Five Qwen3.8 Agents Beat Claude Fable 5 on LiveCodeBench

A multi-agent workflow running five Qwen3.8-27B instances has outperformed Claude Fable 5 on LiveCodeBench Hard coding problems, scoring a mean 9.8 out of 10 compared to Fable 5's solo score of 8.0. The Memdoor implementation reproduces the GVS5H harness architecture, assigning specialized roles to each agent and restricting communication to files written to disk, with no shared chat context between models.

Architecture and Role Assignment

The Memdoor workflow deploys five Qwen3.8-27B agents in distinct roles: a planner, two parallel ideators, a solver, and a manager, plus a repair worker that patches failed solutions. Each agent starts from a fresh context and reads only the specific files it needs for its task. The planner writes a plan file, the solver generates solution.py, the manager analyzes test logs, and the repair worker applies fixes based on error output.

This architecture mirrors the GVS5H harness published by Lee in 2026, reusing the exact prompts from that research. The implementation test validates whether file-based agent coordination can match or exceed single-model performance without shared memory or context windows.

Benchmark Results and Token Economics

Over five passes on the first ten LiveCodeBench Hard problems, the five-agent system achieved a 9.8 mean score, edging past the GVS5H harness's 9.6 and significantly beating Claude Fable 5's 8.0. Across 50 total runs, the team solved 49 of 50 problems, consuming 15.3 million output tokens at a total cost of $35.79. Including 16 incomplete attempts, the full cost reached $38.77.

A median run took 40 minutes and cost $0.39 on Qwen3.8-27B with a 128K reply cap. This represents a fraction of the pricing for equivalent Fable 5 inference, though direct cost comparisons depend on API rate cards and context window usage patterns.

• Mean score: 9.8/10 on LiveCodeBench Hard (first 10 problems)
• Total runs: 50 successful, 16 incomplete
• Token usage: 15.3M output tokens
• Cost: $35.79 for successful runs, $38.77 including failures
• Model: Qwen3.8-27B with 128K context limit

Contamination and Community Scrutiny

The authors flagged potential contamination up front, noting that the LiveCodeBench Hard problems date from March 29 to April 6, 2025, and that Qwen3.8-27B may have seen these problems in training data. The research team stated that contamination cannot be ruled out given the model's training cutoff and the contest dates.

Discussion on Hacker News raised additional questions about whether the file-based communication pattern inadvertently leaks intermediate solutions between agents, effectively giving later agents access to partial work from earlier steps. The workflow's design assumes agents remain isolated except for structured file artifacts, but critics argue that rich error logs and repair-worker patches may encode more information than intended.

Release Date and Open-Weights Context

Officially released on November 1, 2026, the Memdoor GVS5H implementation arrives amid a surge of open-weights Qwen experiments. Recent projects include Lyra's Evil-Qwen jailbreak demonstration and the Nesformer NES-in-Qwen embedding, both published within days of this release. The clustering suggests growing interest in pushing Qwen models beyond standard single-agent inference into specialized architectures and unconventional use cases.

What This Means

The Memdoor workflow demonstrates that coordinated teams of smaller open-weights models can match or exceed frontier proprietary models on structured reasoning tasks, at dramatically lower inference costs. The 9.8 score on LiveCodeBench Hard validates the GVS5H harness architecture and shows that file-based agent coordination scales without requiring shared context windows. However, contamination risks and information-leakage concerns leave open questions about how these results generalize to unseen problems and whether the cost advantage holds at scale across diverse coding benchmarks.