Anthropic's Skills marketplace ships with 200 verified Skills.

The three-tier procurement model

The marketplace ships with three commercial tiers. Tier One is the verified-free layer: 84 of the 200 launch Skills are no-fee installs published by Anthropic itself or by partners who have agreed to a zero-revenue-share placement in exchange for distribution. The free tier is the procurement entry point — a CIO can pilot a verified Skill at zero marginal cost while the enterprise legal team works through the master Skills Agreement. Tier Two is the per-call commercial layer: 89 Skills priced on a metered basis between $0.002 and $0.18 per invocation, with the published pricing routed through Anthropic's existing billing surface and reconciled monthly against the buyer's Bedrock, Vertex, or direct Anthropic contract. Tier Three is the enterprise-license layer: 27 Skills sold on annual contracts ranging from $48,000 to $720,000 per seat-bundle, with Anthropic acting as the licensing agent and the publisher retaining 70 per cent of the contract value. The 70-30 split is more publisher-friendly than the App Store's 70-30 and substantially more publisher-friendly than the OpenAI Apps Store's 75-25 — and that fact is doing visible work inside the announcement.

The launch partners did not arrive at the 70-30 split passively. Stripe's Payments Skill, the Adyen FX-routing Skill, and the Shopify catalog-sync Skill were all conditioned on revenue-share terms that Anthropic conceded in the final 72 hours before the keynote. Patrick Collison, Stripe's chief executive, did not attend the London forum personally but Stripe's API platform lead Will Larson did, and Larson's read on the marketplace mechanics — circulated internally at Stripe and forwarded by at least three sources to procurement counterparts at Block and Shopify — argued that the 30 per cent vendor take is sustainable only if Anthropic absorbs the SOC 2 attestation cost. Anthropic did absorb it. The Skills-specific certification, branded as Skills Verified — Class A, is paid by Anthropic on behalf of the publisher and renewed annually. Class A covers identity, access management, data residency disclosure, prompt-injection resistance, output determinism logging, and a model-version compatibility matrix that ties each Skill to a specific Claude family version. Class B, which the marketplace will introduce in Q3 2026, will require buyer-funded attestation for higher-risk Skills — anything touching protected health information, anything with autonomous spending authority above a threshold, anything that writes to ERP systems of record.

The procurement model rewards the publisher who can demonstrate deterministic output logging. Adyen's FX-routing Skill ships with what the company's product lead Vera Bunschoten described as "audit-grade replay" — every invocation is logged with full input context, full intermediate state, and full output, retained for seven years under Adyen's existing financial services audit framework. That logging discipline is now a default expectation in the Anthropic marketplace. The Tier Two Skills that lack equivalent replay infrastructure were held back from the launch list. Three publishers — two of them named-and-public at the keynote, one which the company declined to identify — withdrew from the launch cohort in the final ten days after failing the determinism logging review. That number, three out of an original 203, is small enough to be incidental and significant enough to indicate that the certification has teeth.

The procurement-side reaction inside Fortune 100 buyers has been faster than Anthropic's go-to-market team modelled. By 21 May 2026, intelligence circulating across procurement networks shows at least 47 CIO-level memos directing legal and risk teams to draft Skills Marketplace participation policies. The memos repeat a common structure: scope of permissible Skills, internal approval thresholds, vendor due diligence overlay, contracting language, and quarterly audit cadence. The procurement teams are treating the marketplace as a new vendor category, not as an extension of the existing Anthropic contract. That treatment matters because it implies a procurement budget line that did not exist 60 days ago, and the existence of that line is the structural change Anthropic has been engineering toward since the Skills primitive shipped in March 2026.

The governance review architecture

The Skills Verified certification is the load-bearing element of the marketplace, and its design reveals what Anthropic believes about the failure modes that will define the next 18 months of agent deployment. Class A certification runs across seven evaluation surfaces. The first is identity and access management: the Skill must declare its required permission scope before invocation and must not request elevated permissions at runtime without explicit user approval. The second is data residency: the Skill must publish a residency manifest that specifies which jurisdictions its computation can run in and which jurisdictions its data persistence layer can touch. The third is prompt-injection resistance: the Skill is tested against a static battery of 1,840 injection vectors derived from a corpus that Anthropic's red team built in partnership with Cure53 and Trail of Bits. The fourth is determinism logging: every invocation must be replayable in a sandboxed environment within a 30-day window. The fifth is model-version compatibility: the Skill must declare the exact Claude family versions on which it has been tested and must fail safely on unsupported versions. The sixth is dependency disclosure: any external API, data source, or third-party model invoked by the Skill must be enumerated in the manifest. The seventh is incident notification: the publisher commits to a 24-hour breach disclosure window via a dedicated Anthropic security inbox.

