Earlier this month, a Washington deputy superintendent said something every leader weighing open-source AI models for schools should hear. Talking with EdTech Magazine about AI policy, Lake Washington School District’s Scott Beebe put it plainly: “The models now change on a cadence of weeks, not months or years.”
He’s right. In the first week of November alone, BenchLM’s news roundup tracked a 65-billion-parameter model released under the permissive Apache 2.0 license, the full weights of a 2.8-trillion-parameter model posted publicly, and two more labs promising open weights soon.
So here’s the question I keep getting from IT directors: if the models change every few weeks, which ones should a district actually use? My answer is that you shouldn’t have to bet on one. Choosing open-source AI models for schools versus commercial ones isn’t a single decision. It’s a set of everyday decisions about which model does which job, and the district should be the one making them.
Key takeaways
- No single AI model is best for every school task; districts get better results and lower costs by matching the model to the job.
- Open-source models running on a district-owned server have no per-user or per-token fee and keep prompts on district hardware.
- Commercial models connected by API are billed per token, paid by the district directly to the provider, and are not included in the server price.
- Even heavy commercial API use, priced at current published rates, typically costs far less than paying per-seat license fees for every staff member.
Open-source AI models for schools: what “local” really means
An open-source (or open-weight) model is one whose trained weights are published so anyone can download and run it. When a district runs one on its own private server, the prompts, the documents and the logs stay on district-owned hardware inside the district network. There’s no meter running.
That makes local models a great fit for the high-volume, everyday work: drafting a parent newsletter, rewording directions at three reading levels, summarizing meeting notes, or answering questions grounded in your own board policies and handbooks. These are tasks where “very good” is good enough and privacy matters a lot.
Two honest caveats. First, not every open model is built for a school server. That 2.8-trillion-parameter release is designed for data-center hardware, and “open” doesn’t mean “fits in your data closet.” Test which models run well on your server with your support team. Second, the very top commercial models still tend to lead on the hardest reasoning and writing tasks. That’s a reason to have both, not to rent everything.
When it’s worth paying for a commercial model
On a private server, the district can also connect top commercial models such as Claude, OpenAI and Gemini through an API and offer them side by side with local models, under the district’s own permissions and logs. The district sets up its own API account with each provider and pays that provider directly for what it uses. Those specific requests go to that provider under its API terms, so the district decides when, and for whom, a commercial model is switched on.
Here’s where tokens come in. A token is a small chunk of text, roughly three-quarters of an English word. Providers charge per million tokens the model reads (input) and writes (output). A typical teacher request, with a prompt, a pasted rubric and a reply, might use about 3,000 input tokens and 1,000 output tokens.
Using Anthropic’s published API prices and Google’s Gemini API price list as of today, here’s the arithmetic for that one request:
- Claude Sonnet 5.5 ($2 input / $10 output per million): 3,000 × $2 ÷ 1,000,000 = $0.006, plus 1,000 × $10 ÷ 1,000,000 = $0.010. Total: about 1.6 cents.
- Claude Haiku 5.5 ($0.10 / $0.50): $0.0003 + $0.0005 = about 0.08 cents.
- Gemini 3.1 Flash-Lite ($0.25 / $1.50): $0.00075 + $0.0015 = about 0.23 cents.
Now a heavy user: 20 requests a day for 20 school days is 400 requests a month. On Sonnet 5.5 that’s 400 × $0.016 = $6.40 a month. On Haiku 5.5 it’s about 32 cents. Compare that with Google AI Pro for Education at $20 per user per month on an annual plan, charged for every licensed adult whether they log in or not.
Let’s push the hypothetical. Say a district has 500 staff, and every single one is a heavy user sending every request to Sonnet 5.5. That’s 500 × $6.40 × 12 = $38,400 a year in tokens, compared with 500 × $20 × 12 = $120,000 a year in seats. If 70 percent of that work stays on free local models, the token bill drops to about $11,500. Real usage varies (long documents cost more; Claude Fable 5.2 lists at $8 / $40), which is why routing matters. You can play with your own numbers on our Education Token Price Comparison page, and I walked through the full seats-versus-use math in this earlier post.
