Skip to main content
Version: Next (unreleased)

Choosing a model and managing cost

GrillMyCode doesn't include its own AI. It uses an AI model of your choice through OpenRouter, which gives you access to models from Google, Anthropic, DeepSeek and many others with one account. You pay OpenRouter directly, from a prepaid balance.

The default is a good choice

GrillMyCode uses Google Gemini 3.5 Flash Lite unless you pick something else. It's fast and cheap, and in testing it wrote the best wrong answers for the multiple-choice quizzes, which is harder than it sounds. For most courses there's no reason to change it.

What it costs

With the default model and the other recommended ones, an assessment usually costs less than one cent. A $5 US balance typically lasts a large class for a whole semester.

What adds to the cost:

  • How often it runs. With the every push trigger, a student who pushes 20 times gets 20 assessments. A submission tag runs once per submission instead.
  • The private answer key. When it's set up, the AI also writes three wrong answers per question for the LMS quiz, which makes each assessment somewhat longer.
  • The model. The more capable models can cost 10 to 100 times as much per assessment.

To keep an eye on spending, check the activity page in your OpenRouter account. You can also give the key its own spending limit.

These have all been tested with GrillMyCode and usually cost under a cent per assessment. They're the choices in the Workflow Wizard's AI step.

ModelGood to know
Google Gemini 3.5 Flash Lite (default)Fast and cheap, with the best multiple-choice wrong answers of those tested
DeepSeek V4 FlashVery cheap, and a reliable alternative
Minimax M2.7Very cheap, and does well with little tuning
StepFun Step 3.7 FlashGood quality for the price
Tencent Hy3Cheap and works, though its writing style varies more
Xiaomi MiMo V2.5 ProGood at following the required question format

Using a more capable model

For advanced courses, a more capable model may ask sharper questions about complex code. In the Wizard, choose Own Choice and enter any model from OpenRouter's catalogue.

Check the price on OpenRouter first, and multiply it by your class size and how often the workflow runs.

Faster or cheaper, same model

Most models are offered by several companies at different speeds and prices, and OpenRouter normally picks one for you. The Wizard's Model routing setting lets you choose what matters more:

  • Balanced (recommended): let OpenRouter choose.
  • Speed: try the fastest providers first. This is worth it if you like the questions but they take too long to arrive, for example when a whole class submits at once. It can cost more, so check the price first.
  • Lowest cost: try the cheapest providers first. They can be slower, or queue at busy times.

The model is the same either way, so the questions are just as good.

When a whole class submits at once

Everyone shares your one OpenRouter key, so a class submitting in the same few minutes shares its limits too. GrillMyCode automatically waits and retries when OpenRouter is busy. If runs still fail with rate-limit errors, check that your OpenRouter balance is above zero; accounts with no credit are limited far more strictly.


Go deeper: OpenRouter: model IDs, routing variants and every OpenRouter-related setting · Recipe: A more capable model