The Short Answer
The agent runtime became free this month, so a per-seat AI fee now has to be justified by everything except the runtime. On ibl.ai you own all the code and the data, run it model-agnostic across any LLM, and pay with no per-seat pricing — so extending AI to every student is a deployment decision rather than a budget event, and you can deploy anywhere, including fully air-gapped.
Two things happened in August that change the shape of a campus AI negotiation.
DeepSeek open-sourced Harness on August 13 under an MIT licence. Within six days it reached 166,992 GitHub stars and 17,813 forks.
Microsoft's Agent Framework Harness reached general availability — the agent loop, planning, memory, context compaction, approvals and telemetry, supported, in Python and .NET.
The runtime that turns a model into an agent is now available free from two directions. Meanwhile a large number of institutions are paying per student, per month, for products built on exactly that layer.
What does a per-seat AI fee actually buy now?
It is a fair question and it has real answers — just fewer of them than it did in July.
Legitimate things a per-seat fee can still buy: permissions-aware retrieval across the SIS and LMS that honours what each student and staff member may actually see. Guardrails calibrated for an eighteen-year-old rather than an enterprise employee.
Audit logging that survives a FERPA review. Integration with Banner, Workday Student, PeopleSoft or Canvas. Someone accountable when it breaks during registration week.
What it can no longer buy is the agent loop. That specific claim moved from differentiator to commodity this month, and any renewal conversation should reflect it.
The useful exercise is simply to ask the vendor to enumerate what is included beyond the runtime — and then check how much of that list is also available as a licence rather than a subscription.
How much is per-seat actually costing an institution?
Enough that the structure, not the rate, is the problem.
Per-seat AI pricing runs roughly $30 per user per month for Microsoft Copilot and about $60 for ChatGPT Enterprise. Applied to a 20,000-student institution at the lower figure, that is $7.2 million a year — for a platform whose core runtime is now MIT-licensed.
The deeper issue is what per-seat pricing does to the deployment. Because every additional student carries full list price, institutions ration licences to a pilot cohort.
The pilot then underperforms, because the students who would benefit most from an always-available tutor are the ones who did not get a seat.
That dynamic is the subject of Per-Student AI Pricing: The Real Math for Universities, and it does not improve when the underlying runtime becomes free — it just becomes harder to justify.
Does a free runtime mean universities should build their own?
No, and this is where the enthusiasm needs a brake.
DeepSeek Harness is a developer preview whose own documentation states there will be compatibility-breaking changes. Building an institutional platform directly on a six-day-old preview is not a plan.
More importantly, the runtime was never the hard part.
What stands between a free agent loop and something a university can actually deploy is permissions-aware retrieval across systems of record, an evaluation set built from institutional data, guardrails, RBAC, audit logging, and the integrations.
That is the same 80% that consumes AI budgets everywhere, which is why 79% of enterprises reported AI cost overruns in the past twelve months and 88% of AI pilots never reach production.
A free runtime removes one component. It does not remove the platform problem, and a university that mistakes one for the other will spend a year discovering the difference.
What is the third option?
License a platform that already exists, own the code, and pay for it once rather than per student.
That is the structure that makes a commoditized runtime an advantage instead of a threat. The platform is already built and running, so the institution is not funding construction.
The source ships under a perpetual licence, so a better free runtime can be adopted underneath rather than triggering a migration. And the licence is flat, so extending AI to every enrolled student is a deployment decision rather than a budget event.
ibl.ai is in production with 1.6M+ users from 400+ organizations and is model-agnostic, so the model layer and the runtime layer both stay replaceable. You own all the code and the data, and it deploys on your cloud, on-premise, or fully air-gapped.
The broader commoditization argument is in The Agent Runtime Just Commoditized. Now What?, and the procurement questions to put to any bidder are in An RFP Checklist for AI Platform Procurement.
What should a CIO do before the next renewal?
Three questions, and they take one meeting.
What in this contract is the runtime, and what is everything else? Ask for the breakdown explicitly. A vendor that cannot separate them is telling you something.
What does this cost at full enrolment rather than at the pilot? Per-seat quotes are usually presented at the licensed subset. Multiply by the population you actually want to serve.
What do we hold if we leave? A subscription ends.
A perpetual licence with source code running on university infrastructure does not — and that distinction now matters more, because the free runtime underneath means the platform above it is the only thing you were ever really buying.