b12 Labs
b12 Labs sells AI software to pharmaceutical and biotechnology chemists planning small molecule synthesis. Selenium takes a target molecule and returns ranked synthesis strategies, each broken into steps with a model confidence, after a group of reasoning agents drafts, critiques and revises the plans; a chemist can add constraints before any lab work. Palladium designs condition screening plates, laying out catalysts, ligands, bases and solvents across a 24 well plate, and Cobalt turns approved conditions into automation protocols for Chemspeed, Opentrons and other instruments without code. Teams can start in the web app with one tool or run all three as one workflow, and Selenium also runs as a plugin inside ChatGPT and Codex.
In a July 2026 company benchmark of 100 targets with published expert syntheses, Selenium recovered the expert strategy in 80% of cases, against about 56% for general coding agents and at most 20% for traditional retrosynthesis software, with no lab validation yet. For customers without their own automation, the Foundry service runs synthesis and data generation projects through a partner lab at the University of Zurich.
Zlatko Jončev and Andres M. Bran founded the company in 2025 out of Y Combinator, with researchers drawn from Roche, Chemspeed, the Max Planck Institute and EPFL, and it is based in San Francisco. Software use is paid in organization credits, and Foundry projects are priced separately.
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
Selenium uses reasoning agents to propose and evaluate routes for making a research compound, Palladium turns a scientist's goals into a plate design for comparing reaction conditions, and Cobalt converts an approved method into a workflow for the lab's instruments. The scientist sets the objectives and reviews what comes back. Each of the three exists to turn a research aim into a plan through a model. Foundry, the partner lab service, adds experimental capacity for customers without the equipment or staff, and is the one part of the offer that rests on chemists and instruments rather than models.
Selenium proposes routes with a model confidence for each option, and a researcher can add constraints and revise a plan before lab work. Palladium generates a proposed experiment layout and Cobalt prepares an instrument workflow from an approved method. The terms require qualified staff to verify the chemistry, instrument instructions and safeguards before a protocol runs, and the customer remains responsible for putting the plan into the lab. No confidence threshold blocks a weak route automatically, and no escalation path is set for a workflow that fails review.
Selenium runs multiple agents on frontier reasoning models: one develops a strategy, another turns it into candidate plans and others evaluate the proposals before the system revises them, with chemists adding lab constraints and new information along the way. Palladium and Cobalt produce their own planning artifacts for experiments and instrument use. The Selenium benchmark sets out its task, reference set and scoring method. No underlying model release is named, training data is not characterized, and no version history links a change in recommendations to a model update.
b12 uses outside AI and research providers, which can receive submitted materials or searches when a function needs them. The Selenium plugin connects to ChatGPT and Codex under their own terms, a link that names the assistant calling Selenium rather than the provider behind its planning agents. Stripe processes website payments and PostHog runs product analytics, yet the model providers for Selenium, Palladium and Cobalt, and the retention terms for the content they process, are not named.
Selenium's July 2026 benchmark took 100 molecules with published expert syntheses and asked whether each system recovered the expert's strategy, with an LLM judge scoring the matches. Selenium did in 80% of cases, general coding agents in about 56% and traditional retrosynthesis systems in at most 20%, and Selenium's routes came within a point of expert feasibility.
None of the 100 plans was run in a lab, and early pharmaceutical, biotech and university users have reported no experimental results. Claims of 70% faster route design and 99% yield on a first plate come without a method.
Structures submitted to Selenium are used for route planning and are not used to train models or shared with other users. The service terms authorize processing of customer content only to provide, secure, maintain and support the requested services, with no general training license.
Customer data is kept for the relationship and at least thirty days after it by default, with exceptions and a separate backup process, and the laboratory terms require qualified personnel to review the output. There is no deletion window for model providers and no defined response to an unsafe recommendation.
b12 processes the research inputs, uploaded materials and generated results an organization's users submit, and its service privacy policy explains when providers receive that material and how an organization controls account access and requests deletion. Selenium's plugin passes on the inputs relevant to each tool call, which can include context from a connected assistant. The product is sold for molecule planning and laboratory work and has no clinical record workflow. No business associate status, BAA tier or explicit exclusion of protected health information from submitted research material is stated.
The trust center lists a 2026 SOC 2 Type 2 report for b12 Labs, released on request, and shows controls for access, regular permission reviews and multifactor authentication. Data is encrypted in transit and at rest, and the control set covers monitoring, backups, incident response and vendor management, with a penetration test within the past twelve months. A buyer sees that summary before contracting; the audit period, and which tools or Foundry workflows the audit covers, come with the report.
