Anthropic just made your Claude commitment spendable at other vendors — without saying how much of it

What shipped

On 23 September 2026 Anthropic put three things that already existed separately behind one door.

The first shelf is connectors and plugins — over 2,000 of them, built on the Model Context Protocol and Agent Skills, the two open standards Anthropic has spent the past two years seeding. Atlassian, Google, Microsoft, Notion and Salesforce are the names in the announcement; the long tail is considerably longer.

The second shelf is products you buy. Claude-powered software from CrowdStrike, Cursor, Harvey, Legora, Lovable and Snowflake, listed for purchase rather than installation.

The third is people. Consultancies and systems integrators from the Claude Partner Network — Accenture, Boston Consulting Group and Deloitte at the premier tier — available to be engaged for rollouts.

Anthropic frames this as a discovery problem solved: “a single destination to find the right tools and services” for customers, “an easier way to reach teams using Claude” for builders. That framing is accurate and slightly beside the point. Anthropic has run a limited-preview marketplace since earlier in 2026, with roughly six partners; the plugin directory has lived inside Claude Code longer still. Catalogues of this kind are not scarce and are rarely decisive.

What is genuinely new sits in one clause, and most of the coverage walked past it.

The clause that matters

Eligible customers can put a portion of their committed Anthropic spend toward those partner purchases, with Anthropic consolidating invoicing under the existing customer relationship.

Anthropic supplies three worked examples: CodeRabbit applying funds toward Vercel, and Power Digital and ThoughtSpot directing theirs to Snowflake.

If that mechanism sounds familiar it is because it is the single most successful commercial construct in enterprise cloud. AWS Marketplace lets customers retire part of an Enterprise Discount Program or Private Pricing Agreement commitment through third-party software purchases. The cap is public: up to 25% of the annual commitment, negotiable to roughly 30–35% for organisations whose stacks lean heavily on third-party SaaS, and since May 2025 restricted to products that actually run on AWS. Google Cloud and Azure run close variants.

The mechanism is elegant because every party gains something immediately. The customer converts commitment risk — money committed and potentially unconsumed — into software it was going to buy regardless. The software vendor gets a procurement path that skips a new vendor-onboarding cycle, which in a large enterprise is measured in quarters. And the cloud provider gets a commitment that is easier to consume, which means customers sign larger ones and renew them more readily.

Anthropic has now built the same instrument. It is a more consequential move than any single model launch, because it changes the shape of the contract rather than the price of a token.

The number Anthropic did not publish

Here is where the launch stops being usable.

AWS tells you 25%. It tells you what qualifies, what does not, and what is negotiable. A finance team can take that to a renewal conversation and model it.

Anthropic says “a portion.” There is no published percentage, no minimum commitment that makes a customer eligible, no statement of whether the cap is standard or negotiated deal by deal, and no disclosure of fees or revenue share with listed partners. The route in is the one every enterprise-sales motion uses at this stage: contact your account team.

That gap is not a rounding error in the story — it is the story. A drawdown allowance of 25% changes which tools an organisation buys and how large a commitment it is willing to sign. A drawdown allowance of 2%, or one gated behind an eight-figure commitment, is a rebate with a press release attached. Those two worlds are indistinguishable from the announcement, and the difference between them is the entire buyer-side question.

Until Anthropic publishes the figure, the correct treatment is straightforward: the Claude Marketplace is a catalogue you can browse today and a budget instrument you cannot yet plan around. Anyone building next year’s tooling budget on the assumption of meaningful drawdown is modelling a number nobody has stated.

Listing is distribution, not assurance

The second thing worth reading carefully is what a listing certifies.

Anthropic’s own plugin directory answers this without hedging. It separates plugins built by Anthropic from external submissions, and external submissions pass a quality and security review before approval. It also carries this: Anthropic “does not control what MCP servers, files, or other software are included in plugins and cannot verify that they will work as intended or that they won’t change.”

That last clause carries the weight. A plugin is not a package of frozen code — it is a pointer to a server a third party operates and can change after you install it. Approval is a gate at the door, not a guarantee about the room.

This is the same exposure that has already produced real incidents. Microsoft’s Copilot connector chain was the vector in the Cosnitch memory-poisoning disclosure, where the trusted surface was not the model but what the model was allowed to read. The Langflow CVE that leaked OpenAI and AWS keys sat in the glue layer between an agent and the services it reached, not in the agent. And as Claude’s MCP surface has become a metered product inside other vendors’ platforms, the number of parties holding a token that can read your systems has gone up, not down.

A first-party catalogue raises the perceived trustworthiness of everything in it. On the evidence of Anthropic’s own disclaimer, that perception runs ahead of the assurance. The install-time review — who runs this server, what scopes does it want, what can it read, what happens when it updates — is still yours, and the self-hosted sandbox and tunnel patterns remain the containment story for anyone who needs one.

What this does to the competitive picture

Two effects are worth tracking.

The first is on partners. A vendor listed here has traded something for the distribution: part of its billing relationship now runs on Anthropic’s paper. That is a familiar bargain — it is how the cloud marketplaces work — and it is usually worth taking. But it concentrates leverage. OpenAI’s Astra for Law launch made the same structural point from the other side: when a lab’s partner roster becomes the market, the roster is a competitive asset the lab controls. Note that Harvey and Legora, direct competitors in legal AI, are both on this list, which is the lab keeping optionality rather than picking.

The second is on the neutral middle. Routers and gateways have argued that buying capacity through a vendor-neutral layer beats buying it from a lab. Drawdown is a direct counter-offer: keep the money inside the lab’s contract and we will make it spend further. The counter-argument has not changed — a commitment you can only spend in one ecosystem is a commitment you have to renew — but the offer is now materially better than it was, which is precisely how this mechanism works.

What to do with this

If you hold an Anthropic commitment, ask your account team for the specific drawdown percentage, the eligibility threshold, and which listings qualify, in writing, before it factors into any plan. Treat “a portion” as zero until it is a number. The right time to establish this is during a renewal conversation, not after.

If you are choosing tools, keep the drawdown out of the shortlist. Pick the coding tool or agent platform that wins on merit — the Cursor versus Claude Code question is decided by how your team works, not by whose invoice it arrives on — and only then ask whether the winner can be bought with committed money. Reversing that order is how organisations end up with a stack selected by a procurement mechanic.

If you are installing plugins, the catalogue changes nothing about your review. Inspect the manifest, identify who operates the MCP server, scope credentials narrowly, and re-check after updates. Developers treating a first-party listing as a substitute for that review are extending trust that Anthropic has explicitly declined to offer.

And note what this is, in sequence. Anthropic spent September cutting effective Opus pricing through defaults rather than list price, tying capability access to verification programmes, and now making the commitment itself more spendable. None of those is a model announcement. All of them make a Claude contract harder to leave — which, on the evidence of how quickly platform layers get retired elsewhere, is the variable worth watching more closely than the benchmark table.

Questions or corrections? Email Pick Right. Want the full list? See all news.