The combination of SOC 2 Type II and Skills Verified — Class A is the procurement-grade credential. SOC 2 covers the publisher's organisational controls; Class A covers the Skill itself. Anthropic's product team has been explicit, in conversation with at least nine Fortune 500 procurement leads, that the two are non-substitutable. A publisher can hold a SOC 2 Type II report and still fail Class A on the determinism logging surface. A publisher can pass Class A on a single Skill and still lack the organisational control attestation to enter a Fortune 100 procurement pipeline. The marketplace insists on both. That insistence is producing a second-order effect: the smaller publishers — typically Series A and Series B companies who hold a SOC 2 Type I but lack the operational maturity for Type II — are being pushed into a parallel certification investment that several have flagged, in conversations cited by Anthropic's developer relations team, as the limiting factor for their 2026 marketplace participation.

The 200-Skill launch cohort skews toward established enterprise vendors precisely because the certification overhead favours that profile. Block contributes three Skills: a Square Reader interface, a Cash App Business sync Skill, and a tax-reconciliation Skill built on top of the Block Inc. fintech taxonomy. Stripe contributes four: a Payments Skill, a Connect onboarding Skill, an Atlas Agent Skill for AI-native business formation, and a Climate carbon-accounting Skill that ties into the Stripe Climate purchase ledger. Adyen contributes two: the FX-routing Skill and a Treasury Skill for high-volume B2B settlement. Shopify contributes five: catalog sync, inventory monitor, fulfilment hand-off, returns triage, and a B2B Shopify Plus Skill for merchant-of-record arrangements. Sanofi contributes one — a clinical trial protocol summariser built in partnership with Anthropic's life sciences team, which is the only Class A Skill in the launch cohort cleared for a regulated clinical workflow. The Sanofi Skill is the one that procurement teams in life sciences buyers have flagged as the most consequential, because it establishes the precedent that a Skills Marketplace asset can be cleared for use inside a clinical trial sponsor's regulated operations.

The governance review apparatus inside Anthropic is the throughput question. The company has staffed the Skills review team at 38 reviewers as of the launch date, with a stated commitment to scale to 90 by year-end. The reviewer profile is mixed: ex-Big Four audit, ex-FAANG security engineering, ex-AppSec consultants, and four reviewers seconded from Cure53 on a 24-month engagement. The review cycle for a new Class A Skill is currently 11 to 14 business days. By comparison, Apple's App Store review cycle for a new app is one to seven days, OpenAI's Apps Store review cycle for a new GPT app is two to five days, and Salesforce's AppExchange review cycle for a managed package is six to eight weeks. Anthropic's 11-14 day cycle sits between the consumer-grade speed of Apple and the enterprise-grade discipline of AppExchange. The procurement audience reads that placement as deliberate. The developer audience reads it as a bottleneck. Both reads are correct.

Conflict with the OpenAI Apps Store

OpenAI's Apps Store, which shipped in March 2026 with 480 launch apps and a 75-25 revenue split, is the visible adjacent venue. The Apps Store and the Skills Marketplace are competing for the same publisher attention and, at the enterprise tier, for the same buyer budget. The two surfaces differ on three structural dimensions. First, the unit of distribution: OpenAI distributes apps as standalone conversational experiences inside the ChatGPT surface, while Anthropic distributes Skills as agent primitives that compose into a larger workflow. Second, the certification posture: OpenAI's app review focuses on consumer safety and brand-compatibility checks; Anthropic's Class A certification focuses on enterprise auditability. Third, the procurement integration: OpenAI's enterprise tier requires buyers to operate inside the ChatGPT Enterprise contract envelope, while Anthropic permits Skills to be invoked from any environment that can call the Claude API, including Bedrock, Vertex, and direct Anthropic endpoints. The third difference is the one that procurement teams cite most often as the deciding factor.

