15M-param agent built in 3 weeks beats Craftax boss 44% of time at 1 cent; GPT-6 wins 1% for $83
hyperparticle · x · 2026-10-08
- Responding to experts who doubted a small neural net could beat Craftax without an LLM's knowledge, the team says their agent meta-harness actually finished the game in a few days, spending the rest on ablations to boost the score and to make the harness discover ideas faster.
- The resulting 15M-param agent beats the final boss 44% of the time, costing about a cent and 10 seconds per run.
- By contrast, GPT-6 Astra won only 1%, taking 24 hours and $83.
Related event: 15M-Parameter Tiny Model Beats GPT-6 on Craftax for One Cent(2 posts)→
More from coding & agent
- mitchellh Calls His OpenAI Dot to Move Benchmarks Across Machines via Tailscale — charliermarsh · 2026-10-08
- A Developer's Map of the Data Tooling Landscape: Ingestion, Storage, Modeling, Governance — bibryam · 2026-10-08
- Operator-1 launches as autonomous document extraction agent for long tables and sprawling PDFs — ycombinator · 2026-10-08
- good-css: 47 modern CSS techniques that make coding agents actually use :has() and subgrid — mattpocockuk · 2026-10-08
- Microsoft launches Azure canvases: an interactive shared workspace for Copilot agents — adnan_hashmi · 2026-10-08
- ChatGPT Work mode vs Codex: same quota, far more tasks done per 5-hour window — sasik520 · 2026-10-08