Reverse engineering a game binary with no symbols. Same model, 10 checks, one run each.
Codex CLI: 8/10 using 7.72M input tokens
nexo: 8/10 using 3.96M input tokens
Same score, half the tokens. Open source, works with any provider.
https://t.co/UW6taboBTp
Limits: one run per tool, a demonstration not a statistic. I wrote the task and the check. Records are public. Independent project, not affiliated with OpenAI.
Repo: https://t.co/UW6taboBTp (results in docs/bench)
Same model, same long task, 10 turns.
Codex CLI compacted its context twice, lost 5 values I gave it early, and failed.
nexo kept them and solved it: 27 calls vs 42, no compaction.
Open source, Rust, any model. One run per tool; limits below 🧵
Stop burning tokens. 🔥
Your coding agent re-sends its ENTIRE history on every call. By request 80: 60k tokens.
nexo curates it → 30k. Half the tokens per session.
Never /compact again. Run long sessions, cache kept.
Open source · Rust · any LLM
https://t.co/UW6taboBTp