Brad Lightcap, OpenAI's chief operating officer, addressed the comparison directly at OpenAI's Tokyo developer forum on 14 May 2026 — five days before Anthropic's London launch. Lightcap argued that the Apps Store distribution model is the higher-velocity surface because consumer-facing distribution drives faster iteration cycles, and that the enterprise integration story will catch up by Q4 2026. The argument is internally coherent. It is also a concession that, on the enterprise procurement surface where the Fortune 100 buyers actually purchase, OpenAI is currently behind. The OpenAI Apps Store has only 14 apps that hold an enterprise-grade SOC 2 Type II attestation, and only three of those 14 have undergone the equivalent of Anthropic's Class A determinism logging review. By the procurement criteria that the CIO memos are using, OpenAI's enterprise-grade app cohort is smaller than Anthropic's by an order of magnitude — even though OpenAI's total app count is more than twice as large.

The publisher economics tilt the other way. The 75-25 OpenAI split is less publisher-friendly, but the OpenAI consumer surface has roughly 4.5 times the monthly active user base that Anthropic's Claude surface has, and consumer-facing publishers like Reddit, Spotify, and the New York Times have prioritised the Apps Store for that reason. The Skills Marketplace will not match OpenAI's consumer surface velocity in 2026. The publishers building Skills know this. They are choosing the procurement tier because the procurement tier is where the enterprise contract value lives, and the enterprise contract value is structurally larger per publisher than the consumer tail. Stripe's Payments Skill at $0.18 per invocation against a Fortune 100 deployment generates an annual gross revenue figure that the same publisher could not reach through a consumer-facing GPT app in any plausible distribution scenario.

The competitive overlap is genuine but the two surfaces are not direct substitutes. They are, in practice, two different go-to-market motions for the agent ecosystem: consumer-distribution-led versus enterprise-procurement-led. The interesting strategic question is what each owner will concede to break into the other's surface. OpenAI has signalled that enterprise procurement parity is a 2026 priority; the Apps Store team has begun publishing technical roadmap items that read as direct responses to Anthropic's Class A certification design. Anthropic, in turn, has begun investing in consumer Skill discovery — a Skills search interface inside the Claude consumer app shipped in beta on 12 May 2026, and the search interface is structurally similar to the OpenAI GPT discovery surface. The convergence is happening from both directions. The structural difference in publisher economics will continue to matter, because the 70-30 versus 75-25 split is not a number, it is a signal about which way the platform owner expects power to flow.

Who owns the agent-skill primitive

The deeper question the marketplace surfaces is who, structurally, owns the agent-skill primitive in the enterprise stack. The candidates are five: the model provider (Anthropic, OpenAI, Google), the orchestration framework provider (LangChain, LlamaIndex, Microsoft Semantic Kernel), the hyperscaler runtime owner (AWS Bedrock Agents, Azure AI Foundry, Vertex Agent Builder), the enterprise platform owner (Salesforce Agentforce, ServiceNow AI Agents, SAP Joule), and the publisher itself. The Skills Marketplace is Anthropic's attempt to anchor the primitive to the model layer — to position the model provider as the registrar of canonical agent capabilities and to make the orchestration framework, the hyperscaler runtime, and the enterprise platform downstream consumers of that registry. The strategy is structurally similar to how Apple positioned the App Store relative to the iOS device manufacturers and to the mobile operators in 2008: the registry-owner captures the platform-defining position even if it does not own all the downstream surfaces.

LangChain's response, telegraphed by Harrison Chase at the company's NYC developer summit on 8 May 2026, is to position LangGraph as the canonical orchestration runtime for Skills regardless of where the Skill is hosted. Chase's pitch is that Skills are portable across model providers and that the orchestration layer is the durable surface. The argument is consistent with LangChain's commercial logic — it monetises orchestration, not model provision — and it is the argument that the LangChain ecosystem has been making since the agent-framework category emerged in 2023. The Anthropic marketplace launch puts pressure on that argument because the Class A certification is Anthropic-specific. A Skill certified against Claude 4.5 family is not, by default, portable to a non-Anthropic model. The publishers building toward the certification are, in effect, accepting a degree of model lock-in in exchange for the procurement-grade credential.

