b12 labs
b12 labs is an AI platform for chemical synthesis, built for pharma and biotech R&D teams. Selenium plans retrosynthesis routes, Palladium designs condition-optimization plates, and Cobalt turns approved methods into automation-ready lab workflows.
What is b12 labs?
b12 labs is the platform published by b-12, Co., a three-person company from the Y Combinator Summer 2025 batch whose founders and research scientist come from Roche, Chemspeed, the Max Planck Institute and EPFL. It builds AI software for chemical synthesis planning, experiment design and lab automation, presented as a full-stack layer covering both what to run and how to run it.
The platform is made of three tools, each usable on its own. Selenium is a multi-agent retrosynthesis system built on frontier reasoning models: it analyses a target, proposes a strategy, expands it into a route, critiques it and iterates. On a benchmark of 100 molecules with published syntheses, the publisher reports that Selenium recovered the expert strategy in 80% of cases against at most 20% for traditional retrosynthesis systems, and closed the feasibility gap to under one point of the expert reference, where general-purpose coding agents recovered the idea roughly 56% of the time but stayed more than 12 points behind. Palladium is a plate designer: it reads the reaction, mines precedents from papers, patents and past screens, adapts to the substrate and lays out a structured plate under high-throughput experimentation constraints, where classical Bayesian optimization or DoE campaigns typically need three to ten screening rounds. The publisher reports 91% yield on a first in-house plate, and above 95% at early partners. Cobalt is an automation copilot that turns plain-language intent into executable workflows without specialised programming.
The audience is discovery and process R&D: medicinal chemists, HTE teams, automation scientists and lab leaders in pharmas, biotechs, CROs and university research groups. No code is required, the approach is claimed to be non-black-box since the references behind each well can be inspected, and a documented REST API covers the route search.
The claims are precise but they come from the publisher. The homepage figures (2 vs 5 steps to target, 70% faster route design, 99% yield on a first plate, over 80% conversion), the benchmark and the partner yields are internal results or early-partner feedback, with no independent verification. The academic record is firmer: twelve selected publications across eight venues between 2019 and 2026, including a 2026 Matter paper reporting 71% alignment with independent expert chemists.
What it does
- Plan a retrosynthesis from a target molecule and return ranked strategies
- Break each strategy down into concrete steps, each carrying a model confidence indicator
- Design a condition-optimization plate (catalysts, ligands, bases, solvents) for a single campaign
- Translate optimized conditions into executable protocols for Chemspeed, Opentrons, Hamilton, Tecan and Unchained
- Identify commercially available starting materials through eMolecules, with CAS enrichment from PubChem
- Assess how feasible a route is, through feasibility bands and judge scores
- Launch asynchronous route searches over a REST API, with a credit quote issued before any spend
When to use b12 labs / When not to
A quick filter to help you decide if b12 labs is the right fit.
When to use b12 labs
- Medicinal chemists and synthesis planners who need ranked, strategy-first routes to complex or multi-step targets
- Process chemists and high-throughput experimentation teams designing condition-optimization plates
- Automation scientists and lab engineers who must turn approved methods into instrument-ready protocols
- Academic and nonprofit researchers, who have been offered free access since 22 July 2026 with no application or demo request
- R&D leaders in pharmas, biotechs, CROs and university groups whose robotic platforms are underused for lack of a software layer
When not to use b12 labs
- Teams working outside synthetic chemistry: nothing here covers biology, clinical work, formulation or general office tasks
- Buyers bound by procurement rules: no terms of service and no privacy policy are published, and both footer "Legal" links are simple mailto addresses
- Organizations that need a price in currency before committing: no monetary amount is published anywhere, and full access is granted on request rather than by self-service checkout
- Users expecting a mobile app, a browser extension or a non-English interface: the platform is web and API only, in English only
- Anyone expecting the AI to run the chemistry itself: it plans and designs, but execution still requires a laboratory, instruments and a chemist to validate every protocol
How to use b12 labs
A typical end-to-end flow, from setup to results.