To be fair, per-seat products bundle polished teacher features, support and training. That has value. But you’re paying for every seat, every year, on the vendor’s schedule and with the vendor’s model choices.
A simple model-routing plan for a district
Here’s how I’d set it up on a private server, using role-based access:
- Students: local open-source models only, inside teacher-built assistants. No commercial API access by default.
- Teachers: local models as the default for daily work, plus one lower-cost commercial model for tasks where quality clearly improves, like feedback on long student essays (with names removed).
- Leaders and specialists: access to a premium commercial model for board summaries, complex analysis and policy drafting, side by side with a local model so they can compare.
- Anything with sensitive student information: local models only. That one rule does more for privacy than any vendor promise.
When a better model comes out next month (and one will), IT adds it, tests it with a pilot group, and swaps it in. No new contract. No migration. That’s what owning the platform buys you, and it pairs naturally with the access and logging controls I covered in the November updates to the private server.
What district leaders can do this month
- List your top ten AI tasks. Sort them into “everyday” and “high-stakes.”
- Write a one-line data rule, such as: “Prompts with student information stay on local models.”
- Pull your current per-seat bill. Multiply seats by price by 12, and then check how many of those seats were actually used last month.
- Price your heaviest users in tokens. Use the arithmetic above with today’s published API prices, not last year’s.
- Add a model-review line to your AI policy. The EdTech article suggests reviewing the operational layer of AI policy at least yearly; with models changing this fast, I’d look quarterly.
- Ask any AI vendor: which models do we get, who decides when they change, and can we bring our own? My five questions to ask before you buy another AI tool are a good companion here.
Why a private server makes model choice yours
A private server isn’t one more chatbot. It’s the district’s own AI platform, where the district chooses the models, sets the rules and controls the cost. Through the program, the server is a one-time $20,000 purchase that includes pre-configured hardware, 12 months of remote support and a year of virtual professional development. There are no per-user licensing fees, and one server supports roughly 200 concurrent users. It can run in a secure cloud, but we recommend on-site: keeping the hardware in your own building is the strongest position for control and data security.
Just as important, it comes through LeadershipAIplus, a nonprofit, educator-led program run by former K-12 administrators. It’s educational leaders supporting educational leaders, not a vendor chasing seats, and districts from Kansas to Washington to Berlin are already running their own servers through it. Tricia and I talked about why leaders need sustained learning, not one-off workshops, in AI in Schools Is Changing: What Leaders Need Now.
FAQ
Can schools run open-source AI models on their own server?
Yes. A district-owned private server runs open-source models such as Llama and Mistral locally, so prompts and logs stay on district hardware with no per-user or per-token fee. Very large open models built for data centers won’t fit a school server, so test before you commit.
Are commercial API costs included in the private server price?
No. The district sets up its own API account with each provider and pays that provider directly for the tokens it uses. At current published prices, a heavy teacher’s monthly usage on a mid-tier model costs a few dollars, far less than a typical per-seat license.
The last word
The models will keep changing every few weeks. Your district’s ability to choose shouldn’t.
So here’s my question: if a better model came out tomorrow, how long would it take your district to put it in front of teachers, and who would decide?
If you want help mapping your tasks, users and costs to the right mix of models, schedule a Systems Review.
Sources
- Alexander Slagg, EdTech Magazine, “Writing and Updating K–12 AI Use Policies That Meet the Moment,” November 4, 2026. Link
- BenchLM Research, “LLM News, November 7, 2026: Kimi K3.1 Pro Weights, GPT-6 Cyber,” November 7, 2026. Link
- Anthropic, “Pricing” (Claude API documentation), accessed November 11, 2026. Link
- Google, “Gemini Developer API pricing,” accessed November 11, 2026. Link
- Google Workspace Admin Help, “Google AI Pro for Education,” last updated October 30, 2026. Link