Build a full-stack SaaS with this set
The model comes from Terminal-Bench on the CLI scenario, chosen by run cost. It is not a second scenario board. The Next.js, Supabase, Convex, and ReactBench evals stay separate and are not merged into one score.
Edited 2026-10-10 · Model basis Terminal-Bench 4.0 2026-10-06
RolePrimaryAlternative · when to switch
ModelGPT-6.1 SolTerminal-Bench #3 · 58.2when you ignore cost and read only the pass rateApp frameworkNext.js + Next.js AGENTS.mdLibrary + AGENTS.mdThe UI layer is React. ReactBench is an ecosystem eval, not an official boardDataSupabase + Supabase skillsLibrary · MITIf you switch to Convex, read its own no-guidelines board. Do not add it to the Supabase pass rate
Take this set
AGENTS.md
# Full stack - App framework: Next.js. The agent reads the repo's Next.js AGENTS.md and does not write the App Router from memory. - Data: Supabase. Load the official skills when needed. Baseline ability is the no-skills eval. - Do not add Next.js, Supabase, Convex, and ReactBench pass rates together.
Install and use
npx create-next-app@latest npm i @supabase/supabase-js
Not for
Interface only Page quality belongs on the web frontend set.
Adding the evals Each framework eval measures a model together with that framework. The samples are small and the front is near the ceiling, so they do not become one ranking.
Treat with-guidelines as the baseline With Guidelines is another arm. Rank still follows no-guidelines; the change lives on the Convex AI rules page.