Cypris vs PatSnap: how they compare in 2026
Cypris and PatSnap both sell R&D and IP intelligence built on patent data, with AI agents for technology scouting, prior art and freedom to operate work, mostly to corporate research and IP teams. Cypris sits in the top two bands on eleven of fifteen axes and PatSnap on six of fifteen, identical on four. Cypris publishes what a buyer needs before uploading research. It names OpenAI, Anthropic and Google as its model providers, keeps data in the United States with tenant isolation, and states that customer data trains no model. Its published agreement commits to written notice and an effort to obtain a protective order before any compelled disclosure, and indemnifies customers against infringement claims. PatSnap's lead is measurement and reach. It publishes PatentBench, a benchmark with its methodology, reporting a 77 percent hit rate on design freedom to operate, and offers REST APIs and MCP servers to developers. PatSnap publishes no position on training, hosting region or model provider, and its SOC 2 attestation is Type 1.
At a glance
All 15 axes, side by side
The same grid applied to every vendor in the index, graded from public sources. Hover a grade to see what the letter means on that axis.
AI Centrality
How much of the product is actually AI. Whether the machine learning is the mechanism the buyer is paying for or a feature layered onto conventional software, and whether the vendor is specific about which is which.
The platform is organized around agents and priced by model usage. The current estate leads with AI agents for every stage of the R&D process, purpose-built for technology scouting, literature review, competitive intelligence, prior art, regulatory research and white space analysis, with agents that retain project context and can be configured to a team's own methods. Semantic retrieval across the corpus runs through a proprietary ontology rather than keyword matching, and the agreement makes the metering explicit: AI credits are allocated annually and consumed per session at a rate reflecting the processing cost of the model chosen. Remove the models and a large licensed database remains, which is what the vendor positions itself against. The human limb is real and separate: Research Briefs are written by the vendor's analysts, with the estate calling them built for decisions where accuracy matters too much to leave to AI alone. Verified 20 September 2026.
A data platform that existed for over a decade before the models, with a substantial AI layer built on top of it. Founded 2007, and the asset underneath is the corpus: more than 2 billion structured data points and over 210 million patent, science and technology records with search, landscaping, dashboards and analytics that predate generative AI and function without it. Eureka is described by the vendor itself as a capability layer rather than a separate platform, which is an accurate self assessment and is credited as such. The AI layer is genuinely substantial rather than decorative, spanning domain specific agents for prior art, freedom to operate, drafting, translation and life sciences work, and the vendor describes a four stage data pipeline built to feed it. Graded B on the same basis as Relativity, Everlaw and Lexis+ AI: a mature platform hosts the model layer rather than depending on it. Remove the models and PatSnap remains a working patent analytics business. This is the first record in ip-and-patents below A on this axis, and the reason is age rather than weakness.
Citation Accuracy and Hallucination Disclosure
Whether the vendor publishes measured accuracy on citations and assertions, grounds output to primary sources, and says plainly what its system does when it does not know. Legal has a documented public record of fabricated citations reaching filed briefs, so an untested claim of accuracy is not evidence.
Grounding is documented in unusual detail and the vendor's own error rate is still not published. Every AI output is described as anchored to source citations, and the published comparative study makes that checkable: it states its method, names the date, preserves the outputs, gives the verbatim queries, and says the patent numbers and assignees were checked against USPTO Patent Center and WIPO PATENTSCOPE. The study also discloses that the vendor commissioned it and has a commercial interest, and records a place where its own tool fell short, failing to fully separate two related bodies of work. What is absent is a measurement of the platform's own precision or recall against a gold standard: the study counts what competitors missed rather than what Cypris misses. The agreement runs the other way, disclaiming any warranty that results will be accurate, complete or error free. Verified 20 September 2026.
THE FIRST PUBLISHED BENCHMARK WITH A STATED METHODOLOGY AND A NAMED COMPARATOR IN 51 RECORDS. PatentBench is published as a benchmark built for real patent work, stated to measure specific tasks against structured data, legal context and source linked evidence, with a full methodology the vendor links rather than summarizes. Published results: a 77 percent hit rate on design freedom to operate, and a claim of being 83 times more accurate than ChatGPT on that measure. Naming the comparator and the task is what separates this from the unfalsifiable claims elsewhere on this index, because a reader can locate the methodology and disagree with it. Grounding is described architecturally as well: retrieval augmented generation over a governed corpus, a four stage pipeline the vendor states reduces hallucinations through governance at each stage, and source linked evidence as a stated property of the benchmark. Graded A because this axis rewards verifiable evidence checkable by an outsider without contacting the vendor, and a published methodology with a stated result meets it. Held short of perfection and recorded plainly: the benchmark is the vendor's own rather than independent, the 83 times figure is a ratio against a general purpose chatbot on a task it was never built for, which flatters the comparison, and the methodology page itself was not read in this pass, making this a correction candidate in both directions.
Autonomy and Oversight Model
What the system decides on its own, what a lawyer must approve, and whether the vendor documents where the review point sits. A tool that drafts under review and a tool that files without one are different products and different risks.
A human limb exists as a separate product and no oversight mechanism is described inside the agents. Agents are published as executing research workflows across large datasets in parallel, generating deliverables and monitoring continuously, which is a real degree of autonomy over research rather than over a client matter. Against that, the vendor sells Research Briefs as human-in-the-loop work written by trained analysts, and frames them as being for decisions too important to leave to AI alone. That is a statement about a different product, not a checkpoint in the automated one. Nothing published describes what an agent decides on its own, at what point a person reviews an agent's output before it informs an investment or filing decision, or what the system does when its evidence is thin. The agreement is silent on oversight, and the vendor reserves the right to change the available models at any time. Verified 20 September 2026.
Agent language is used extensively and no oversight model is published. The platform is described as agentic across IP and life sciences, with named agents including an FTO Agent, a Lead Compound analyser, an Antibody Target Predictor and a Markush Claim Drafting agent, and a published customer account describes completing infringement analysis for new products in hours rather than weeks. Data governance is described at each stage of the pipeline, which is governance of inputs rather than of agent behavior. Nothing published states what an agent may complete without human review, whether any output can be relied on without verification, what confidence signal accompanies an agent result, or where a human checkpoint sits in an agentic workflow that runs from query to report. The gap matters because the outputs described are decision grade: a freedom to operate assessment that clears a product for launch, and a novelty judgment that decides whether a molecule is pursued. Checked the Eureka product pages, the life sciences landing page, the home page and the Open Platform material on 29 Aug 2026.