- Open the web application at app.b12-labs.com, where Selenium, Palladium and Cobalt are reachable separately
- Sign up: academic and nonprofit researchers are offered free access with no application file and no demo request, while other accounts start in trial mode with a limited number of remaining runs
- With Selenium, submit or describe a target molecule (SMILES) and receive ranked strategies broken down into steps
- Review the proposed routes, their step-by-step confidence indicators and the commercially available starting materials suggested for them
- With Palladium, describe the optimization objective in plain language and receive a plate layout (for example 24 wells in a 4x6 grid) with ligands, bases and solvents, each well backed by an inspectable reference
- With Cobalt, convert the selected conditions into an executable protocol for the laboratory instrument in use
- For programmatic use, create a B12 API key and read the documentation and the OpenAPI schema; the Codex skill package can be installed for agent-driven work
- Call /v1/capabilities and /v1/credits, request a quote, then submit the route search with a stable Idempotency-Key
- Poll the job according to the retry_after_seconds value returned, then fetch the results with GET /routes
- Buy prepaid credit packs through Stripe from the API console, and write to the publisher for full access or for an instrument that is not yet covered
Pros & Cons
Pros
- Published, quantified comparison results: the expert strategy recovered in 80% of 100 benchmark targets, against at most 20% for traditional retrosynthesis systems
- Traceability is a design claim rather than a black box: the references and the reasoning behind each proposal can be inspected
- No programming needed, since the whole platform is driven in natural language
- Instrument-agnostic output, covering Chemspeed, Opentrons, Hamilton, Tecan, Unchained and manual setups
- Free access for academic and nonprofit research, which makes evaluation nearly costless for a research laboratory
- A documented API with an OpenAPI schema, a quote issued before any spend and idempotency keys to avoid double charges
- An explicit statement that submitted structures are not used to train models, together with TLS 1.3 in transit and encryption at rest
Cons
- No terms of service and no privacy policy are published: both "Legal" links in the footer are mailto addresses
- No price in any currency is published, anywhere: there is no pricing page and the sitemap lists only eleven URLs
- No GDPR mention at all, no Article 27 EU representative and no data processing agreement
- No hosting country or region is disclosed, so the jurisdiction the data sits under is unknown
- Full access is granted on request rather than through self-service sign-up, and no postal address is published
- The performance figures come from internal tests and early-partner feedback, with no independent verification
- A very young company with a three-person team, on a domain registered on 30 June 2025, and a narrow scope: synthesis chemistry only, web and API only, English only
Pricing & Plans
A free plan exists, but it is targeted: since 22 July 2026 free access has been open to academic and nonprofit researchers, with no application file and no demo request, and the application additionally grants trial credits on verified sign-up. No lowest paid price point can be stated, because no amount in any currency is published anywhere: not on the website, whose sitemap lists eleven URLs and no pricing page, not in the application, and not in the API. Consumption is metered in credits only, at four route search tiers: Scout 10 credits, Basic 50, Standard 100 and Exhaustive 300. Credits are bought as prepaid packs through Stripe and the packs do not expire. Every search is quoted in credits before submission, the quote itself debits nothing, and credits are refunded if a search is cancelled before it starts. Full access is obtained on request from the publisher.
- Free access for academic and nonprofit researchers
- announced on 22 July 2026
- with no application file and no demo request
- trial credits granted on verified sign-up
- with a limited number of remaining runs
- 10 credits
- 50 credits (1 researcher run
- 3 route families
- 1 route
- 30 minutes)
- 100 credits (5 researcher runs
- 8 route families
- 3 routes
- 90 minutes)
- 300 credits (12 researcher runs
- 24 route families
- 6 routes
- 360 minutes)
- Prepaid credit packs bought through Stripe
- with no expiry date
- Contracted custom profile
- arranged through the quote endpoint
- none of these tiers carries a published price in any currency
Data, GDPR & hosting
A consolidated view of how b12 labs handles your data.
GDPR overview
There is no GDPR mention of any kind. A search across the 36 archived text and HTML files returned zero occurrences of "GDPR", "RGPD" or "General Data Protection". No privacy policy is published: the footer "Privacy Policy" link is a mailto to security@b12-labs.com, and so is the "Terms of Service" link. No Article 27 EU representative and no data protection officer are named, and nothing is said about legal basis, retention periods, data subject rights or sub-processors. The publisher is United States-based (San Francisco, Y Combinator Summer 2025 batch). The closest public statement is a security and non-training paragraph in a blog post dated 22 July 2026. A European user therefore has no published information on how personal data is handled or transferred, and no documented route to exercise GDPR rights.
Who owns the data?
No terms of service are published, so no contract states who owns the data: both "Legal" links in the footer are mailto addresses to security@b12-labs.com. The only public statement is a blog post dated 22 July 2026, not an enforceable document. It says the user's chemistry stays theirs: structures submitted are used only to plan that user's routes, are not used to train models and are not shared with other users, with TLS 1.3 in transit, encryption at rest and access protected by strong authentication. No data processing agreement, no licence clause and no governing jurisdiction is published, so nothing published binds the publisher beyond that single post.
