Forge benchmark: Codex solves 37/50 on GPT-6 Astra, Forge 35/50 at 36% lower cost

ahmet cetinkaya @cettocdxOriginally posted 2026.10.18

Open original on X

We benchmarked Forge against Claude Code and Codex using the same model in each matchup. Claude Opus 5.5, 290 real GitHub issues: Forge solved 268, Claude Code 262 — at a 32% lower cost per resolved issue. GPT-6 Astra, 50 issues: Codex solved 37, Forge 35 — with Forge costing 36% less. All data is public: http://github.com/cogniphic/forge-bench⁠