Operational and Outcome Evidence
Named, dated evidence that the product works in production at real firms or legal departments. Case studies with figures and identified customers count. Unattributed testimonials and launch announcements do not.
Named organizations, named people, and no measured outcome. Customers are named across three industry sections with what each uses the platform for: Johnson & Johnson for cross-domain research intelligence spanning materials science and pharma, Yamaha for scouting emerging technologies and landscaping suppliers, Oerlikon for monitoring compound activity and tracking competitive filings across global patent offices. Government and research customers named elsewhere on the estate include NASA, the US Air Force and Los Alamos National Laboratory. Two testimonials carry a name and title, Chuck Wright, Corporate R&D Manager at NOV, and John Parks, Senior Footwear Innovation Developer at Brooks Running, each with a published customer story. Scale is given as hundreds of R&D teams and thousands of researchers. What is missing is measurement: the strongest figure is that work which took weeks now takes days, with no baseline, period or method behind it. Verified 20 September 2026.
Named customers of the highest recognisability, a named executive, and independent third party presence, all dated and checkable. Named customers include NASA, Tesla, Vodafone, MIT and General Electric, reported in independent trade coverage rather than only on vendor surfaces, plus NCH Corporation with a named executive, Michael Schuster, VP of Product Development and Innovation, quoted by name and title on vendor material. Corporate history is a matter of public record: founded 2007, unicorn valuation reached in 2021 following a $300m Series E led by SoftBank Vision Fund 2 and Tencent, with Jeffrey Tiong named as chief executive. Independent evidence extends beyond a press release: a Gartner Peer Insights listing, a SoftwareOne marketplace listing, and a hands on product review by a named trade publication that obtained a trial and published its assessment. An outsider can verify the company, the customers, the funding and the product without contacting the vendor. Held short of perfection because outcome claims carry no methodology, being productivity boosted by up to 60 percent and workflows 90 percent faster, and one customer quote describing hundreds of thousands of dollars saved is not attributed to a named organization.
Privilege and Confidentiality Posture
How client confidences are handled: attorney client privilege and work product treatment, segregation of one client matter from another, whether client data trains any model, and what the vendor commits to in writing rather than in marketing.
The commitments are in a published agreement and they are specific. Confidentiality is mutual, covers information whether or not marked confidential, restricts disclosure to employees with a need to know, and survives five years after disclosure with trade secrets protected for as long as the law protects them. On termination the receiving party must return or destroy every copy and certify the destruction in writing. Compelled disclosure requires written notice to the other party first and a reasonable effort to obtain a protective order. The customer owns all right, title and interest in its data, and the vendor's carve-out extends only to data derived from monitoring use of the service. Around that sit tenant isolation, row-level security for the most sensitive data, least-privilege access, background checks and audited access. The gaps are a missing subprocessor list, no data processing addendum referenced, and no treatment of the disclosure risk that unfiled invention material carries. Verified 20 September 2026.
General enterprise confidentiality is evidenced and the legal professional dimension is absent, and the buyer profile explains why without excusing it. Published: strict data isolation, ISO certified security policies, encryption in transit and at rest, consent management, GDPR and CCPA compliance, and a positioning as a privacy first partner for enterprises. Those are real controls covering the confidential research and portfolio material customers upload. What does not appear anywhere: attorney client privilege, work product, or any acknowledgment that a law firm user's material carries professional obligations distinct from a corporate research team's. The distinction from category peers is instructive rather than incidental. Patlytics and Solve Intelligence sell primarily to patent practitioners and both reached A here by engaging the professional rules; PatSnap sells across research and development, strategy, life sciences and legal, and addresses confidentiality as an enterprise data question throughout. That is coherent for its market and it leaves the practitioner question unanswered. Checked the trust center material, the home page, the Eureka pages and the security statements on 29 Aug 2026.
UPL and Professional Responsibility Posture
Whether the vendor is clear that it supplies a tool rather than legal advice, who its audience is, and how it addresses unauthorized practice of law, competence and supervision duties, and jurisdiction limits. ABA Formal Opinion 512 is the reference point. Where the advice line is not the duty a product raises, the axis is read through the nearest professional duty it does raise: judicial conduct rules and the reviewing duty for products sold only to courts, and the duty to bill for time actually spent for products that draft time entries.
A warranty disclaimer stands in for a position on advice, and the audience is left genuinely unsettled. The agreement disclaims all warranties and states specifically that no warranty is given that results will be accurate, complete or error free. Nothing says the output is not legal advice, nothing addresses an attorney-client relationship, and nothing addresses who should review a freedom-to-operate assessment before a company acts on it. That matters because the output is legal in character: per-patent freedom-to-operate risk ratings, novelty and prior art assessment, and white space identified for new filings. On audience the estate says two different things. Product surfaces address IP teams, a patent committee and IP counsel directly. The vendor's own comparison writing says the platform is built for researchers rather than IP attorneys and points readers to other tools for prosecution and IP legal work. Both are the vendor's own words and both are recorded. Verified 20 September 2026.
Not located. The platform generates freedom to operate assessments, novelty and design patent risk analyses, and drafts patent claims including Markush structures, all of which are outputs a practitioner signs off or a business relies on, and the customer accounts published describe non lawyers in research and development using them for risk management decisions. Nothing published addresses whether output constitutes legal advice, what the role of qualified counsel is where an engineer or a chemist runs a freedom to operate agent directly, or any professional conduct framework. The question is arguably more live for this vendor than for its category peers precisely because its named buyers include scientists and engineers rather than only attorneys. Checked the home page, the Eureka pages, the life sciences landing page, the trust center material and the site navigation on 29 Aug 2026.
AI Governance and Bias Disclosure
Published governance over model behavior: who owns it inside the vendor, what is tested before release, and what is disclosed about disparate output across matter types, parties, or populations.
A thoughtful published position on what AI can and cannot do here, with no mechanism behind it for the vendor's own system. The estate carries a substantial argument that general-purpose models cannot do patent intelligence, resting on training data limits, the closure of the web to crawlers and the absence of ontological frameworks for claim scope, prosecution history and assignee normalization. It commits to being model-agnostic and to updating its ontology continuously. What it does not publish is governance of itself: nobody is named as accountable for model behavior, nothing describes what is tested before an agent or ontology change ships, and no disclosure addresses whether coverage or ranking differs across technology domains, languages or patent offices. On a platform whose value rests on completeness, the absence of any published evaluation of its own coverage gaps is the gap. Verified 20 September 2026.