Research laboratories use the software to propose synthesis routes, design chemistry experiments and prepare instrument workflows, and organizations can commission compounds or data through Foundry. The terms present generated plans as computational proposals, with no claim that they establish experimental success, suitability or regulatory compliance, and leave qualified researchers to judge the lab work.
There is no diagnostic or treatment indication and no device clearance, and nothing in the product touches prescribing, participant care or medical decisions, even when the molecules under study are potential medicines.
Qualified staff must review generated chemistry and instrument instructions before anything runs, and Selenium's ranked proposals let a chemist compare alternatives before choosing one. Route planning is scored against published expert strategies, with an LLM judge deciding whether a plan matches, and the terms allow b12 to restrict service for abuse and security threats.
The 100 target result is a single aggregate: no breakdown by scaffold family, reaction class or data coverage shows where the planning models are weaker, and there is no independent AI audit or model governance certification.
A chemist can examine and revise a proposed workflow before sending it to an instrument. b12's standard terms treat the output as an experimental proposal and disclaim any guarantee of a research outcome, and they cap each party's liability at the fees paid or payable for the affected service in the preceding twelve months, with exceptions such as fraud and willful misconduct. b12 corrects verified billing errors and reviews certain refund requests. Billing errors and refunds are the only remedies, and the terms carry no output warranty and no duty to correct or compensate for a failed experiment.
Cobalt writes workflows for automation platforms such as Chemspeed and Opentrons, Selenium runs as a plugin for ChatGPT or Codex tied to an existing b12 organization, and Foundry has an API for project work. Every connection serves chemistry and lab execution, since users work with molecules and experiment data rather than patient charts. There is no clinical record connector, HL7 or FHIR interface, or way to read or write clinical documentation, which fits a research product with no clinical workflow.
Selenium, Palladium and Cobalt run in a hosted web application. Connecting Selenium to an outside assistant authenticates the user's existing b12 organization and sends the tool input to b12 for processing, and b12 stores the requests and results it needs to provide the service. An organization manages its own users and permissions.
The privacy terms let providers process information in countries other than the customer's, and Foundry's partner lab sits in Zurich, which places the experiments rather than the software. No storage region, inference location or customer hosted option is stated for the software.
Usage is billed in credits held by the customer's organization. A chargeable operation shows its credit cost before it runs, the terms hold the price of an accepted operation quote, and purchased credits stay with the organization that bought them. Scientists can start in the web app and connect the Selenium plugin to their organization's balance, with Stripe taking card payments on the website. Credit packs carry no public dollar rate, so the cost of a route or a plate only appears inside the purchase flow, and Foundry synthesis and data projects are scoped and priced one by one.
The software is built for medicinal chemists, process chemists and lab automation specialists in early drug discovery. A team picks the tool that matches its work, from planning a route to comparing conditions or preparing an instrument workflow, with Cobalt supporting Chemspeed and Opentrons and Foundry taking work from labs without compatible equipment. Early users span pharmaceutical, biotech, contract research and university groups, none named. The 100 target benchmark covers route planning against published chemistry, not a therapeutic area, a named pharmaceutical lab or the full workflow from plan to plate.
Alternatives to b12 Labs
Other products that do the same kind of work as b12 Labs, and what each one does.
Iktos is a Paris generative AI and robotics company for small molecule drug discovery, founded in 2016.
Variational AI develops Enki, a generative model for designing small molecules against a research team's target product profile.
Stanford spinout, now trading as Genesis Molecular AI, building GEMS, a small molecule discovery platform its own chemists and its partners' programs run on.
Matterworks, a Somerville, Massachusetts AI lab, sells Pyxis, a platform that reads high resolution mass spectrometry data, returns the identity and concentration of small molecules and lipids, and helps researchers work out what the changes mean biologically.
Latent Labs builds AI models that design proteins and antibodies from scratch, and runs them as a web platform that researchers can open in a browser.
PostEra is the medicinal chemistry counterpart to the biology platforms that dominate this category. Where Immunai, Cellarity and Enveda build models of cells, tissue or natural product chemistry to decide what to make, PostEra builds models of how to make it.
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 |
|---|---|---|---|---|
|
Usage credits
|
The software bills by operation against the organization's credit balance. A chargeable operation displays its credit requirement. Dollar rates for credit packs are unavailable before entering the purchase flow. | The service handles research content and has no stated business associate tier. | Researchers can access the software through the web application. Foundry experimental work is scoped and priced separately. | Vendor Published |
Stripe handles web payments. The Selenium plugin draws on existing account credits. Credits are tied to the purchasing organization. Foundry supports separately arranged compound synthesis and experimental data projects.