What SUTD actually rolled out
On 14 September, SUTD launched the SUTD LLM Gateway, a centralised platform that replaces what the university describes as fragmented, individually funded subscriptions. Previously, staff and students paid for different providers out of course or project budgets. Now there is one "common front door".
The catalogue includes GPT, Claude, Gemini, Qwen and Kimi, spread across 15 models from six providers in the primary chat interface. An API key extends access to more than 500 additional models in a broader marketplace — which is a very different product from the chat window, and a much harder one to govern.
The privacy design is where the engineering effort went. The gateway automatically redacts sensitive data such as ID numbers before a request reaches a provider, does not retain prompts, and contractually prevents providers from training on university data. SUTD also stresses that students should critically assess AI output rather than trust it.
That last sentence is easy to skim past. It is also the only part of the announcement that is genuinely about education rather than procurement.
Fifteen models, one bill — and a spread nobody budgets for
A single gateway is usually sold as convenience. In practice it is a cost-control instrument, because the models sitting behind it are not remotely comparable in price. Using current list prices per million tokens:
| Model | Input / 1M | Output / 1M |
|---|---|---|
| GPT-6 Astra | $10.00 | $50.00 |
| Claude Fable 5.1 | $10.00 | $50.00 |
| Grok 4.6 | $2.00 | $6.00 |
| Gemini 3.8 Flash | $0.75 | $3.75* |
| Mistral Medium 3.5 | $0.50 | $1.50 |
| DeepSeek-V4.1-Flash (off-peak) | $0.15 | $0.60 |
| Muse Spark 1.3 (blended) | ~$0.10 | — |
*Introductory rate through 31 December 2026. GLM-5.3 has no announced standalone per-token price yet.
Read that table as ratios, not numbers. GPT-6 Astra costs 13.3x more per input token than Gemini 3.8 Flash and 66x more than DeepSeek-V4.1-Flash at off-peak pricing. On output — which is where most chat workloads actually burn money — the gap between the flagship tier and DeepSeek is over 83x. The same university, the same gateway, the same toggle in a dropdown.
Our own back-of-envelope, clearly labelled as such
SUTD has not published usage volumes, so this is our arithmetic, not their data. Assume each of the 3,000 accounts sends 200,000 input tokens and 50,000 output tokens per month — a modest load for someone writing code with an assistant.
That is 600 million input tokens and 150 million output tokens per month. Routed entirely to GPT-6 Astra: $6,000 in, $7,500 out — $13,500 per month. Routed entirely to Mistral Medium 3.5: $300 plus $225 — $525. Routed entirely to DeepSeek-V4.1-Flash off-peak: $90 plus $90 — $180.
The gap between the flagship-only strategy and the mid-tier European option is roughly $12,975 per month, about $155,000 a year. And European buyers pay VAT on top of US list prices — 21% in Czechia — which many budget spreadsheets quietly forget until the invoice arrives.
This is exactly the trade-off we try to quantify on our own rig. The AI Arena benchmark setup exists because "which model is better" and "which model is affordable at volume" are two different questions, and vendors answer only the first one loudly. In production, the honest answer is usually a routing mix: a cheap model for classification, summarisation and first drafts, an expensive one only for the genuinely hard 5% of requests.
The European angle: what SUTD does not have to solve
Singapore's approach is worth copying, but not transplanting. Three things change on this side of the world.
GDPR. A redaction layer and a no-training clause are prudent engineering. For an EU institution they are also contractual artefacts: the university remains a controller, needs a lawful basis, needs Article 28 processor agreements and — where prompts leave the EEA — a Chapter V transfer mechanism. A gateway makes those obligations easier to evidence, but it does not discharge them. Redaction reduces risk; it does not replace paperwork.
The AI Act. General-purpose AI obligations have applied since 2 August 2025, and AI literacy duties under Article 4 have been in force since 2 February 2025. A university handing 3,000 students a model menu is a deployer, and the literacy obligation is the one most institutions are still ignoring. SUTD's "critically assess the output" line is, functionally, an AI literacy policy. Europe requires roughly the same thing — with a legal hook attached.
Procurement. A European university rarely buys directly from five US vendors. It buys through national framework agreements, research consortia and increasingly through sovereign or EuroHPC-backed capacity, which is where the current political momentum sits. A single internal gateway is precisely what makes that switchable later: if the backend is one interface, changing the supplier underneath it is a configuration task, not a migration project.
What is actually worth copying
Four things, in this order. One: put a redaction layer in front of every provider rather than trusting users to paste carefully. Two: make no-training and no-retention contractual, in writing, per provider. Three: route by task, not by brand loyalty — the price spread above is not a rounding error. Four: keep the chat interface and the API catalogue separate, because 500 extra models is a governance surface, not a feature.
What is not worth copying is the assumption that a well-built gateway answers the legal question. It answers the engineering question. In Europe, those are two different documents.
Can a smaller European organisation build the same thing without a university-scale budget?
Yes, and mostly without buying anything new. A self-hosted router in front of two or three APIs — one cheap model for bulk work, one premium model on demand — is a weekend project for a competent developer. The hard part is not the software; it is the per-provider contracts covering data retention and training use, which smaller buyers rarely have the leverage to negotiate.
Does using a gateway mean the institution sees everything users type?
Not necessarily, and the two should be decided separately. SUTD's design, as described, does not retain prompts. But redaction and logging policy are configuration choices, not inherent properties of a gateway. Any institution rolling one out should publish what is logged, for how long, and who can read it — staff will ask anyway, and vague answers destroy trust faster than an honest logging policy does.
Why does off-peak pricing matter for universities?
Because student workloads are already batch-shaped. Assignment deadlines, overnight grading runs and bulk document processing can be shifted a few hours, and some providers price that movement explicitly — DeepSeek's off-peak rate is half the peak rate. For a fixed-cost institution, scheduling flexibility is free money.
