GPT-5 mini

303 HackerNews comments (top 25 by upvotes shown) — a developer-heavy, benchmark-literate crowd. Dominant thread: OpenAI emphasized GPT-5's Tau²-bench Telecom gain while other domains were 'overlooked'; a self-identified OpenAI employee defends the Telecom emphasis, but peers remain skeptical ('a huge bump like this is not expected — I would expect OpenAI to opt…' / 'I don't trust these benchmarks fully'). Second thread: a tester boosted scores by having Claude rewrite the benchmark's agent policies into clearer structure (decision trees, explicit tool-call parameters, binary conditions). Peers push back hard: rewriting 'negates some of the efficiency and latency benefits of using mini', risks 'teaching to the test', may leak task hints, and 'will not work well for other cases like medical or social advice'. Open question raised: is gpt-5-mini smart enough to rewrite prompts itself, or would complex policies cause 'cognitive overload'? Separately, one user running tau2 on gpt-5 vs 4.1/o3 praises gpt-5 as 'really good at looking at tool results and interleaving those with thinking… too good at figuring out the right tool to use in one go' — the clearest capability positive, but attributed to full GPT-5, not mini. Surrounding comments debate coding agents generally (expected value of hands-off agents, Rust onboarding friction, 'Git mastery now is about concepts'), plus one user who bought a Max subscription after a health scare — context on how devs pay for models, not mini-specific. Net: users see real family-level capability gains but treat mini's headline numbers as prompt-engineering-dependent and possibly benchmark-contaminated; mini's core appeal is cost/latency for continuous interaction.

No aggregate ratings were found for this product during the last harvest.