Data governance is described in operational detail and model governance is not. The vendor publishes that a four stage pipeline processes over 2 billion data points across 20 specialized domains with rigorous governance at every stage, stating this is what reduces hallucinations and improves output reliability, and names its retrieval architecture. PatentBench, graded on the Citation Accuracy axis, is a published evaluation instrument and its existence is evidence that measurement happens. What is absent is governance of the models themselves: no AI policy, no model card, no bias or fairness testing, no accuracy monitoring beyond the benchmark, no drift statement, no named governance body, no ISO 42001 and no EU AI Act positioning. The distinction being drawn is between governing the corpus and governing the system, and this vendor documents the first thoroughly and the second not at all. Compare Patlytics at B with an ISO 42001 certificate, and Solve Intelligence at B with a published continuous evaluation process and an EU AI Act self classification.
AI Safety and Data Stewardship
Retention, deletion, access control, and what happens to prompts and documents after they are processed. Whether the vendor states its subprocessors and its incident practice, or leaves the buyer to assume.
Most of the ground is covered and the retention question is not. Published: encryption in transit and at rest from the interface to the storage layer, HTTPS with modern TLS, hosting on major cloud infrastructure with network and data segmentation, firewalls and automated threat detection, least-privilege production access that is regularly reviewed, row-level security for the most sensitive data, multi-factor authentication and single sign-on, employee background checks, annual penetration testing by independent firms and continuous vulnerability scanning. Data is stored within United States borders with no offshore routing. Customer data is never used to train models and is not shared with model providers for training. What is absent: no retention period is published for prompts, outputs or uploaded documents, no subprocessor list was located, and the agreement carries no breach notification commitment, so what happens after an incident is unaddressed. Verified 20 September 2026.
Isolation and compliance are stated and the training question is unanswered. Published: strict data isolation, ISO certified security policies, encryption in transit and at rest with named protocol and cipher, GDPR and CCPA compliance with stated consent management, and a self description as privacy first. Consent management is a genuine element and few records on this index name it. What was not located is any statement on whether customer uploaded documents, search histories, proprietary formulations or research materials are used to train or improve the domain models, any retention period, or any deletion right. The gap is material for this product because customers upload their own novel compounds, formulations and unpublished technical material into an analysis platform, which is competitively sensitive in a way patent search over public data is not. Compare Solve Intelligence at A and DeepIP at B in this category, both of which state a retention or training position plainly. Checked the trust center material, the home page, the Eureka pages and the security statements on 29 Aug 2026.
AI Liability and Recourse
What the vendor stands behind contractually when its output is wrong. Indemnities, caps, carve outs, insurance, and whether any of it is published or only reachable through a negotiated agreement.
A real published position, with the exposure this product creates sitting outside it. The agreement gives the customer an indemnity against third-party claims that the service infringes or misappropriates US patents, copyrights or trade secrets, with stated conditions, two carve-outs and the vendor's right to modify, replace or terminate if it cannot cure. The customer indemnifies the vendor for negligence, unauthorised use and unapproved combinations. Liability is capped at fees paid in the twelve months before the claim, applied to both the infringement indemnity and the agreement as a whole, with consequential, incidental and punitive damages excluded, along with loss of data and breach of data or system security. What is not covered is a wrong answer: the warranty section disclaims any assurance of accuracy or completeness, so a freedom-to-operate assessment that misses a blocking patent carries no recourse. No insurance or service level is published. Verified 20 September 2026.
No published position located. Nothing was found on liability for AI output, warranty, service levels or remedy. The exposure here is unusual on this index because the consequential outputs are business decisions rather than filings: a published customer account describes the FTO Agent transforming risk management by completing infringement analysis in hours, and another describes lead compound screening acting as the first gate in a research and development process, so a false clearance sends a product to market and a false negative on novelty kills a viable molecule. The vendor publishes a 77 percent hit rate on design freedom to operate, which by its own arithmetic means roughly one in four is missed, and no published position addresses who carries that. Checked the home page, the Eureka pages, the life sciences landing page, the trust center material and the site navigation on 29 Aug 2026.
Practice Systems Integration Depth
How deeply the product reaches into the systems legal work already lives in: document management such as iManage and NetDocuments, Word and Outlook, contract lifecycle management, matter management, e-billing, and court filing systems.
Named connections with described mechanisms, into the tools this buyer actually uses. The agreement treats API and Model Context Protocol access as part of the service rather than an add-on. Integrated Intelligence connects a customer's own Copilot, Claude or ChatGPT to the patent, scientific, chemical, regulatory and market data, with prebuilt stage-gate agents or custom agents built around the customer's workflows, and the vendor describes enterprise API partnerships with OpenAI, Anthropic and Google. Journal access runs through a named third party with published terms: the Cypris Journal Connection links a customer's existing journal subscriptions through OpenAthens single sign-on, with the OpenAthens flow-down terms incorporated by reference. Single sign-on is supported with leading identity providers. What is not yet there is stated as such: integrations into electronic lab notebooks and enterprise tools are described as expanding, and no developer documentation was located. Verified 20 September 2026.
THE MOST COMPLETE INTEGRATION DISCLOSURE ON THE INDEX. The Patsnap Open Platform is a dedicated developer product, not a connector list, offering four distinct integration surfaces: REST APIs, MCP servers, UI widgets and agent skills, with a published Developer Center and documentation. Coverage is enumerated rather than claimed: more than 20 patent APIs, more than 25 corporate APIs and more than 16 infringement APIs. The MCP server offering is the notable element, since it exposes the corpus and its agents as tools to external AI systems, and the vendor names the frameworks it targets including LangChain, AutoGen and function calling workflows, which is integration aimed at how AI systems are actually assembled in 2026. Agent skills as a published surface goes further still. A buyer can obtain an API key from the site. Graded A because a developer can assess, obtain and build against this without contacting sales, which is the top of this axis, and because the surfaces are enumerated with counts rather than asserted. Held short of perfection because no named IP management or docketing system integration was located, which is the one thing a prosecution practice would ask for and which DeepIP states.
