Compacting earlier cuts agent costs 16-23%; GLM 5.3 costs 30x less than Grok 4.7
yb2698 · x · 2026-11-02
evalstate benchmarked compaction thresholds in fast-agent on a Terminal-Bench 4.0 subset (19 tasks × 5, 95 trials):
- GLM-5.3-Flash: compacting at 200k vs no compaction (1M window) cut total cost from $26.15 to $20.06 (-23%) and agent time from 73.8h to 56.1h (-24%), with score +2.1 ± 6.6 pts (within error); $0.21 vs $0.28 per trial
- Grok 4.7: compacting at 190k vs the default 425k cut cost from $721 to $605 (-16%) with score +2.1 ± 7.2; $6.37 vs $7.59 per trial
Takeaway: context size before compaction is a big lever in harness optimization. Bonus: GLM 5.3 costs 30x less than Grok 4.7 for the same work.
More from coding & agent
- When Codex writes 15 components on single lines: SonarQube flags agent-built code debt — bibryam · 2026-11-02
- Convergence theory shows AI agent teams can collaborate: 2x faster routing, 30% time cut — peter_richtarik · 2026-11-02
- Andrew Ng: 100% of my tasks now run through AI agents — 2C_ornot2C · 2026-11-02
- kamchatka: a Rust terminal agent with no system prompt that edits its own context — yb2698 · 2026-11-02
- As Grok bot limits tighten, open-source self-hosted agent open-duet hits 6.5k GitHub stars — matchaman11 · 2026-11-02
- Solo Dev Codes Entire Fishing Game With Opus in 3 Months, Demo Out Free — RUSuper · 2026-11-02