The Short Answer
Academic-integrity scoring, early-alert flags, admissions triage and help-desk routing are classification, not writing, and at published September 2026 prices a million such decisions cost $12,500 on GPT-6 Astra, $125 on GPT-6 Luna and $42 on TypeSafe's Jev. The 99% saving is real but mostly comes from model choice, not a new model class. With ibl.ai you own all the code and the data.
A university that moves its judgment work off a frontier model captures most of that saving today, without adopting anything new.
What share of a university's AI workload is judgment rather than writing?
Searching for a published figure does not produce one. A number of "60–80% of AI volume is classification" is circulating alongside the System One story, and no survey, vendor telemetry or institutional study surfaces behind it.
Treat it as a hypothesis about your campus, not a finding about campuses. It is also a measurable one — every AI platform logs its own calls, so the split is a query against data you already hold.
What is measured is the volume underneath. The Digital Education Council's AI in Higher Education Global Survey 2026 drew 45,398 responses across 35 countries and found 92% of students actively engaging with AI.
The tasks divide cleanly once you look at them. A tutoring exchange, a draft feedback comment and a research summary need prose. An integrity score, a risk flag, an eligibility determination and a ticket route need a category and a confidence number.
The second group is where the volume lives, because it runs per submission, per ticket and per application rather than per conversation.
What does one classification decision cost at published 2026 model prices?
Priced from the vendors' own rate cards, the spread is three orders of magnitude. The table below holds the workload fixed and changes only the model.
| Model | $/MTok in | $/MTok out | 1M decisions |
|---|---|---|---|
| GPT-6 Astra (frontier) | $10.00 | $50.00 | $12,500 |
| Claude Opus 5.5 (frontier) | $4.00 | $20.00 | $5,000 |
| Claude Haiku 4.5 (small) | $1.00 | $5.00 | $1,250 |
| GPT-6 Luna (small) | $0.10 | $0.50 | $125 |
| Jev (TypeSafe, System One) | $0.042 | unmetered | $42 |
Sources for every rate in that table are linked at the foot of this post: OpenAI's and Anthropic's published API pricing, and TypeSafe's launch post for Jev.
Does a judgment model really cut university AI costs by 99%?
Against a frontier model, yes — and that is the misleading part of the claim.
Astra to Jev is a 99.7% reduction on the table above. But Astra to Luna, the same vendor's own small model reached through the same API, is already 99.0%. Jev then takes $125 to $42, a further 66%.
So the headline percentage is almost entirely an argument against running classification on a frontier model. It is not an argument that a new model class is the thing standing between a university and a smaller bill.
That reframing matters because the two conclusions imply different work. One is a procurement decision about a vendor you do not yet have. The other is a routing decision inside the platform you already run, available this week.
The expensive mistake is not late adoption of System One models. It is sending an integrity score to whichever model the platform defaults to, because the platform only offers one — the structural problem behind the real math on per-student AI pricing.
One correction to the story as it is being told. TypeSafe's CEO Diogo Almeida is frequently described as a co-creator of ChatGPT.
TypeSafe's own team page says he co-invented RLHF and InstructGPT, "the methods that lead to ChatGPT and GPT4," and that he was previously at Google Brain. The narrower claim is the one the company makes.
Which university workloads are classification rather than generation?
Six, and each is high-volume, repetitive and scored rather than written.
- Academic-integrity scoring — a similarity or provenance signal with a confidence value, not a written verdict. A person drafts the explanation later, for the cases that escalate.
- Early-alert risk flags — a probability over engagement, grade trajectory and attendance. The advising outreach that follows is generation; the flag that triggers it is not.
- Admissions triage — completeness checks, prerequisite matching and rubric scoring against a defined band. Thousands of runs per cycle, each returning a category.
- Course routing — ranked options under prerequisite, schedule and degree-audit constraints, with a score per option.
- Aid eligibility — rule-heavy determination with edge cases, where the required output is a bracket and a reason code.
- Help-desk ticket routing — a queue, a priority and a confidence. The highest-volume item on most campuses and the easiest to measure.
Every one of these returns a structured value. Asking a text model for it means paying for a decoding step whose output you immediately parse back into a field.
What has anyone actually measured about judgment models in production?
One narrow public test, published five days after launch.
LangChain ran an agent-evaluation comparison and published the results on 20 September 2026. Across 500 pass/fail decisions on a weather agent, Jev agreed with the human reviewer 100% of the time at $0.00035 per call and 0.44 seconds average latency.
The comparison judges were GPT-5.6 Luna, GPT-5.6 Terra and Claude Sonnet 4.6, the test having published on 20 September, two days before GPT-6 Luna shipped. Terra agreed 99.8% of the time, Luna 96.4%, Claude Sonnet 4.6 80.0%.
Total spend was $0.34 for Jev against $28.17 for the Claude judge.
LangChain states the limitation plainly: they still need to see whether the result carries to other agents and production workflows, and note that cheap judging can amplify an error across far more decisions.
That is the right way to read it. One agent, one task family, one vendor's model under test — a direction, not a procurement case. What Jev is and what it is not was covered when it launched, in generation is commoditized and judgment is the new frontier.
How does ibl.ai route university judgment work to the model it needs?
With ibl.ai you own all the code and the data.
The platform runs inside the university's own perimeter with full source code under a perpetual license, is model-agnostic across any LLM so each agent uses the model its task actually requires, is usage-based with no per-seat pricing against a credit ceiling the institution sets, and you can deploy anywhere — your own cloud, on-premise, GovCloud, or a fully air-gapped network.
The swappable model layer is the point here: it lets an institution pick the right model per task and migrate without code changes, so an integrity scorer and a tutoring agent do not have to share a price tag.
A September 2026 release made that layer self-maintaining. The model catalogue now auto-syncs from the gateway every six hours.
A discovered model stays inactive until a verification gate has tested multi-turn recall, streaming, tool calling, image input and a real non-zero cost. So a cheaper model reaches the picker without a redeploy, and an unverified one never does.
Pricing follows the same shape. Self-serve runs on prepaid credits with a hard cap; pilots start at $15K; deployment is a one-time $25K–$80K engagement; full codebase ownership is a custom quote. None of it is priced per student.
1.6M+ users across 400+ organizations run the platform this way, including NVIDIA, MIT, and Syracuse University. The higher-education deployment and Agentic OS pages cover the rest.
ibl.ai is family-owned and operated from New York, NY.
Related reading: per-student AI pricing: the real math for universities — what the same workload costs when the meter is headcount instead of tokens · generation is commoditized, judgment is the new frontier — what Jev is, what it claims, and who should own the rubric.
Sources: Jev's $0.042/MTok input price, unmetered output and 40×–200× speed range from TypeSafe's launch post; Diogo Almeida's description from TypeSafe's team page. Model rates: GPT-6 Astra and GPT-6 Luna from OpenAI's API pricing; Claude Opus 5.5 and Haiku 4.5 rates from Anthropic's pricing page. Also: the 500-decision agreement, cost and latency figures from LangChain's 20 September 2026 test; adoption figures from the Digital Education Council's AI in Higher Education Global Survey 2026.