Deployment Model and Data Residency
Where the software runs and where the data sits. Multi tenant cloud, single tenant, private deployment, on premises, and whether region of residence is a published option or an enterprise conversation.
Both halves are answered plainly, and neither is offered as a choice. Residency: the platform is built and operated in the United States, with US-based infrastructure across datacentres, hosting and operations, and the estate states there is no offshore data routing with no exceptions. All data is stored within US borders. Tenancy: tenant isolation is described as by design, with customer data logically separated between tenants and row-level security applied to the most sensitive data, hosted on major cloud infrastructure with network and data segmentation. The company describes itself as US-headquartered with US-based investors and leadership, and the agreement carries export control obligations and the federal commercial-item clauses that government buyers need. What is not offered is optionality: no single-tenant, private or on-premise deployment is published, no cloud provider is named, and a buyer outside the United States has no region to choose. Verified 20 September 2026.
Nothing located. No hosting provider is named, no region or data residency commitment is published, and no single tenant, dedicated or on premise option is described. The absence is more consequential for this vendor than for most on the index: the company is headquartered in Singapore with named investors including SoftBank and Tencent and serves customers including NASA, MIT and General Electric, so the jurisdiction in which research material and unpublished technical data is processed is a question those customers' own compliance functions would ask directly, and it is unanswered in public material. Stated GDPR and CCPA compliance establishes that regimes are addressed and says nothing about where data sits. Compare DeepIP at A for on premise deployment and Solve Intelligence at A for customer selectable jurisdiction. Checked the trust center material, the home page, the Eureka pages, the Open Platform material and the site navigation on 29 Aug 2026.
Security Certifications and Trust Center
Independent attestation a buyer can pull without a sales call: SOC 2, ISO 27001, penetration test summaries, a trust center with current reports and named scope rather than a badge image.
The standard is named, the portal is real and populated, and the evidence behind the certificate is still not published. The vendor states it maintains SOC 2 Type II compliance, independently audited each year, with continuous controls monitored across availability, security and confidentiality, and it points buyers to a trust center at a stable address where evidence of compliance is available. That portal, read on 20 September 2026, shows SOC 2 marked compliant and lists more than fifty named controls across product, data, network, application, endpoint and corporate security, together with more than thirty-eight named policy and procedure documents including access control, asset management, acceptable usage, endpoint security and data protection, behind a request-access button. Annual penetration testing by independent firms is stated separately. What is absent is what would lift this higher: no audit period, no auditor named, no subprocessor list, and no way to confirm from outside that a request is fulfilled. Verified 20 September 2026.
A published trust center with named standards and annual audits, held below the top by the type of the SOC report. Published on a dedicated trust center: certification by the AICPA, ISO 27001:2022 with a statement that systems undergo rigorous annual external audits, encryption using TLS 1.3 and AES-256 in transit and at rest, GDPR and CCPA compliance, and consent management. Naming the ISO revision year and committing to annual external audit addresses currency, which most records on this index omit. THE LIMITING FACT: the SOC attestation is announced and described as SOC 2 Type 1. Type 1 assesses the design of controls at a single point in time; Type 2 tests whether those controls operated effectively across a period. Every other record on this index that reaches A on this axis holds a Type 2 or an equivalent, and the difference is not cosmetic. Held at B on that basis, alongside the absence of a named auditing firm and any certificate date for the ISO. Calibration within this category: Patlytics reaches A on three certifications with named penetration testing partners, and this record publishes a trust center with a weaker attestation type.
Model Supply Chain Disclosure
Which models sit underneath, whose they are, where they run, and whether the vendor commits to telling customers when that changes. A legal buyer inherits every dependency it cannot see.
Three providers are named and the commercial mechanics are unusually visible. The estate states that agents run on models from OpenAI, Anthropic and Google, describes the platform as model-agnostic and multi-model against competitors locked to one provider, and refers to enterprise API partnerships with all three. The agreement makes the arrangement concrete: AI credits are allocated annually by subscription value and each model consumes them at a rate reflecting its processing cost, so a buyer can see that model choice is a priced variable. A data commitment accompanies it, that customer material is not shared with third-party model providers for training. What is not published is anything version-level: no model identifiers, nothing on which model serves which agent, no statement of where inference happens, and the opposite of a change commitment, since the vendor reserves the right to adjust available models, credit allotments and per-prompt costs at any time. Verified 20 September 2026.
Nothing located. No foundation model provider, model family or version is named, and no subprocessor list was found. The vendor describes its architecture in technical terms, naming retrieval augmented generation and what it calls retrieval augmented thinking, and describes domain specific and domain trained models, all of which characterize method rather than provenance. Independent product review notes an evolving AI backbone, which implies underlying models change over time and makes identification more rather than less relevant for a buyer assessing consistency. For a platform whose Open Platform exposes agents to external systems through MCP servers, the identity of the models behind those agents is a question a developer integrating them would ask. Checked the home page, the Eureka pages, the Open Platform material, the trust center material and independent review material on 29 Aug 2026.
Commercial Transparency
Whether a buyer can learn what this costs without entering a sales process: published rates, the unit being charged, what sits behind an enterprise tier, and what implementation adds.
No price is published anywhere, and the structure behind it is documented better than most. There is no pricing page in the navigation or the footer and the only route is a demo request. What the published agreement does give a buyer is the shape of the deal: fees are set in an order form, seats are counted as named authorized users with a cap stated in that form, the initial term runs at least a year and renews automatically unless either side gives forty-five days notice, and late payment carries 1.5 percent monthly interest. Three consumption units are defined. AI credits are allocated annually by subscription cost and consumed per model session, expiring at the end of the year. Research paper credits are valued one to one with US dollars for buying closed-access papers. Research entitlements are capped at one report and two briefs a month. None of it is attached to a number. Verified 20 September 2026.
The pricing structure is disclosed and no figure is. Published through an independent analyst platform and consistent with vendor material: a subscription model with multiple tiers varying by features, user numbers and access levels, with details provided on request based on customer requirements, and a named Enterprise plan that includes specified agents such as the Antibody Target Predictor and Markush Claim Drafting. Knowing that the meter runs on seats and feature tier, and knowing which capabilities sit behind the top tier, tells a buyer how the bill will grow and which plan they would need, which is more structural disclosure than most of this index provides. Two self serve entry points exist and are credited as reducing the barrier: a free trial of Eureka and an API key obtainable from the Open Platform. Held at C because no price, range or entry point figure is published at any level and every commercial route ends in a quote request. Source basis recorded as Third Party Estimated because the tier structure detail comes from an independent analyst listing rather than a vendor pricing page, which was not located.