The LangChain ecosystem has begun to fracture along that lock-in line. Two of the largest LangChain enterprise customers — a Tier-1 European bank and a US healthcare technology company, neither of whom Anthropic has named publicly but both of whom presented at Anthropic's London forum on customer panels — have signalled that they will require Skills certification on top of LangGraph orchestration for any new agent deployment after 1 July 2026. That requirement reduces LangChain's bargaining power, because it forces LangGraph into a downstream-consumer position relative to the Skills registry rather than an authoritative-runtime position. Chase has not commented publicly on the dynamic. Internal LangChain documentation, surfaced through procurement-side conversations with three buyers who had reviewed the company's product roadmap, indicates that LangChain is now investing in a parallel certification surface — branded as LangChain Trust — designed to provide a model-agnostic equivalent of Class A. The launch target is Q4 2026, which would be six months behind Anthropic in the enterprise procurement window.

The hyperscaler runtimes are taking a different position. AWS Bedrock Agents, in its 30 April 2026 release notes, added native support for invoking Anthropic Skills from within Bedrock workflows — but did so without conceding the registry function. Bedrock will host the runtime; Anthropic will host the catalogue. The split mirrors the AWS-Anthropic commercial relationship in which AWS hosts the inference and Anthropic owns the model IP. Vertex Agent Builder, by contrast, has positioned itself as a parallel registry for Gemini-compatible agent components and has been more openly competitive with the Skills concept. Google's developer relations leadership has signalled that the Vertex registry will be the canonical Gemini agent surface and that interoperability with Skills will be partial, not full. Azure AI Foundry sits between the two postures: Microsoft has committed to runtime support for Skills inside Foundry and has not yet declared a registry stance. The enterprise platform layer — Salesforce, ServiceNow, SAP — has uniformly chosen the runtime-not-registry position, accepting Anthropic's catalogue while preserving their own packaging and contracting layer above it.

The structural conclusion the marketplace forces is that the agent-skill primitive is being contested at the model layer rather than at the orchestration or runtime layers, and Anthropic has taken the first authoritative position. The position is contestable. Google can mount a counter-registry around Gemini agent components. OpenAI can extend the Apps Store enterprise tier toward parity by Q4 2026. LangChain Trust can ship and find buyers. But every counter requires either replicating the certification investment or accepting a downstream-consumer position relative to Anthropic's catalogue. The procurement memos circulating across the Fortune 100 are reading the marketplace launch as the moment the agent-skill primitive became a registered, governed, contracted unit — and they are budgeting accordingly.

What the 200 launch Skills tell us

The composition of the 200 launch Skills is itself an indicator. The largest single category is payments and commerce, with 38 Skills spanning Stripe, Adyen, Block, Shopify, Rakuten, Klarna, and Wise. The second is enterprise systems integration, with 31 Skills covering SAP, Oracle Cloud, Workday, NetSuite, ServiceNow, and Salesforce. The third is developer tools, with 27 Skills spanning GitHub, Atlassian, GitLab, Linear, Notion, and Anthropic's own internal tooling. The fourth is data and analytics, with 24 Skills covering Snowflake, Databricks, dbt, Fivetran, Hex, and Mode. The fifth is communications and collaboration, with 19 Skills covering Slack, Microsoft Teams, Zoom, Loom, Asana, and Monday. The remaining 61 Skills distribute across customer support, marketing automation, HR, finance and accounting, legal and contracts, supply chain, and a small clinical and life sciences cohort. The distribution is the procurement-grade enterprise surface, not the consumer-grade tail. That is the inventory choice Anthropic made.

The launch cohort also reveals the partner concentration. The top 10 publishers account for 71 of the 200 Skills — more than a third of the inventory. Stripe, Shopify, Block, Salesforce, Microsoft, Atlassian, ServiceNow, Workday, Snowflake, and Databricks dominate the visible launch surface. That concentration is a strategic choice: Anthropic prioritised depth-of-partnership over breadth-of-partnership for the launch, betting that procurement buyers will trust the registry faster if the launch publishers are the buyers' existing enterprise vendors. The long tail of smaller publishers will follow once the certification throughput scales and the marketplace's procurement reputation stabilises. The choice is conservative in the same sense that the Apple App Store's launch was conservative — 500 apps from a controlled set of partners, before opening to the broader developer community in iteration two.

