Direct from the Lab or via a Reseller – What's the Difference?
Anthropic and OpenAI sell directly from the lab: clearer contracts, day-one access to new models, and token pricing without markup, but EU residency is built via a hyperscaler. GitHub Copilot and Azure AI Foundry are resellers: one contract and one governance model across several models, against inherited terms, slower deprecation, and documented markups.
Behind every AI procurement lies a dividing line that shapes both price and governance: do you buy directly from the lab that trains the model, or via a reseller that packages it? Anthropic and OpenAI sell directly. GitHub Copilot and Azure AI Foundry are, in practice, resellers – they sell the same underlying models plus their own additions. Both paths are reasonable; they just optimize for different things. The dividing line is older than AI – direct versus through a distributor – but the pricing dynamics in AI make it unusually expensive to ignore.
The direct path: closer to the lab
A direct deal gives you a shorter chain between you and the model:
- Clearer contracts. The DPA, zero data retention, and sub-processor chain are negotiated with one party – the lab itself.
- Day-one access. New models are available right at release, not once the reseller has managed to certify them.
- Token pricing without markup. You pay the lab’s list price for actual consumption.
The trade-off: you build it yourself. The EU residency leg runs via a hyperscaler or dedicated EU environments, integration with your systems is your responsibility, and each lab is its own contract. The direct path suits organizations with their own platform expertise and a few, heavy use cases. Just don’t underestimate the build: residency configuration, key management, and monitoring are real work items, and they need staffing even after launch.
The reseller path: one contract, several models
The platform path optimizes for governance and simplicity:
- One contract, one invoice. Several labs’ models under the same commercial umbrella.
- One governance model. Identity, permissions, logging, and policies in one place, often within your existing cloud.
- Ready-made residency tools. Region selection and data zones are included in the platform’s configuration.
The trade-off here: inherited terms from the labs underneath, deprecation at the reseller’s pace, and documented markups. For the Azure path, the markup has been estimated at roughly 15–40 percent of total cost of ownership, and Copilot’s credit-based billing can multiply the cost several times over for agent-heavy flows compared with pure token pricing. The markup buys something real – governance, consolidation, and a single counterparty – but it should be a deliberate purchase that shows up in the calculation, not a surprise after the fact.
| Aspect | Direct from the lab | Via a reseller |
|---|---|---|
| Contract and DPA | Direct with the lab, short chain | One contract – with the lab's terms inherited underneath |
| New models | Day one | At the reseller's pace |
| Price | Token pricing without markup | Markup or credit model |
| EU residency | Built yourself via a hyperscaler | Included in the platform's configuration |
| Governance | One contract per lab | One model across several labs |
A worked example: agent flows and credits
The difference shows up most clearly in agent-heavy flows. A coding agent solving a task doesn’t make one call – it reads files, proposes changes, runs tests, and iterates, often dozens of calls per task.
With a direct deal’s token pricing, you pay for tokens actually consumed, no more, no less. Under a credit model, every premium call gets billed in credits, and the multiplier means the monthly cost per developer can end up several times higher than the estimate built on “normal” chat usage. Same team, same model – completely different final bill depending on the purchase path. The mechanism also works the other way: for an organization where most people just ask an occasional question per day, a packaged license can be cheaper than building and running your own API solution.
The takeaway isn’t that credit models are always wrong, but that the calculation has to be run on your own usage profile. And since both markups and pricing models keep changing: use current price lists, not last year’s blog post.
How to choose
A few rules of thumb that hold in most cases:
- Few models, high pace, in-house platform expertise – a direct deal gives the best price and fastest model access.
- Many users, existing cloud governance, need for a single invoice – the reseller path pays off in administration and control.
- Regulated operations – weigh the direct path’s short contract chain against the platform’s ready-made residency configuration; it’s the documentation burden that differs, not the possibility.
- Often the answer is hybrid – platform for breadth, direct deal for the core product.
Also write into the decision which metrics would make you switch paths – cost per user, cost per task, time to new model. Then reconsideration becomes a routine instead of a negotiation.
At Weapp we’ve built AI solutions along both paths and are happy to help with the calculation for your own volume and usage profile – get in touch and we’ll run the numbers together.
Frequently asked questions
What does it mean that the reseller inherits terms?
The reseller sells models it didn't train itself, so the lab's terms stay underneath – retention, deprecation pace, and model restrictions carry through into the reseller's package. You negotiate with one party but depend on two, and that should show up in your sub-processor chain.
Is the markup always more expensive overall?
Not necessarily. One contract, one invoice, and a shared governance model save on administration and legal work, which can offset the markup at moderate usage. At high volumes or agent-heavy flows, though, the markup grows fast in kronor – calculate it against your own usage profile.
Can we combine a direct deal and a reseller?
Yes, and it's common: for example a platform path for the breadth of users and a direct deal for your own product where price and model access matter most. The requirement is a clear division of labor so data and cost flows don't get mixed up.
How big is the markup at resellers?
For the Azure path, the markup has been estimated at roughly 15 to 40 percent of total cost of ownership compared with a direct deal, and credit-based models like Copilot's can multiply the cost for agent-heavy flows several times over. Levels change – always check current price lists.
Why do agent flows get so expensive under credit models?
An agent doesn't make one call per task but dozens – it reads, tries, reruns, and iterates. Under a credit model, every premium call gets billed, so the multiplier hits the monthly cost directly. With token pricing, you pay for actual consumption with no markup layered in.