Firm and Practice Coverage
Who the product is actually built for. AmLaw, midlaw, small firm and solo, in house departments, government and courts, and which practice areas are supported rather than merely claimed.
Three industries are described with the work each does and a named customer beside it, and no boundary is drawn. Materials and chemicals teams get compound-level intelligence linking patent filings, scientific literature and regulatory data, with prior art search, freedom-to-operate analysis, competitive monitoring and chemical structure mapping, and Oerlikon is named. Industrial and manufacturing teams get technology landscaping, patent analytics, supplier scouting and literature review, with Yamaha named. Life sciences and pharma teams get deep patent family analysis, citation tracking, clinical and scientific literature review and pipeline monitoring, with Johnson & Johnson named. Buyer types run from Fortune 100 companies to government agencies and national laboratories, and the agreement carries the federal commercial-item clauses that public buyers require. What is absent is any statement of where the platform does not reach, whether by right, jurisdiction or workflow. Verified 20 September 2026.
The largest and most specifically characterized corpus in the category, stated on the dimensions a researcher checks. Scale: more than 2 billion structured data points and over 210 million global patent, science and technology records, spanning patents, scientific literature, litigation records, clinical data and technical knowledge across 20 specialized domains. Breadth beyond patents is the distinguishing element, since prior art is not confined to patents and this is the only record in the category to state coverage of scientific literature and clinical data alongside patent documents. Industry coverage is enumerated across agriculture and chemicals, consumer goods, food and drink, life sciences, automotive, oil and gas, professional services, aviation and aerospace and education. Domain depth is demonstrated rather than claimed in life sciences, with named capability for sequences, Markush structures, structure activity relationship extraction, clinical trial comparison and druggability. Patent translation trained on global patent data addresses cross jurisdiction work directly. Graded A because coverage is enumerated with counts, domains and named artifact types a buyer can test against their own work. Held short of perfection because no update lag or refresh frequency is published and no patent office list is given.
The 12 legal signals, side by side
Recorded rather than graded. These are the questions a practitioner has to answer before a tool touches a client matter, and the answers are taken from public material only.
Client Data in Training
Can material a lawyer puts into this product be used to train a model?
The position is absolute and it sits in a policy page rather than the contract. The security page answers the question directly: asked whether Cypris uses customer data to train AI models, the answer is no, and the commitment extends to searches, interactions and data, covering both training its own models and sharing with third-party model providers for training. The published Terms of Service, read in full, contain no training clause at all, in either direction.
What the agreement does give is ownership, stating that the customer owns all right, title and interest in its data, with the vendor's carve-out limited to information derived from monitoring use of the service. So a buyer gets a clear answer and no contractual remedy if it is departed from, and the agreement can be amended by the vendor posting revised terms. Moving the commitment into the order form is the obvious ask.
Silent. The quoted phrase is the closest the vendor comes to addressing customer content in the AI context and it describes isolation and compliance rather than use. No statement in either direction was located on whether customer uploaded documents, proprietary formulations, search histories or research materials are used to train or improve the domain specific models. The gap is material because of what this platform ingests: customers upload novel compounds, unpublished technical material and internal documents for analysis, and the vendor separately markets domain trained models across 20 specialized domains, so the question of whether one feeds the other is directly raised by the product design.
Recorded as silent, not as a negative commitment. Correction candidate: the trust center was identified but not entered in this pass. Checked the Eureka pages, the home page, the life sciences landing page, the trust center summary and the Open Platform material on 29 Aug 2026.
Prompt and Output Retention
How long does the product keep what a lawyer typed, and can that be set to zero?
How long anything is kept is not addressed. Nothing on the estate states a retention period for prompts, agent outputs, saved searches, uploaded documents or the internal project material the Knowledge Management product is built to hold, and nothing describes deletion on request during the term. The agreement touches the question only at the edges: on termination each party must return or destroy the other's confidential information and certify it in writing, which covers material exchanged between the parties rather than the working data a team accumulates in the platform over years.
The retention commitment a buyer might expect to find beside the no-training promise is not there. That gap is sharper here than on most records because the product's stated purpose is accumulation, with intelligence that compounds across projects and institutional knowledge captured so it survives departing employees.
Not addressed. No retention period is published for uploaded documents, agent queries, generated reports or analysis outputs, and nothing indicates whether retention is configurable or whether a customer can require deletion. Consent management is named as a capability, which addresses the collection of personal data under GDPR and CCPA rather than the duration for which uploaded research material is held. Compare DeepIP at disclosed fixed with a stated zero retention policy, Solve Intelligence at customer configurable zero, and Patlytics at disclosed fixed with a 90 day window, all within this category. Checked the trust center summary, the Eureka pages, the home page and the Open Platform material on 29 Aug 2026.
Ethical Walls and Matter Segregation
Does retrieval respect the firm’s ethical walls, or can the model read across them?
Separation between customers is described as an architectural property and backed by named controls. The estate states tenant isolation by design, with customer data logically separated between tenants and row-level security enabled for the most sensitive data. Access to production systems follows least-privilege principles and is reviewed regularly, multi-factor authentication is enforced for all employees and administrators, employees are background-checked, and access privileges are audited.
Single sign-on is supported with leading identity providers so a customer can apply its own controls. The trust center lists named controls for production system user review, production database access restriction and identity validation. What is not described is separation inside a customer: nothing published sets out how one project, division or research program is walled from another, which matters where a company runs competing programs or holds material from a partner.
Claimed and not documented. The vendor states strict data isolation alongside ISO certified security policies and consent management, so a separation model is asserted. Nothing documents it: no statement of whether isolation operates between customers, between users within an organization, or between projects, and no description of how the boundary applies to the agent layer when an agent runs analysis across a corpus.
The scenario this product raises is different from a law firm conflicts question and is equally live: the platform is sold across research and development, strategy and legal within the same enterprise, and to competing enterprises in the same technical field, and nothing published describes what prevents one customer's uploaded formulation or unpublished compound from informing anything surfaced to another. No integration with a customer permission system was located that would let the platform inherit an existing model.
