Google Gemini (3.8 Flash / Gemma 4): the cheapest long context, and how a Hong Kong school actually reaches it

Gemini 3.8 Flash is the cheapest long-context model among the major providers and suits high-volume document work — yet the detail Hong Kong schools most often miss is that inside Edor.ai it is reachable only through Poe, and today's low rate is introductory pricing that doubles in January 2027.

Google's models are built by Google DeepMind and reach the outside world through three doors: the consumer Gemini App, the developer Gemini API, and Google Workspace for Education, which many schools have already met. All three sit on the same underlying models, but their terms, data handling and available features differ — the most common source of confusion in a procurement discussion. For Hong Kong schools there is a second distinction: Gemini (cloud API) and Gemma 4 (open weights you can self-host) are different things and procurement papers need to say which one is meant.

Flash first, Gemma on-prem

Google's strategy over the past year has been unambiguous: concentrate on the Flash line. Between May and September 2026 Flash shipped four versions, 3.5, 3.6, 3.7 and 3.8, while the Pro line's last public update remains Gemini 3.1 Pro in February 2026. What Google is selling is not the biggest or strongest model but one that is fast enough, capable enough and considerably cheaper.

The second track is Gemma 4: open weights Google publishes for download onto a school server, so no data leaves the campus network. It arrived on 2 April 2026, with a 12B version following on 3 June 2026 that can be self-hosted through Ollama. Sizes run from E2B and E4B on a laptop up to 12B, 26B and 31B on a workstation — procurement must state whether the school is buying cloud Gemini or self-hosted Gemma.

Schools that already hold Google Workspace for Education often run both tracks: teachers use Google's own tools inside Workspace for everyday documents and a light assistant, while the school platform picks a provider per subject with data and audit records kept separately, and privacy papers should account for each on its own.

Model line-up at a glance

The September 2026 Google line-up is one Flash trunk and one open-weight branch:

Gemini 3.8 Flash's $0.75 / $3.75 is roughly a thirteenth of GPT-6 Astra or Claude Fable 5.1 — processing the same forty-page school document, the input cost gap is real money. It must be read alongside the introductory deadline: from 1 January 2027 the standard rate is $1.50 / $7.50. There is also a Gemini 3.8 Flash Cyber variant offered only to trusted defenders through the Fairwind Program, which is not relevant to schools.

Three Flash iterations inside six weeks (3.5 → 3.6/3.5-Lite → 3.7 → 3.8) is itself a factor for a school hoping to standardise on one version for three years.

What you pay beyond Workspace

Google Workspace for Education AI features and the Gemini API are different contracts with different billing. If a school platform needs Gemini for batch teaching materials or student conversation, the current route inside Edor.ai is through Poe, not a direct Google API key.

The billing logic has three layers:

Introductory pricing has an end date. The $0.75 / $3.75 rate is not permanent. A three-year budget written at today's rate goes wrong in year two. Edor.ai's fixed annual fee means the platform carries that movement.

Poe adds a third party. The request reaches Poe first and then Google, so a privacy assessment must cover both, not Google's DPA alone.

Self-hosted Gemma trades hardware for control. Hardware cost and IT maintenance sit with the school, but student conversations never leave the campus network. Mid-sized Gemma 4 builds still trail the cloud flagships on long-form Chinese marking — that is a trade-off, not a flaw.

Writing $0.75 / $3.75 into a grant application as if it were permanent automatically underestimates the bill by half from January 2027 — the introductory deadline belongs in the budget footnotes.

Two routes inside Edor.ai

Google is not a native provider in Edor.ai. The five native routes are OpenAI, Azure OpenAI, Anthropic, Poe and local Ollama.

Route one: Gemini through Poe. An administrator adds a Poe key and selects the Gemini family from the supported model list. This covers chat models only — Google's own embedding and speech models are not in what Poe serves, so the school knowledge base and oral practice still need an OpenAI or Azure OpenAI key.

Route two: self-hosted Gemma 4 on Ollama. This suits schools with strict data residency requirements. An IT coordinator can run a proof of concept on one workstation with gemma4:12b or 26b before deciding on GPU purchase.

On either route, conversations pass through four-layer safety gating and ai_audit_logs, and student personal-data fields are encrypted at rest. Through Poe, the content-safety posture is determined jointly by Poe and Google and the school cannot tune it item by item, which makes the platform's own gating especially important on this route.

Bulk work and the privacy floor

Worth it for: summarising and consolidating large document volumes (the unit price is low — input cost on a forty-page school document is roughly a thirteenth of the flagships); turning video or image material into written key points (native multimodality: text, image, audio and video fed in directly); generating or rewriting questions in bulk; schools already on Google Classroom and Drive where training cost matters.

Students: only inside a gated, monitored environment. Through Poe the safety posture cannot be tuned item by item, so the platform's four-layer gating is the main line of defence.

Not worth it when: you expect one Google key to cover embedding and speech — Poe does not carry them; you budget long term at the introductory rate without noting the January 2027 doubling; you need version stability but pick the Flash line with three releases in six weeks; data residency is strict but you still route through Poe instead of self-hosting Gemma; you publish AI-consolidated school documents without teacher review — Hong Kong curriculum detail and assessment terminology still go wrong in the same confident tone as any other model.

A junior secondary science teacher told a colleague at break: "I fed the demonstration clip into Gemini and it listed three observation points for the pupils — but I still marked the refractive index column 'teacher to verify' before putting it on the worksheet." That is normal use.

Copy this into an Edor.ai teacher tool to test:

You are a Hong Kong junior secondary science teacher. My Secondary 2 class will run a practical on the refraction of light using a semicircular acrylic block, a ray box and a protractor.
Output: the procedure written in short sentences a student can follow, a results table with columns for angle of incidence and angle of refraction, three post-lesson thinking questions, and a teacher-only safety note.
Constraints: write in Hong Kong English conventions; do not give away answers; do not assume the school has digital sensors; for any specific refractive index value, write "teacher to verify" and say which kind of reference to consult.

The trade-off for schools

• If you need high-volume material processing and already use Google Workspace, treat Gemini as a bulk tool, but the budget table must note the standard rate of $1.50 / $7.50 from 1 January 2027 — do not estimate three years at $0.75 / $3.75.
• If data residency is strict, assess self-hosted Gemma 4 before reaching Gemini through Poe — the latter involves two overseas third parties, Poe and Google.
• If you need the knowledge base and oral practice running inside Edor.ai alongside Gemini, keep an OpenAI or Azure key even when Gemini is your main chat model — the Poe route does not cover embedding or speech.