Octomind benchmark: Codex beats Claude Code on cost and speed, trails on solve rate

#What happened

Octomind published a benchmark that ran OpenAI Codex and Anthropic Claude Code, each with stock settings, on the same 25 tasks taken from real merged pull requests. The tasks span Python, PHP, Rust, C++, and JavaScript, and pass or fail was decided by each project's own hidden tests.

Codex, using gpt-5.6-sol, solved 21 of the 25 tasks for $14.86 in about an hour. Claude Code, using claude-opus-5, solved 23 of 25 for $81.79 in 6.7 hours. That works out to $0.71 per solved task for Codex versus $3.56 for Claude Code.

The two-solve gap came from thoroughness traps: both agents failed the same two hardest cases, and Claude Code's two extra solves were ones where Codex had attempted a fix but its fix failed. On tasks both agents passed, quality was close, with average judge scores of 93.0 for Codex and 92.1 for Claude Code.

The piece also compares the plans behind each tool. Both start at $20 per month: ChatGPT Plus for Codex and Claude Pro for Claude Code. Heavier tiers run ChatGPT Pro at $100, $200, and $500 versus Claude Max 5x and 20x at $100 and $200. Codex ships its OS sandbox on by default while Claude Code has it off by default, using a classifier to review actions instead. Codex CLI is open under the Apache-2.0 license; Claude Code is proprietary. It also notes that GPT-5.5 leaves Codex subscription use on Oct 14, with users urged to move to gpt-6-sol or gpt-6.1-sol.

#Why it matters

This is the first apples-to-apples measurement of the two leading agentic coding harnesses, and the arithmetic favors a cascade strategy: run Codex first and escalate only its failures to Claude Code. In Octomind's numbers that reproduces Claude Code's 23 solves for $26.89, about a third of Claude Code's $81.79 bill, in 93 minutes. The rule they propose: escalate by detectability, not difficulty. Send Codex anything a test suite or reviewer will check before merge; give Claude Code the first pass where nothing checks the work.

Verified: 2026-10-16.