Third Party Request and Subpoena Notice
If someone subpoenas the vendor for a firm’s data, does the firm hear about it first?
The commitment is contractual, mutual, and goes beyond notice to resistance. The confidentiality section permits disclosure to the limited extent required to comply with a court order or other governmental body, or as otherwise necessary to comply with applicable law, but only where the disclosing party has first given written notice to the other party and made a reasonable effort to obtain a protective order. Both limbs matter: the customer learns of the demand before anything is produced, and the vendor must itself try to narrow or block it rather than leaving that to the customer.
Because the clause runs both ways it binds the vendor as recipient of the customer's confidential information. What is not published is operational detail, a named contact, a stated timeframe, or any transparency reporting on how often demands arrive.
Not addressed. No government or law enforcement request clause, no commitment to notify a customer before producing their data, and no transparency report were located. The question carries particular weight for this vendor and the note records why plainly rather than implying anything: the company is headquartered in Singapore with named investors including SoftBank Vision Fund 2 and Tencent, and its named customers include NASA, MIT and General Electric, so the jurisdictions in which a request could be served and the parties who could receive one are a live diligence question for customers uploading unpublished technical material.
Nothing published addresses it in any direction. Checked the trust center summary, the home page, the Eureka pages and the site navigation on 29 Aug 2026.
Primary Law Corpus Provenance
Where does the law in this product come from, and does the vendor have the right to use it?
The corpus is itemized with unusual precision and its licensing basis is mostly unstated. Published counts: 198 million patent documents from more than 150 countries, 341 million research papers from 274,000 journals and repositories, 125 million indexed chemical compounds with structure, synonym, trade name and vendor resolution, more than a million tracked news sources across 200-plus countries, and over a billion live commercial, market and regulatory sources.
Data is described as real-time and taken direct from global patent offices, publishers and trusted sources. Where licensing is addressed it is addressed properly, at the edges: research paper credits let a customer buy closed-access papers, with the vendor stating it takes measures to comply with publisher copyright restrictions and passing all publisher rights and restrictions through to the customer at the point of purchase, and journal access runs through OpenAthens under incorporated flow-down terms. For the indexed corpus itself no license basis is stated.
Named with the most specificity in the category, and no license basis stated. The corpus is enumerated rather than gestured at: more than 2 billion structured data points and over 210 million global patent, science and technology records, spanning patents, scientific literature, litigation records, clinical data and technical knowledge across 20 specialized domains, with a four stage processing pipeline described as applying governance at every stage.
Coverage of scientific literature and clinical data alongside patents is the distinguishing element and no category peer states it. What is absent is the basis on which any of it is held. Patent documents are published by offices as a condition of grant, but scientific literature is substantially copyrighted and clinical trial data is licensed or regulated, and nothing states what rights support their inclusion, from which publishers or registries, under what terms.
No update lag, refresh frequency or historical date range is published either, despite the pipeline governance claim. The larger and more heterogeneous the corpus, the more the licensing question matters, and this is the largest corpus in the category.
Good Law Verification
Does the product tell you when the authority it just cited has been overruled?
Read through the patent analog, a currency mechanism exists and stops short of a validity opinion. The nearest equivalent to whether an authority is still good is whether a patent is still in force, and the estate addresses it. The source named is the registers themselves: data is described as real-time and taken direct from global patent offices, across more than 150 countries, with no commercial citator or third-party status product anywhere in the path.
The status is live rather than periodic, and outputs carry current legal status alongside assignee and filing date; continuous monitoring tracks changes across filings. The vendor's published critique of general-purpose models turns on exactly this, that they cannot verify whether a patent has been assigned, abandoned or subjected to terminal disclaimer since their training data was collected. What that gives is status currency rather than subsequent treatment: nothing published tells a user that a claim has been narrowed in prosecution, challenged at the board or limited by a court, and the freedom-to-operate risk ratings are the platform's own assessment rather than a record of what has happened to a claim.
Not addressed, and applicable in the patent specific form. The platform performs freedom to operate analysis, design patent risk assessment, infringement risk identification and novelty assessment, all of which turn on the legal status of the references surfaced: whether a patent is granted and in force, lapsed for unpaid fees, expired, amended in reexamination, or invalidated in post grant proceedings. A freedom to operate clearance that treats a lapsed patent as a live blocking right, or misses one that has been revived, is wrong in a way no amount of retrieval quality fixes.
Nothing published states whether legal status is tracked, surfaced or incorporated into the risk scoring, and litigation records are named as part of the corpus without any statement that outcomes are linked back to the patents they concern. Fourth consecutive record in this category to leave this signal unanswered, and the pattern is now firm across every ip-and-patents vendor built.
Refusal and Uncertainty Behavior
What does the product do when the answer is not in the corpus?
The vendor diagnoses this failure in other systems and publishes nothing about its own. Its comparative study identifies the central risk of general-purpose models as producing incomplete outputs with high confidence, noting that at no point did any model indicate the boundaries of its knowledge or flag that its results represented a fraction of the record, and that the better the format the less likely a user is to question completeness.
Nothing published says what Cypris does differently. There is no confidence score on a result, no stated threshold below which an agent declines to answer, and nothing describing what a user sees when coverage in a technology area is thin. The freedom-to-operate risk ratings are classifications of the patents found, not statements of confidence in the search behind them. Checked the home page, the security page, the about page, the product summaries and the Terms of Service on 20 September 2026.
Not addressed as a behavior, with an aggregate accuracy figure published instead. The vendor publishes a 77 percent hit rate on design freedom to operate through PatentBench, which is an honest headline number and is graded on the Citation Accuracy axis. It describes performance in aggregate and says nothing about the individual result in front of a user: no confidence indication accompanying an agent output, no flag on a low certainty match, no statement of whether an FTO Agent will report that it could not reach a conclusion rather than returning one, and no description of what a user sees when the corpus does not support an answer.
The distinction matters most on precisely the number published, since a stated 77 percent hit rate means roughly one in four is missed and nothing tells the user which. Compare Patlytics at documented in this same category, on color coded confidence indicators surfaced at the point of reading.
Fabricated Citation Record
Does a public court record exist addressing fabricated or hallucinated legal citations in output from this product?
