Halv's context wrapper cuts Codex tokens 51% per fix

Halv published benchmark results on September 4 showing its context-compression wrapper beat vanilla Codex on a 20-pair SWE-rebench test. For anyone paying by the token, fewer tokens per correct fix means lower cost per bug closed, if the result holds beyond this one test.

Halv's benchmark writeup ran gpt-5.6-luna at medium reasoning through Codex 0.152.0, with identical model and settings on both arms. Vanilla Codex solved 7 of 20 task pairs. Halv's wrapped version solved 10.

Total token use fell 30.2%. Halv also solved three more tasks than vanilla Codex, so tokens per correct answer fell further, down 51.1%.

The wrapper compresses context and filters shell command output before it reaches the model. The idea is that most of what a coding agent reads back from grep or ls is noise it pays for and never uses.

Halv is upfront about the limits. The company calls twenty pairs directional evidence, not a proven number, and says it won't generalize across other repositories or other models without more runs.