LatchBio
LatchBio runs a cloud workspace where biotech, pharma and academic research teams store, analyze and visualize omics data, alongside an AI research lab that benchmarks how well agents analyze biological data. In the Latch Console, Latch Data holds files, Registry ties samples to their metadata, Workflows runs repeatable pipelines, Pods gives on demand compute for notebooks, and Plots brings transformations and interactive figures into one workspace. The console's analysis assistant, built around Plots, helps prepare datasets and organize outputs for a scientist to review, and an optional connector lets customers use Anthropic's Claude with their Latch workspace. Scientists can run conventional analyses or inspect every result in a notebook.
Doloromics used the platform to give staff access to its chronic pain research data and internal single cell analyses, and its scientific lead reported validated targets in its pipeline rising from four or five to 12 to 15 over the next year. LatchBio says it serves more than 300 pharma, biotech and academic teams. Since December 2025 it has released a run of agent benchmarks, from SpatialBench and scBench to TxBench, OncologyBench and the BioSecBench biosecurity series, testing frontier models on real datasets. The San Francisco company was founded in 2021 out of Berkeley, raised a $28 million Series A in 2022, and has Kenny Workman leading its benchmark research. Platform access and professional services are sold by order form.
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
The assistant is one feature on a data platform that works without it. Built around Plots, it helps a scientist organize data and prepare analysis artifacts, and the version a Berkeley student group evaluated worked inside Plots on discrete research tasks. Everything else stands on its own: the registry for samples and metadata, interactive visualization, notebook compute in Pods and a workflow engine that runs pipelines with no model involved.
Those tools carry the data management, reproducibility and collaboration a customer buys. The company's agent benchmarks put models at the center of its research, but a customer can deploy the whole workspace and never call the assistant.
Scientists control what goes into an analysis and can inspect inputs, outputs and provenance in the workspace, and the terms make the customer responsible for evaluating information and deciding whether it fits its use. In the evaluation of the original console agent, a scientist reviewed the work at each stage before it advanced. The current assistant keeps outputs reviewable and owned by the customer, but no approval step, stopping rule or escalation is set out for it.
The company's own research points to why one matters: its November 2026 respiratory virus countermeasures benchmark found the best agent configuration passed 30.2% of tasks and concluded that current agents are not reliable on that work.
Plots provides a notebook style environment with data transformations and interactive figures, and in the early console agent evaluation the system wrote code and reports for discrete research tasks. The terms allow the platform to use third party AI tools. Separately, the Anthropic connector lets customers use Claude with their Latch workspace, a route the customer chooses.
The model inside the built in assistant is not named, and neither are its training sources or how it changes between releases. The company's benchmark papers test many named frontier models, Gemini and Claude among them, across agent harnesses, which describes the field it studies rather than the model it ships.
The platform agreement says customer content can pass to third party services that supply AI tools so they can deliver the service. Research data rests in Amazon S3, and customers can authorize connections to their own research applications. Anthropic lists LatchBio as the maker of an optional Claude connector, which settles the supplier for that one route.
For the built in console assistant, no model provider is named, and no retention or training terms are set for the parties that receive a prompt and the data attached to it. The hosting route, the customer authorized integrations and the assistant's inference path are three parts of the chain, and only the first two are described.
Doloromics reported its pipeline of validated chronic pain targets growing from four or five to 12 to 15 in the year after it adopted LatchBio, with shared access to data and internal analyses across research staff. That result belongs to the platform as a whole. A Berkeley student group tested the first console agent over ten weeks on research data workflows and limited its conclusions to that version.
The company's benchmarks use deterministic grading across hundreds of evaluations and publish pass rates by model and harness, but they measure third party frontier models rather than the assistant in the console. No measured effect is tied to the current release.
The privacy policy says customer data is not used to train machine learning models. The platform terms allow customer content to be shared with providers of AI tools so they can deliver the service, and an enterprise processing agreement can replace the standard treatment. Personal information is kept as long as needed and then deleted or anonymized, with no fixed schedule for inference inputs, logs or backups.
Encrypted storage, workspace permissions and immutable activity logs protect research data. Scientists can review analysis artifacts before relying on them, and no procedure is set out for an assistant that produces a wrong analysis or for an incident at an AI tool provider.
The privacy policy reaches the service and covers uploaded experimental data, and it distinguishes the standard policy from enterprise agreements that can govern processing on a customer's behalf. Enterprise material has described HIPAA support, and the platform offers access controls and encryption. The service terms put responsibility for the right to upload personal data on the customer.
No business associate agreement is offered in the standard terms, no tier is named as eligible for one, and nothing says how such an agreement would reach the AI tool providers the terms allow. A HIPAA claim without the agreement behind it leaves the position to enterprise contracting.
The company describes annual penetration testing and a SOC 2 Type II audit for the platform, along with encryption, granular workspace permissions, immutable activity logs and monitoring. Those controls cover the data and compute service. The trust center asks visitors to contact the company for access, so the report's dates, auditor and scope come only on request, including whether the console assistant sits inside the audited system. The SOC 2 type is named.