No record was located of this product's output being found fabricated or inaccurate in a proceeding, a regulatory action or a published account. Searches on 20 September 2026 across the vendor's estate, press and directory profiles returned nothing of the kind. The relevant failure here would be a missed blocking patent rather than an invented one, since results are anchored to filings a user can open in USPTO, Espacenet or WIPO PATENTSCOPE.
The vendor's own published study records one shortfall in its output, an incomplete separation between two related bodies of prior work in a test landscape, which is disclosed rather than found against it.
None located, with the instrument named. General web searches combining the vendor and product names with court, order, sanction, fabricated citation and patent terms returned nothing on 29 Aug 2026, and no named docket database, patent office record system or court record tracker was searched. Recorded as a statement about what this search found, not as a clearance. The exposure shape differs from the drafting products in this category: this platform generates analyses and reports supporting business decisions rather than documents filed at a patent office, so the analogous adverse finding would be a proceeding addressing reliance on a defective freedom to operate clearance, which would surface as commercial litigation rather than in a prosecution file wrapper.
Bar Guidance Alignment
Has the vendor engaged in public with the ethics opinions its buyers are bound by?
Nothing on the estate engages professional or ethical obligation. No rule of professional conduct, ethics opinion or bar guidance is named, and the duty is not engaged in general terms either: the Terms of Service carry no statement that output is not legal advice and create no position on an attorney-client relationship, offering only a disclaimer of warranties. The omission is worth noting on this record because of what the product produces and who receives it.
Freedom-to-operate risk ratings, novelty assessment and white space for new filings are legal determinations in substance, the estate names IP counsel and a patent committee as recipients of that work on one surface and describes the platform as built for researchers rather than IP attorneys on another, and the buyer is frequently an engineer. Checked the Terms of Service, the home page, the security page, the about page and the footer on 20 September 2026.
Not addressed. No named ethics opinion, no USPTO Rules of Professional Conduct reference, no 37 CFR citation and no bar guidance of any jurisdiction was located. The absence is coherent with the vendor's market position rather than an oversight: PatSnap sells across research and development, strategy, life sciences and legal, and the majority of its named customers are corporations rather than law firms, so its material addresses enterprise buyers rather than practitioners.
That explains the silence without curing it, since the platform still generates freedom to operate assessments and drafts claims, and legal is a named buyer. Third of four records in this category at this value, with Patlytics the only exception. Checked the home page, the Eureka pages, the life sciences landing page and the site navigation on 29 Aug 2026.
Billing and Fee Posture
Does the vendor address what happens to the bill when the work takes an hour instead of six?
No lawyer's fee sits in this product's path. The buyer is a corporate R&D or IP department licensing a subscription for its own internal use, with the agreement limiting access to authorized users who are the customer's own employees, contractors and agents. Nothing is billed on to a client, no outside counsel relationship is mediated, and the vendor's own commercial framing is about research throughput rather than legal spend.
What the agreement does publish about money runs to its own charges: fees set in an order form, annual terms renewing automatically, AI credits allocated by subscription value and consumed per model session, research paper credits valued one to one with dollars, and capped entitlements of one analyst report and two briefs a month. None of that bears on disclosing machine-assisted work to a client or on how such work is billed.
Savings claims only, and the published figures describe enterprise outcomes rather than firm labor, which is a category first. Standard efficiency claims: productivity boosted by up to 60 percent and workflows completed 90 percent faster. The distinctive one is a customer quote stating the platform has literally saved hundreds of thousands of dollars in the context of evaluating partnerships and acquisitions, which is a claim about avoided transaction cost rather than about hours, and a second describing infringement analysis moving from weeks to hours.
None carries methodology, baseline or period, and the six figure saving is not attributed to a named organization. Nothing appears on the client's side of the equation: no position on how AI assisted analysis should be recorded where a firm bills a client for freedom to operate work, and no exportable record showing what portion of an assessment was machine generated.
Outside Counsel Guideline Readiness
Can a firm get this vendor through a client’s AI clause without a bespoke negotiation?
A diligence reviewer gets further here than on most records in this lane, and stops at the door. The trust center publishes the certification, more than fifty named controls across product, data, network, application, endpoint and corporate security, and more than thirty-eight named policy and procedure documents including access control, asset management, acceptable usage, endpoint security and data protection, all obtainable through a request-access route rather than a sales conversation.
The security page answers the standard questionnaire directly on training, encryption, access, penetration testing, tenant separation and single sign-on, and the agreement carries confidentiality, compelled-disclosure notice, US export compliance and the federal commercial-item clauses a government buyer needs. What is not available is the evidence itself without asking: no subprocessor list, no audit period or auditor, no data processing addendum referenced, and no breach notification commitment anywhere.
On request, through a published trust center with named standards behind it. The trust center is a defined destination and the disclosures a firm could cite are specific: AICPA certification, SOC 2 Type 1, ISO 27001:2022 with stated annual external audits, TLS 1.3 and AES-256 encryption, GDPR and CCPA compliance, strict data isolation and consent management. Naming the ISO revision year and the audit cadence is more than most records manage.
Held at on request rather than higher because nothing is downloadable and the pack is incomplete: no subprocessor list, no named model provider, no data processing agreement, no auditor identified, no certificate dates, and no route to the SOC report itself was located, with the trust center gate not entered in this pass. The SOC being Type 1 rather than Type 2 also limits what a firm can represent onward to a client, since Type 1 speaks to control design at a point in time rather than operating effectiveness over a period.
Court Disclosure Support
If a judge’s standing order requires an AI disclosure, can the product produce one?
Nothing published addresses disclosing the machine's involvement to anyone outside the company. No court or tribunal sits in this product's path, and the nearest forum question, whether AI-assisted prior art or freedom-to-operate work needs to be disclosed to a patent office or relied on in litigation, is not raised anywhere on the estate. What exists is evidentiary rather than procedural and is real as far as it goes: outputs are anchored to patent numbers, assignees and publication dates that a practitioner can verify independently in USPTO, Espacenet or WIPO PATENTSCOPE, and the vendor preserved and published the outputs of its own comparative study.
But nothing describes an exportable record of how a conclusion was reached, nothing distinguishes an agent's determination from an analyst's in the Research Briefs product, and no format or guidance is offered for any proceeding.
Partial record, with a stronger source limb than most and the same two gaps. Source linked evidence is named as a property of the PatentBench methodology and the platform is built on a governed corpus with a four stage pipeline, so an assertion in a generated report is traceable to a document in a way a reader can follow, and the Open Platform exposes the underlying data through APIs so a claim can be independently re-queried against the same corpus.
