, biomarker decisions or a diagnostic submission, and adapting a model to a drug or assay happens jointly with Zephyr inside a development program. Clinical validation and deployment decisions stay with the program teams. No sign off step, confidence threshold or route for flagging a doubtful prediction is described.
AIM runs Zephyr's own multimodal models for clinical, molecular and pathology inputs, adaptable to a single drug or assay. AIM-Ex derives expression estimates from DNA panels that can feed response modeling, and the models are built to show the biology linked to each prediction. The AACR 2026 pathology proof of concept used embeddings from the GigaPath and UNI foundation models, with comparisons against conventional biomarkers. The production model versions and the full data lineage behind each partner's adapted module are not published.
Zephyr's own models drive AIM, with outside foundation model embeddings such as GigaPath and UNI used in research, and data comes from Aster Insights and the ORIEN network of cancer centers. The privacy terms let service providers, which may operate globally, process information under Zephyr's instructions and confidentiality safeguards. No hosting provider, inference supplier or subprocessor list is named for the licensed modules.
The AIM response model was validated retrospectively across tissue sequencing, liquid biopsy and research pathology images. The AACR 2026 abstract describes a source cohort of about 3,000 patients across 13 tumor types, and in held out laboratory test data the model separated outcome groups with a hazard ratio of 0.23, with a 95% confidence interval of 0.14 to 0.41. Guardant Health is a named research partner for drug response predictions and biomarker insights. No prospective study or independent evaluation of treatment selection has been reported, and no biopharma program outcome is public.
De identified clinical cases from approved research collaborations feed Zephyr's work, and its privacy terms allow de identified information to be used for internal validation and development. Zephyr must get written consent before contacting a patient directly about an opportunity found through a commercial research collaboration, service providers process information under its instructions and confidentiality safeguards, and ORIEN data rests on consent and controlled access. Whether identifiable partner records can train models, how long they are kept and when a deployed model's data is deleted are left to each program agreement.
HIPAA compliance is claimed on the platform page, and the privacy terms reach Zephyr's software and APIs, allowing de identified information to be used for quality work, internal validation and research and requiring written consent before Zephyr contacts a patient about an opportunity found through a commercial collaboration. ORIEN data rests on patients' lifetime consent and controlled access.
Business associate status is not stated and no agreement is published, so a diagnostics partner sending identifiable results and a sponsor using prepared research data each work under what their own contract says.
SOC 2 compliance is claimed on the platform page beside HIPAA, with no report type, audit period, system boundary or trust portal, so its scope is a question each partner raises. Service providers must follow Zephyr's instructions and apply confidentiality and security measures, and the ORIEN research data sits behind controlled access. Security for a given program also depends on where its model and data are hosted, which the program agreement sets.
Research use is the current position. AIM supports biomarker research, patient stratification and companion diagnostic development, AIM-Ex signatures are offered for research use on existing panels, and models move into validation programs before any clinical deployment. AIM-CDx is co developed with partners for specific drugs and assays, so any finished diagnostic would carry its own intended use and submission.
No FDA authorization covers an AIM companion diagnostic or a treatment recommendation today, and the platform page describes AIM-Bx and AIM-io as able to serve as AI companion diagnostics for partner drugs.
Models are validated across cohorts, tumor types and input types, from tissue sequencing to liquid biopsy and pathology images, so performance can be compared across data sources, and ORIEN governs consent and access for its longitudinal records. No results are broken out by ancestry, sex, age or treatment setting, which matters for response models trained largely on data from major cancer centers, and no monitoring of deployed partner models is described.
Partners use the software to compare cohorts, study treatment response and decide whether a biomarker merits further validation, and a companion diagnostic program carries its own intended use and trial evidence. The privacy terms address rights to information, and separate agreements govern the services delivered. Nothing published says what happens when a model is wrong, such as a trial cohort stratified incorrectly, and any remedy sits in the negotiated program agreement.
Nexus links record, claims, genomic and pathology data into research datasets, diagnostics partners plug research signatures into existing reporting flows with no new assay, and partners add their own trial and validation data as a program moves forward, with the privacy terms covering Zephyr's software and APIs. No named record system, standard interface or certified connection is offered; the platform builds research datasets rather than exchanging data with a clinical record.
Pilots launch on Zephyr's managed infrastructure within weeks and scale into production without being rebuilt, with outputs fed into the partner's research or diagnostics workflow. The privacy terms let service providers operating globally process data under confidentiality and security requirements. No hosting provider, country or region is committed for program data, and where a partner's model and data reside after a pilot is set in the implementation agreement.
Engagements start as pilots or proof of concept studies and grow into full development programs, on Zephyr's stated approach of starting small and scaling quickly. AIM modules can be licensed alone or together, and Nexus work is scoped to a cohort and an evidence question, sometimes combining partner data with Zephyr's linked sources. Software charges, billing units and the fees for preparing data or adapting a model are not published, and nothing indicates how they are calculated.
Oncology biopharma and diagnostics developers are the buyers, for research on small molecules and immune checkpoint inhibitors, patient stratification, trial optimization, indication expansion and companion diagnostics, and Nexus also covers chronic disease populations and health economics. ORIEN contributes longitudinal data from leading cancer centers, with deeper molecular and pathology profiling for its Avatar subset. Retrospective validation spans 13 tumor types and three data modalities. Use in routine patient care is outside the current offer.
Alternatives to Zephyr AI
Other products that do the same kind of work as Zephyr AI, and what each one does.
ConcertAI's Accelerated Clinical Trials platform helps pharmaceutical sponsors and contract research organizations design, plan and run oncology studies.
Oncology clinical trial matching and enrollment company based in Boca Raton, Florida.
Tissue intelligence company for drug development, applying computer vision and machine learning to H&E, immunohistochemistry, multiplex imaging and spatial transcriptomics and linking tissue features to clinical outcomes.
PathAI sells AISight, a cloud image management system that anatomic pathology laboratories use to manage cases, read slides on screen and run AI, along with its own algorithms and a menu of partner algorithms.
Guardant Health is a publicly traded precision oncology company (Nasdaq: GH), founded in 2012 and based in Palo Alto, California, that runs blood and tissue cancer tests in its own laboratories.
Precision oncology company in Waltham, Massachusetts, combining sequencing in its own CAP accredited, CLIA certified laboratory with models of tumor and immune biology.
Pricing
Vendor-published figures are labeled as such. Figures labeled “Estimated” are derived from third-party sources and have not been confirmed by the vendor.
| Entry Price | Pricing Basis | BAA Tier | Implementation | Source |
|---|---|---|---|---|
|
Contact for program pricing
|
Software is delivered through scoped partner pilots and development programs. The AIM modules can run individually or together. Scope can include testing against a trial population and adapting a model to a specific drug or diagnostic assay. | Zephyr makes a HIPAA compliance commitment. A business associate agreement for the platform is unavailable to review. | Implementation and validation fees depend on the scope of the partner program. | Vendor Published |
The software, research dataset and validation scope are agreed through the partnership process. Rates and billing units are negotiated privately.