LatchBio also runs a biosecurity program, benchmarking how agents handle pathogen surveillance and risk assessment, which addresses misuse of models in biology rather than the security of customer data.
The platform is research and development software. It keeps versions, provenance and related records for drug development and regulatory work, with validation and traceability controls for teams preparing an investigational filing, so researchers can revisit the inputs and outputs of an earlier run rather than reconstruct it. The material separates the customer's eventual submission from what the platform provides.
The assistant analyzes research data, makes no diagnostic or treatment claim and carries no device authorization claim. Its use sits upstream of any patient, in discovery and preclinical analysis, which matches how the product is sold.
Research outputs sit beside their inputs and provenance so a scientist can review the work, workspaces limit who can see and change data, and versioning preserves the evidence behind a result. The company's research arm studies agent performance and alignment in the life sciences across a growing benchmark series, including a paired performance and refusal measure for biosecurity risk. That work shows how the company thinks about model behavior.
It does not give a console user an evaluation of the deployed assistant by data type or assay, or a change control process for when the assistant's model is updated, so governance of the shipped feature rests on traceable workflows.
The terms include a limited warranty that the platform will operate materially as described when used according to its documentation, with a reasonable effort to repair the affected part as the remedy. The same terms disclaim warranties about the accuracy of results and disclaim liability for third party AI tools, and aggregate liability is capped at what the customer paid in the 12 months before the claim, subject to applicable law.
A scientist can inspect provenance or rerun a workflow to check an output. The repair commitment covers the platform working as documented, so a wrong scientific conclusion from the assistant falls outside it.
The platform connects laboratory and research data through a workspace, a registry, an API and software development kits, and moves data between the research environment and tools such as Benchling. Results go to authorized collaborators, and the assistant's main output is an analysis artifact stored beside its data. The company also bundles kits and machines for laboratory work. Every connection serves research and the lab. No EHR is named, there is no route to read or update a clinical chart, and no record system receives the assistant's outputs, which fits a product used in discovery rather than care.
The hosted console keeps research data in Amazon S3, and enterprise customers can keep data in their own cloud environment. Workspaces and organizations govern who can reach each dataset, and storage and backups are encrypted with recovery across regions. The privacy policy allows transfers to the United States or other locations depending on service providers. That covers where research storage can sit and how access is divided. The regions themselves are not named, and the assistant's inference route, which third party AI tools may handle, is left to the deployment terms.
Access and professional services are sold through order forms under the platform terms, and the console still offers account signup. Compute and storage were once priced through resource credits, but those rates no longer stand as a quote, and the former pricing address now leads to the company's research homepage.
Integration, onboarding and training are bought separately on a professional services order, and the platform now comes bundled with laboratory kits and machines, which widens what a single order can include. A buyer can tell that resource use, services and hardware drive the cost. Current rates, assistant allowances and minimum terms come only through contact.
The platform serves biotech, pharma and academic research groups that share molecular and experimental data: single cell, spatial, epigenomic and proteomic datasets that can run to terabytes per experiment. Doloromics used it to open chronic pain research data and internal single cell analyses to scientists with different informatics experience, and the company counts more than 300 teams.
Its benchmark program spans spatial biology, single cell, epigenomics, pharmacology, oncology, neuroscience, rare disease and biosecurity. The named customer supports a discovery setting. The current assistant has no performance figures tied to any research specialty or stage, even though the company's benchmarks test agents in exactly those areas.
Alternatives to LatchBio
Other products that do the same kind of work as LatchBio, and what each one does.
Causaly supplies scientific AI software for pharmaceutical, biotechnology and academic research. Its agents combine biomedical literature, knowledge graphs and an organization's internal evidence to investigate questions, compare findings and produce assessments.
Edison Scientific is the commercial spinout of the nonprofit research lab FutureHouse, formed in November 2025 to run its AI scientist platform.
Medra builds a Physical AI Scientist for preclinical drug research: robotic lab workcells and scientific AI that design, run, interpret and improve experiments in a closed loop.
BenchSci develops EMET, an AI research environment for scientists in biopharma and academia. The platform combines scientific literature, experimental databases and internal research with a biomedical knowledge graph.
Lila Sciences makes very large claims about what its platform can do, and as of September 2026 it has begun publishing evidence for some of them.
South San Francisco company founded in 2023 and launched publicly in April 2024 with more than $1 billion in committed funding, co led by ARCH Venture Partners and Foresite Labs.
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 |
|---|---|---|---|---|
|
Custom quote
|
Platform fees and professional services are set by order form under the platform terms. Earlier resource credit rates for compute and storage no longer serve as a quote, and assistant pricing is not listed. | The standard platform terms offer no business associate agreement and name no eligible tier. | The terms put integration, onboarding and training into a separate professional services order form when purchased. The scope and fee depend on the engagement. | Vendor Published |
The console remains open for account signup, while the former pricing page leads to the research homepage. Earlier credit prices do not show what customers pay today for access, compute or the assistant.