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

Short answer: Gemini 3.8 Flash is the cheapest long-context model among the major providers and suits high-volume document work, but inside Edor.ai it is not a native provider and is reachable only through Poe, and today's low rate is introductory pricing that doubles in January 2027.

What this is: a vendor that has put everything behind Flash

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 the one a school has most likely met already, Google Workspace for Education. All three sit on the same underlying models, but their terms, data handling and available features differ, and that is the most common source of confusion in a procurement discussion.

Google's product 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.

There is a second track worth a school's attention: Gemma. This is Google's lightweight open-weight family, which can be downloaded and run on a school server so that no data leaves the campus network. Gemini and Gemma are different things and procurement papers need to say which one is meant.

The current line-up: one Flash trunk and an open-weight branch

Gemini 3.8 Flash is the current workhorse, released on 2 September 2026 with roughly 1.05M tokens of context, 65K output, and a thinking_level that can be set to low, medium or high. Pricing is $0.75 per 1M input tokens and $3.75 per 1M output tokens. That is an order of magnitude below the OpenAI and Anthropic flagships, but it has to be read alongside one sentence: this is introductory pricing, and the standard rate of $1.50 / $7.50 begins on 1 January 2027. There is also a cybersecurity variant, Gemini 3.8 Flash Cyber, offered only to trusted defenders through the Fairwind Program, which is not relevant to schools.

Behind it sit 3.7 Flash (13 August 2026), 3.6 Flash and 3.5 Flash-Lite (21 July 2026), and 3.5 Flash (19 May 2026). Three iterations inside six weeks is itself a factor for a school hoping to standardise on one version for three years.

On the open branch, Gemma 4 arrived on 2 April 2026, built for reasoning and agentic workflows, with a 12B version following on 3 June 2026 that can be self-hosted through Ollama. The full list of versions, dates and prices is in the automatically generated series table further down this page.

Strengths: the four that matter most to a school

1. The lowest cost per token. Processing the same forty-page school document, Gemini 3.8 Flash's input cost is roughly a thirteenth of GPT-6 Astra or Claude Fable 5.1. For batch work over teaching materials or past papers, the difference is real money.
2. The most complete native multimodality. Text, image, audio and video can all be fed in directly. A teacher who wants the model to watch a demonstration video and draft observation points does not have to convert anything first.
3. A genuine open-weight branch. Gemma 4 can be downloaded and self-hosted, in sizes from E2B and E4B on a laptop up to 12B, 26B and 31B on a workstation, so a school can match the model to its hardware budget.
4. It sits naturally beside existing Workspace use. Where Google Classroom and Google Drive are already in daily use, teachers know the interface and account model, which lowers the training burden.

Limits: what to say before anyone signs anything

• It is not a native provider here. Edor.ai natively supports OpenAI, Azure OpenAI, Anthropic, Poe and local Ollama. Gemini is reached through Poe, which puts an additional third party in the path and means a privacy assessment must cover Poe as well.
• The price doubles in January 2027. The $0.75 / $3.75 rate is introductory, not permanent. A three-year budget written at today's rate goes wrong in year two.
• Poe does not carry embedding or speech. Google itself publishes embedding and speech-recognition models, but they are not in what Poe serves. So even with Gemini as the main model, the knowledge base and oral practice still need OpenAI or Azure OpenAI.
• Versions move very quickly. Three releases in six weeks is a burden on a system a school wants to keep stable. A prompt template tuned this term may need retesting after the next version.
• Hallucination and local detail are unchanged. Gemini gets Hong Kong curriculum specifics, local assessment terminology and school names wrong in the same confident tone as any other model. Outward-facing documents and marking results need teacher review.

Where each tier fits

Teachers: summarising and consolidating large volumes of documents, turning video or image material into written key points, and generating or rewriting questions in bulk. Because the unit price is so low, high-volume work of moderate difficulty is where it pays off most.

Students: only inside a gated, monitored environment. Note that when calls go 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 four-layer gating more important on this route than on others.

IT coordinators: if data residency is the main concern, assess self-hosted Gemma 4 before reaching Gemini through Poe. The first keeps everything inside the campus network; the second involves two overseas third parties. A sensible first step is a proof of concept running gemma4:12b or 26b on a single workstation before deciding on purchase scale.

Prompts worth testing

Copy these into a Edor.ai teacher tool or any Gemini interface. The constraint lines are what make the output usable:

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.

Below is a class-level summary of a Primary 5 English writing task, which I compiled after marking each script. Please do two things.
First, identify the three most common error types, each with two examples that actually occurred in the class.
Second, design a 10-minute remedial classroom activity for each error type, stating what the teacher does and what the pupils do.
Constraints: write the explanation in Hong Kong English conventions and keep example sentences in the pupils' original wording; do not rewrite anything beyond the quoted sentences; if you are unsure what the Hong Kong curriculum expects for a grammar point, mark it "teacher to verify".
[paste summary]

Availability 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, so a school wanting Gemini inside the platform reaches it through the Poe provider. Once an administrator adds a Poe key, the Gemini family can be selected from the supported model list.

The second route is self-hosting Gemma 4 on Ollama. This suits schools with strict data residency requirements: the weights are downloaded onto a school server and student conversations never leave the campus network. The cost is hardware plus capability — the mid-sized Gemma 4 builds still trail the cloud flagships on long-form Chinese marking.

On either route, remember that Poe carries no embedding or speech models. The school knowledge base and oral practice need a separate OpenAI or Azure OpenAI key. Schools that already hold Google Workspace for Education can run it alongside the school platform, but the data and audit records are kept separately and procurement and privacy papers should account for each on its own.

Alternatives

• For one key covering chat, embedding and speech, see OpenAI.
• For a natively supported provider with steadier Chinese writing feedback, see Anthropic Claude.
• To keep data inside the campus network while retaining strong Chinese-language ability, see Alibaba Qwen or Meta open weights, and read the open versus closed cost and privacy trade-off.
• For every provider's pricing and positioning in one place, read the 2026 all-model comparison.