A Lisp coding agent is beating Codex in token efficiency by 10x

An X user posting as Tech Friend AJ reported that the open-source Lisp-based coding agent NODECODE used 7–14x fewer input tokens and finished 2.5x faster than Codex CLI on an identical cloud-VM task: building a playable 3D pelican kart racer with up to 60 minutes to improve it. On GPT-6.1 Terra, Codex took 9:56 and 3.29M input / 31.6K output tokens versus NODECODE's 3:56 and 226K / 13.0K; on GPT-6 Luna, Codex took 20:29 and 7.62M / 74.0K versus NODECODE's 8:24 and 1.07M / 23.8K. All four games loaded, drove and passed the tester's check, though NODECODE stopped earlier and produced plainer games, with one run per configuration.

Tech Friend AJ on X: "Tested NODECODE against Codex CLI. Same model, same prompt, identical cloud VMs: build a playable 3D pelican kart racer, with up to 60 minutes to keep improving it. GPT-6.1 Terra: • Codex: 9:56, 3.29M input / 31.6K output tokens • NODECODE: 3:56, 226K / 13.0K GPT-6 Luna: • Codex: 20:29, 7.62M / 74.0K • NODECODE: 8:24, 1.07M / 23.8K All 4 games load, drive and pass my check. NODECODE used 7–14x fewer input tokens and finished 2.5x faster, but it also stopped earlier and its games are plainer. One run each." Tested NODECODE against Codex CLI. Same model, same prompt, identical cloud VMs: build a playable 3D pelican kart racer, with up to 60 minutes to keep improving it. GPT-6.1 Terra: • Codex: 9:56, 3.29M input / 31.6K output tokens • NODECODE: 3:56, 226K / 13.0K GPT-6 Luna: • Codex: 20:29, 7.62M / 74.0K • NODECODE: 8:24, 1.07M / 23.8K All 4 games load, drive and pass my check. NODECODE used 7–14x fewer input tokens and finished 2.5x faster, but it also stopped earlier and its games are plainer. One run each. Introducing NODECODE: an open-source agent that rewrites itself as it works. In LISP; the AI-native language. McCarthy imagined a program you’d improve “by making statements to it.” Now you can, it’s yours and it will evolve over time. MIT. Works in Discord, TG, etc. Built for Tested NODECODE against Codex CLI. Same model, same prompt, identical cloud VMs: build a playable 3D pelican kart racer, with up to 60 minutes to keep improving it. GPT-6.1 Terra: • Codex: 9:56, 3.29M input / 31.6K output tokens • NODECODE: 3:56, 226K / 13.0K GPT-6 Luna: • Codex: 20:29, 7.62M / 74.0K • NODECODE: 8:24, 1.07M / 23.8K All 4 games load, drive and pass my check. NODECODE used 7–14x fewer input tokens and finished 2.5x faster, but it also stopped earlier and its games are plainer. One run each. Introducing NODECODE: an open-source agent that rewrites itself as it works. In LISP; the AI-native language. McCarthy imagined a program you’d improve “by making statements to it.” Now you can, it’s yours and it will evolve over time. MIT. Works in Discord, TG, etc. Built for