That last property is unusual and is a genuine disclosure strength: a party can verify the underlying data without the vendor's cooperation. The familiar limbs are absent. Nothing indicates that output records which model or agent produced it, which matters more here because independent review describes an evolving AI backbone, so the system behind a report generated last year may not be the one running today. And no human verification record is captured, so a party relying on a freedom to operate assessment in a dispute cannot evidence that a qualified person reviewed it.
The questions both sides leave open
Derived from the records above rather than written, so it cannot favor either vendor. Take these into both conversations and ask each side the same question.
- Prompt and Output Retention
- Refusal and Uncertainty Behavior
- Bar Guidance Alignment
Which one fits
Choose Cypris if
- You must tell leadership whose AI reads your research. Cypris names OpenAI, Anthropic and Google as its model providers, meters them as AI credits whose rate reflects each model's cost, and states that customer data is never used to train models or shared with providers for training.
- Your data must stay in the United States. Cypris states that all data is stored within US borders with no offshore routing, isolates tenants with row level security for sensitive data, and carries the federal commercial item clauses government buyers need.
- You want a person to write the high stakes brief. Cypris sells Research Briefs written by its own analysts alongside its agents, with a set number of reports and briefs each month, and connects its data to your own Copilot, Claude or ChatGPT through an MCP server.
Choose PatSnap if
- You want a published accuracy test before trusting an FTO agent. PatSnap publishes PatentBench, a benchmark for real patent tasks with a stated methodology, reporting a 77 percent hit rate on design freedom to operate.
- Your developers will build on the data. PatSnap's Open Platform offers REST APIs, MCP servers, UI widgets and agent skills, with more than 20 patent APIs, 25 corporate APIs and 16 infringement APIs, and an API key obtainable from the site.
- You work in life sciences. PatSnap's Eureka Life Sciences suite covers lead compound analysis, structure activity relationship extraction, clinical trial comparison, antibody target prediction and Markush claim drafting, over a corpus that includes scientific literature and clinical data.
In summary
Cypris
Cypris, sold by IP Web, Inc. of New York, is an R&D and IP intelligence platform unifying 198 million patent documents, 341 million research papers, 125 million chemical compounds and market and regulatory sources, with AI agents for technology scouting, literature review, prior art, freedom to operate and white space analysis, and analyst written Research Briefs. The AI Legal Index grades it in the top two bands on eleven of fifteen capability axes, with an A on AI centrality. It names OpenAI, Anthropic and Google as model providers, stores data in the United States, and states SOC 2 Type II. It names Johnson & Johnson, Yamaha and NASA among customers. As of 20 September 2026 the index located no published price or retention period.
PatSnap
PatSnap, founded in 2007 and headquartered in Singapore, is an innovation intelligence and patent analytics platform for IP, R&D and innovation teams, over more than 210 million patent, science and technology records. Its Eureka layer provides agents for prior art, novelty and freedom to operate, drafting and translation, with a life sciences suite for biopharma, and its Open Platform exposes APIs and MCP servers. The AI Legal Index grades it in the top two bands on six of fifteen capability axes, with A grades on citation accuracy, operational evidence, integration depth and coverage. It publishes the PatentBench benchmark and names NASA, Tesla and MIT among customers. As of 29 August 2026 the index located no training position, hosting region or named model.
Questions buyers ask
Cypris vs PatSnap: which is better for R&D intelligence?
Cypris sits in the top two bands on eleven of fifteen AI Legal Index capability axes and PatSnap on six of fifteen, identical on four, largely because Cypris publishes its model providers, data location and contract terms. PatSnap publishes a benchmark with its methodology, a deeper developer platform and a larger corpus. Teams uploading unpublished research have more to read from Cypris; teams building their own tools, from PatSnap.
What is PatentBench?
PatentBench is PatSnap's own benchmark for patent tasks, which it says measures specific work against structured data, legal context and source linked evidence, with a published methodology. PatSnap reports a 77 percent hit rate on design freedom to operate and claims to be 83 times more accurate than ChatGPT on that measure. The benchmark is not independent, and ChatGPT is a general tool rather than a patent product. Graded by AI Legal Index against 15 capability axes and 12 legal signals, including privilege handling and citation accuracy, from each vendor's own published materials, verified September 27, 2026. No vendor pays for placement.
Which AI models does Cypris use?
Cypris states that its agents run on models from OpenAI, Anthropic and Google, and its agreement meters them through AI credits consumed at a rate reflecting each model's processing cost. It names no versions, does not say which model serves which agent, and reserves the right to change available models. PatSnap names no model or provider. Graded by AI Legal Index against 15 capability axes and 12 legal signals, including privilege handling and citation accuracy, from each vendor's own published materials, verified September 27, 2026. No vendor pays for placement.
Where do Cypris and PatSnap store data?
Cypris states that all data is stored within United States borders with no offshore routing, on major cloud infrastructure with tenant isolation and row level security. PatSnap, headquartered in Singapore, names no hosting provider, region or residency commitment, and states GDPR and CCPA compliance without saying where data sits. Graded by AI Legal Index against 15 capability axes and 12 legal signals, including privilege handling and citation accuracy, from each vendor's own published materials, verified September 27, 2026. No vendor pays for placement.
What do Cypris and PatSnap both leave unpublished?
A price and a retention period. Neither publishes a figure for any tier, and neither states how long uploaded research, prompts or agent outputs are kept. Neither describes an oversight point inside its agents before a result informs a filing or launch, and neither addresses whether its output is legal advice. Graded by AI Legal Index against 15 capability axes and 12 legal signals, including privilege handling and citation accuracy, from each vendor's own published materials, verified September 27, 2026. No vendor pays for placement.
Three readings to weigh. PatSnap's benchmark is its own rather than independent, and its 77 percent hit rate means roughly one design freedom to operate question in four is missed; its methodology page was not read by this index. Cypris's comparative study was commissioned by Cypris, as it discloses. Cypris's commitment not to train on customer data sits on its security page, not in its terms, and neither vendor publishes a retention period. Cypris was verified on 20 September 2026 and PatSnap on 29 August 2026. Neither vendor reviewed this page.
Neither vendor paid for inclusion, placement or a grade, and neither reviewed this page before it published. Everything above comes from public material on the dates shown. How the index grades.