Building a free educational simulation app with Gemini — how to handle high-quality generation without unsustainable API costs?

Community
, ,
1

Hi everyone,

I’m a student developer working on a project called CreateSim — a free educational app aimed at helping middle and high school students understand science concepts through interactive simulations.

The idea is:

  • User inputs a concept (e.g., osmosis, pendulum, reflection)

  • The system generates a smooth, interactive simulation + guided explanation

  • Focus is on visual learning and accessibility, not replacing textbooks but enhancing them

But, I’ve run into a major limitation with model capability vs cost:

  1. Model capability gap

    • Models like Gemini 2.5 are sufficient for basic structuring and visual guidance

    • But for actual simulation generation (logic, behaviour, realism), they’re not reliable enough

    • Higher reasoning models (Pro-level) perform significantly better

  2. Cost constraint

    • If I use my own API key with higher-end models, all generation costs fall on me

    • This is not sustainable as the app scales

  3. User-side API is not viable

    • Asking users (students/teachers) to enter their own API keys creates too much friction

    • Most users won’t know how to obtain or manage API access

  4. The manual library is not scalable

    • I’ve built a few simulations manually and integrated them

    • But creating a large library of concepts manually isn’t feasible, given time constraints

I want to keep CreateSim:

  • Completely free

  • Accessible to students regardless of background

  • Scalable without per-user high compute cost

I’d really appreciate guidance on:

  1. Are there any Google programs/grants/credits for student or social-impact AI projects like this?

  2. Best architectural approach:

    • Is a hybrid system (prebuilt simulation engines + lighter model mapping) the right direction?

    • Any recommended patterns for reducing reliance on high-cost generation?

  3. Efficient use of higher-end models:

    • How can I limit the usage of Pro-level models while still maintaining quality?

    • Caching/reuse strategies that work well in practice?

  4. Any examples or similar projects that have solved this kind of problem?

This is a non-commercial, student-led project focused on education accessibility, so I’m trying to design this in a way that avoids passing cost to users while still delivering a high-quality experience.

Thanks in advance — even high-level guidance or pointers would really help.