A line on Hacker News on Saturday evening makes you look twice: a team of five small models beats Claude Fable 5 on LiveCodeBench Hard. 433 points, and the numbers hold up. The repository itself tells a more careful story than the headline does.
The setup is called team-solve, a workflow for Memdoor, a CLI tool for agent pipelines.
Six roles work on one programming problem: a planner, two ideators, a solver, a manager and
a repair worker. Every role starts from a fresh context, they talk only through files on
disk, and whether a run succeeded is decided by a script that executes the sample tests.
All of it lives in eight YAML files, with no code and no model named. The method comes from
a Lee paper published this year.
The numbers, and what they cover
The run used the same model as the original authors, Qwen3.8-27B, with their prompts, their loop and a 128K reply cap. The basis is the first ten problems of the authors’ LiveCodeBench Hard list, five passes each, graded against every hidden test:
| mean solved (out of 10) | |
|---|---|
| this workflow | 9.8 |
| authors’ harness | 9.6 |
| Claude Fable 5, one call | 8.0 |
All 50 runs are in the repository, each with its requests, replies, program, grader output and cost. A median of 40 minutes and $0.39 per run, $35.79 for all 50.
The caveats are in the repository, not in the headline
That is the part I like most about this project. Ten problems against the authors’ 100, and 9.8 versus 9.6 sits inside the noise, so the claim is parity with the authors’ harness, not superiority. The contests date from spring 2025, and contamination is ruled out for no column. Fable 5 ran one call per problem, taken from the authors’ files, not inside this harness. Across the authors’ full 100 problems it is 92.4 against 90.4.
Even the one lost problem gets explained: the manager invented a wrong counterexample, and the team rewrote a program that had been correct. Twice elsewhere a correct program was overwritten and then recovered. Both are gaps in the authors’ loop, and they were left in on purpose.
This compares a team against a single call
Which is why “beats Fable 5” is the wrong reading. The comparison is 40 minutes of work by six roles against one request to a frontier model. That the pipeline wins is not surprising, it is the actual finding: discipline in the loop, meaning fresh context, file-only communication and a script as referee, pulls more out of a 27B model than a direct call to a large one gives you.
The number gets interesting the moment someone runs the same loop with Fable 5 inside it. Until then the comparison shows what a harness is worth, not which model reasons better.