The price points are clustering above the consumer-app norm. Tier Two metered Skills are pricing between $0.002 and $0.18 per invocation. The median is $0.045 — substantially above what the comparable OpenAI Apps Store metered apps charge. Three Skills are priced above $0.10: the Sanofi clinical trial protocol summariser at $0.16, the Adyen FX-routing Skill at $0.14, and the Snowflake schema-evolution Skill at $0.12. The high-end pricing reflects two things: the cost of compute for high-context Claude invocations, and the willingness of enterprise buyers to pay procurement-grade rates for procurement-grade Skills. The pricing data also tells us where the publishers expect their gross margin to come from. The Adyen and Stripe payments Skills are pricing for high-volume, low-margin invocation patterns. The Sanofi clinical Skill is pricing for low-volume, high-value invocation patterns. Both patterns are sustainable inside the Tier Two structure. The Tier Three enterprise-license pricing — $48,000 to $720,000 per seat-bundle annually — is set for vertical-specific Skills where the addressable buyer base is small but the per-buyer willingness-to-pay is high.

The implication for the broader market is that Anthropic has created a procurement category that did not exist 90 days ago. Inside the Fortune 100, every CIO office is now allocating budget against a Skills Marketplace line. The procurement teams are writing internal policies. The legal teams are drafting marketplace participation agreements. The risk teams are mapping the Class A certification against existing vendor risk frameworks. None of this infrastructure existed before 19 May 2026, and most of it will be in production by 30 September 2026. That speed is the consequential signal — not the 200 Skills themselves, but the fact that Anthropic has compressed what would normally be a 12-to-18-month enterprise procurement category buildout into 90 days. The compression is possible because the buyer pull is real. The buyer pull is real because Skills, as a primitive, solves a problem that the orchestration framework category has been promising to solve and has not yet solved at procurement scale.

What to watch

The Skills Marketplace is six days old. The next 90 days will determine whether the launch establishes the registry-owner position that Anthropic is targeting or surfaces a constraint that the competition can exploit.

• Whether Anthropic clears the throughput question — the company has committed to 90 reviewers by year-end, and the procurement pipeline currently sitting in Anthropic's submission queue, by intelligence from at least four publishers, exceeds 480 Skills awaiting Class A review; if throughput slips, the procurement category that the marketplace has created will outpace the supply of certified Skills and create an opening for the OpenAI Apps Store enterprise tier or for LangChain Trust.
• Whether Class B certification, scheduled for Q3 2026, lands on a credible cost-and-cycle envelope; if Class B requires buyer-funded attestation in the $80,000 to $250,000 per-Skill range — the figure that has been circulating in publisher conversations — it will compress participation to the largest publishers and reduce the long-tail distribution that the marketplace needs to reach the inventory targets Anthropic has signalled for 2027.
• Whether Google's Vertex registry counter-launch produces a parallel certification framework or accepts interoperability with Skills; the next consequential signal will be the Google I/O 2026 keynote on 9 June 2026, which is widely expected to include a Vertex agent registry announcement that will either align with or diverge from the Class A certification architecture.
• Whether LangChain's Trust framework ships on the Q4 2026 timeline and finds buyer adoption; LangChain's argument is that orchestration-layer certification is the durable abstraction, but the procurement teams currently writing Skills Marketplace participation policies are not writing parallel LangChain Trust policies, and the policy infrastructure inside a procurement office is sticky once it exists.
• Whether the Tier Three enterprise-license publishers — the 27 launch Skills that priced into the $48,000 to $720,000 annual range — close enough seven-figure contracts in Q3 2026 to validate the high-end pricing tier; if the seven-figure contract closure rate runs above 30 per cent of the launch Tier Three publishers, the marketplace economics will pull more vertical-specific publishers into the registry; if it runs below 15 per cent, the high-end tier will compress toward the metered Tier Two median and the structural revenue projection that Anthropic shared with the launch partners will need revision.

Frequently asked

What is Skills Verified — Class A, and how does it differ from SOC 2 Type II?