Reuse rights
No terms of service exist, so no document grants or restricts what the user may do with the outputs. That absence is a gap, not a permission. In practice the outputs are synthesis routes, plate layouts and machine-ready protocols meant to be executed in the user's own laboratory, and the REST API lets routes be retrieved and downloaded as JSON through /v1/route-searches/{job_id}/routes. The only published usage instruction is an agent guideline asking not to expose internal identifiers unless debugging support requests them. Nothing published forbids reuse and nothing published authorises it in writing, so a team with a compliance requirement should obtain a written statement from the publisher before relying on the outputs.
Data retention & training
Hosting summary
No hosting country and no hosting region is published. There is no trust or security page, no privacy policy and no terms of service, so the jurisdiction under which user data is stored and processed is simply unknown. The only public technical elements are the ones stated in a blog post of 22 July 2026: TLS 1.3 in transit, encryption at rest, and access protected by strong authentication and security controls. The website resolves to a Cloudflare anycast address (104.26.14.62, AS13335), which locates a CDN node and says nothing about where user data lives; that distinction matters and should not be read as a hosting disclosure. The API sits on the api.b12-labs.com base and its /v1/ responses are returned with Cache-Control: no-store. The publisher, b-12, Co., is a United States company from the Y Combinator Summer 2025 batch, based in San Francisco, which makes United States hosting plausible but unconfirmed.
Things to keep in mind
Risks and trade-offs to weigh before adopting b12 labs.
- No published contractual framework: submitting proprietary structures without terms of service or a data processing agreement exposes trade secrets with no written recourse
- The only non-training guarantee comes from a blog post, not from an enforceable document, and a blog post can be edited or removed without notice
- No hosting country is declared, so it is impossible to know which jurisdiction the submitted structures pass through, and no GDPR information is available to a European user
- Over-trusting a proposed route: the performance figures come from internal tests and early-partner feedback, without independent verification
- Chemical safety: a generated protocol still has to be validated by a chemist, and the tool does not replace laboratory risk assessment
- Erosion of planning skills if junior chemists delegate retrosynthesis without understanding the reasoning behind it
- Credits are committed at submission and a misread quote is paid for, while depending operationally on a one-year-old, three-person company carries its own risk
Setup & Integrations
Technical difficulty
Two very different paths. Through the web application, setup is immediate: nothing to install, no code, and the publisher states the platform can be used straight away, with lab automation claimed to be learnable in hours rather than weeks. Through the API, it is developer work: create a key, read the OpenAPI schema and handle an asynchronous flow of quote, submit with an idempotency key, poll and fetch. Downstream, real use still requires an instrument park and a chemist to validate the protocols, and full access must be requested, self-service being limited to trial mode and academic access.
Deployment
Integrations
Supported languages
Behind b12 labs
Fundraising
Social
Resources
All the official URLs gathered for verification and reference.
Frequently asked questions
What kinds of molecules can Selenium plan?
How do Selenium, Palladium and Cobalt fit together?
Can each product be used separately?
Which instruments does Cobalt support?
Do I need to write code?
Is there free access?
How much does it cost?
Is there an API?
Are the structures I submit used to train models?
Are there terms of service or a privacy policy?
Should you pick b12 labs?
b12 labs is a narrow tool with a serious scientific footing. Twelve selected publications, a benchmark of 100 molecules, and a team drawn from Roche, Chemspeed, the Max Planck Institute and EPFL make it more than a demo, and its strongest structural idea is the coverage of the whole chain: planning the route with Selenium, designing the optimization plate with Palladium, then handing an executable protocol to the instrument through Cobalt, all without writing code.
The reservation is not scientific but contractual, and it is severe. Nothing legal is published: no terms of service, no privacy policy, no GDPR mention, no data processing agreement, no postal address and no declared hosting jurisdiction. The two "Legal" links in the footer are mailto addresses. The reassuring commitments on non-training and encryption exist, but they live in a blog post dated 22 July 2026, which no legal department can enforce. For a laboratory submitting proprietary structures, this is the deciding point.
Commercially, the picture is just as open. No amount in any currency appears anywhere; usage is counted in credits, at 10 for Scout, 50 for Basic, 100 for Standard and 300 for Exhaustive per route search, with prepaid packs bought through Stripe, and full access granted on request. Budgeting therefore requires a conversation with the publisher. The company itself is barely a year old, with three people and a domain registered on 30 June 2025: a continuity risk worth weighing before an industrial dependency.
The sensible path is the one the publisher itself opens. Free access for academic and nonprofit researchers makes an evaluation almost costless, and the quantified claims can be tested on molecules a team already knows. Promising, worth trying, and not yet ready for a procurement-bound purchase.
- Choosing a selection results in a full page refresh.
- Opens in a new window.