Skip to content
LEAPDocs
2026.08

Models & providers

Which model does what, and why the platform does not use one model for everything.

Four tiers

Models are selected per job. Curriculum structure, bulk content, vision and rubric grading are different problems with different cost and quality profiles.

SmartUsed for: Syllabus parsing, structure extractionWhy this tier: Hardest reasoning in the pipelineRuns once per course. Gets the largest token budget because the unified extraction payload — gates plus every lesson with examples and exercises — must not truncate mid-structure.
FastUsed for: Lessons, scripts, questions, grading rationaleWhy this tier: Volume work with strict output shapeRuns many times per course. Chosen for reliable structured output and predictable latency rather than raw capability.
CheapUsed for: Vision: PDF and image parsing, photo gradingWhy this tier: Reading pixels, not reasoningExtracting text from a scan does not need a frontier model.
GradingUsed for: Rubric grading of long handwritten answersWhy this tier: Band judgement is its own skillSelected specifically for rubric-band quality on answer books, and given a large budget because it emits a verbatim detected answer plus reasoning per criterion.

How models are reached

All model traffic routes through a single provider gateway, which means a model can be swapped without touching application code. Model choice is configuration, and it has changed several times as the trade-offs shifted.

Why choices change

A worked example

The content tier was originally a mid-cost model, moved to a much cheaper one for obvious reasons, then moved back. The cheaper model was around fourteen times cheaper on input but produced quiz generation at a median of 73 seconds with a long tail — and one truncated JSON response forced a whole generation to retry. Cost per token is not cost per completed course.

Structured output

Every generation call expects a strict shape and is validated against it. A malformed or truncated response fails the job rather than being partially written — which is why a model’s reliability at emitting valid structure matters more here than its benchmark scores.