GPT-5.5 in Hermes Agent: data cleanup and code review

This first look tests GPT-5.5 in Hermes Agent through a Codex account. The speaker reviews OpenAI's benchmark and efficiency claims before trying two tasks. He reports a 400,000-token context window in Codex and describes a planned API offering with a one-million-token window. These are launch details presented in the video, rather than independently verified specifications.

The first task combines CSVs, screenshots and PDFs from the speaker's 2024 DLM liquidity pool farming. He asks for an executive memo and supporting spreadsheet, with key numbers checked, conflicts recorded and missing values left uninvented. He reports that the AI agent finishes in a few minutes without follow-up prompts. His initial review finds no glaring errors, though this is not a full accuracy audit.

The second task reviews his TurboQuant PyTorch repository. The speaker reports that GPT-5.5 reproduces the import failure in issue 17, identifies a packaging problem and proposes restructuring the repository with a pyproject.toml file. After a further instruction, it implements the patch and opens PR 19. A subsequent check says the PR fixes the package import bug but does not resolve every open issue.

The comparison with Opus 4.7 rests partly on an earlier spreadsheet screenshot that the speaker says Opus could not read. He treats the results as preliminary and leaves the PR for closer review before merging.