---
schema: 1
type: term
title: _LFGBench_
description: >-
  _LFGBench_ is a coding benchmark in which a model is given four tasks and must
  plan, build, and test in a loop until each is done, scored by the percentage
  of planned features it completes.
canonical: 'https://liquidgold.cc/channel/everyinc/explore/term/lfgbench-8f6f5e94eb3b'
markdown: 'https://liquidgold.cc/channel/everyinc/explore/term/lfgbench-8f6f5e94eb3b.md'
llms: null
parent: 'https://liquidgold.cc/channel/everyinc/llms.txt'
channel: everyinc
updated: '2026-10-23T22:12:35.346Z'
language: null
children: []
---


# \_LFGBench\_

Canonical: https://liquidgold.cc/channel/everyinc/explore/term/lfgbench-8f6f5e94eb3b

Parent: [Channel index](https://liquidgold.cc/channel/everyinc/llms.txt)

## Definition

_LFGBench_ is a coding benchmark in which a model is given four tasks and must plan, build, and test in a loop until each is done, scored by the percentage of planned features it completes.

## How the channel uses **_LFGBench_**_LFGBench_ is the hands-on coding benchmark the Every team runs to see how much of a planned build a model can actually carry through on its own — the plan → build → test loop described below.### What it measuresThe structure is simple and demanding:- A model is given **four tasks**.- For each, it must **plan**, **build**, and **test** in a loop **until done**.- The score is the share of planned features actually completed.This makes _LFGBench_ a test of _long-horizon follow-through_, not single-shot cleverness: can the model keep a whole feature set moving without abandoning it or drifting?### What it revealedIn the Opus 4.6 vs. Codex 5.3 head-to-head, the results split sharply:- **Opus 4.6 completed 100%** of planned features.- **Codex 5.3 completed around 64%**.The team's read: Codex thrives on precise specifications, while Opus is better at filling in the blanks — with the risk of occasionally doing the wrong thing confidently.In the same stream, Kieran Claassen described his personal **`/LFG` command**, which runs a similar loop; with Anthropic's agent swarms enabled, its runtime dropped from 20 minutes to 5 — though he noted swarms shine in implementation, where the plan is clear, less so in planning phases.### Key takeaways- _LFGBench_ = **plan → build → test in a loop until done, across four tasks**, scored by completed features.- It favors models with **long-horizon stamina** and willingness to fill in gaps.- Pair it with other checks — speed, variance, design, writing — since a high _LFGBench_ score alone doesn't capture a model's consistency.