Class A is Anthropic's Skill-specific certification, covering seven evaluation surfaces: identity and access management, data residency disclosure, prompt-injection resistance, determinism logging, model-version compatibility, dependency disclosure, and incident notification. SOC 2 Type II is an organisational controls attestation that covers the publisher's overall security, availability, and confidentiality posture. The two are non-substitutable. A publisher can hold a SOC 2 Type II report and still fail Class A on the determinism logging surface, and a publisher can pass Class A on a single Skill while lacking the organisational controls required for Type II. The Skills Marketplace requires both for Tier Two and Tier Three placement.

How is the marketplace revenue split structured, and how does it compare to the OpenAI Apps Store?

Anthropic operates a 70-30 split, with 70 per cent of the gross contract value retained by the publisher and 30 per cent by Anthropic as the platform owner. The OpenAI Apps Store operates a 75-25 split favouring OpenAI. Both splits apply to Tier Two metered Skills and Tier Three enterprise-license Skills. Tier One free-layer Skills do not generate revenue share. The 70-30 split is one of the most publisher-friendly splits in any major enterprise software marketplace and was a deliberate choice by Anthropic to attract launch partners; Stripe, Adyen, and Shopify negotiated the 70-30 floor in the final 72 hours before the keynote.

Are Skills portable across model providers, or do they lock buyers into Claude?

A Class A certified Skill is tested and certified against a specific Claude family version and is not, by default, portable to a non-Anthropic model. Publishers can build cross-model Skill implementations, but the Class A certification will be model-specific. The model-version compatibility surface in the certification requires the publisher to declare which Claude versions the Skill has been tested on and to fail safely on unsupported versions. Procurement buyers who require model portability are using LangChain or LlamaIndex orchestration on top of Anthropic-certified Skills, treating the Skill as the canonical capability and the orchestration as the portability layer. LangChain's planned LangChain Trust framework, targeted for Q4 2026, aims to provide a model-agnostic certification equivalent.

What is the Skills review throughput, and how does it compare to the Apple App Store or AppExchange?

Anthropic's Class A review cycle for a new Skill is currently 11 to 14 business days. The Apple App Store review cycle for a new app is one to seven days. The Salesforce AppExchange review cycle for a managed package is six to eight weeks. The OpenAI Apps Store review cycle is two to five days for a consumer-grade GPT app. The Anthropic cycle sits between consumer-grade speed and enterprise-grade discipline, reflecting that Class A is a procurement-grade certification rather than a consumer safety check. The review team is currently staffed at 38 reviewers with a commitment to scale to 90 by year-end, which determines whether the cycle holds at 11-14 days or expands as submission volume grows.

How does the marketplace affect the LangChain ecosystem and the broader agent orchestration category?

The marketplace puts pressure on LangChain's positioning. LangChain's commercial argument is that orchestration is the durable abstraction and that Skills are portable across model providers via LangGraph. The Anthropic certification is model-specific, which means publishers building toward Class A are accepting a degree of Claude lock-in in exchange for the procurement-grade credential. Two of LangChain's largest enterprise customers — a Tier-1 European bank and a US healthcare technology company — have signalled that they will require Class A certification on top of LangGraph orchestration for new agent deployments after 1 July 2026, reducing LangChain's bargaining position. LangChain's response is the LangChain Trust framework, targeted for Q4 2026, which would offer a model-agnostic equivalent. The six-month gap between Anthropic's launch and LangChain Trust's expected ship date is the procurement window in which the registry position consolidates.

What does the launch tell us about who structurally owns the agent-skill primitive?

Anthropic has taken the first authoritative position at the model layer. The marketplace makes the model provider the registrar of canonical agent capabilities and positions the orchestration framework, the hyperscaler runtime, and the enterprise platform layers as downstream consumers. The strategy is structurally analogous to how Apple positioned the App Store relative to iOS device manufacturers and mobile operators in 2008 — the registry owner captures the platform-defining position even without owning all the downstream surfaces. The position is contestable: Google can launch a counter-registry around Gemini agent components, OpenAI can extend the Apps Store enterprise tier toward parity, and LangChain Trust can ship and find buyers. Each counter requires either replicating the Class A certification investment or accepting a downstream-consumer position relative to Anthropic's catalogue. The next 12 months will determine whether the registry position consolidates around Anthropic or fragments across multiple registries.