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.
An unverified guidelines score The public Convex With Guidelines query did not return scores. The rules can still be installed. Do not write a change figure.