HN’s GPT-5.5 read: the real question is whether it finishes the job

HN cared about delegation, not just raw IQ

The Hacker News thread on GPT-5.5 moved fast because people were not reading it as another benchmark drop. The real question was whether this model changes what you can safely hand off on a computer. Comments kept circling back to the same practical test: can it plan, use tools, check its own work, and stay with a messy task long enough to finish it.

OpenAI’s release makes that exact case. GPT-5.5 is pitched as stronger at agentic coding, computer use, online research, document work, and long-running knowledge tasks. On the published evals it reaches 82.7% on Terminal-Bench 2.0, 78.7% on OSWorld-Verified, 84.9% on GDPval wins-or-ties, and 81.8% on CyberGym. Just as important, OpenAI says it matches GPT-5.4’s per-token latency while using fewer tokens on the same Codex work.

That mix of autonomy and efficiency is what gave the thread its energy. One HN commenter flagged that the rollout in ChatGPT and Codex was gradual, another pointed out that API access is still pending, and another immediately framed the model in cybersecurity terms rather than as a general-purpose toy. In other words, the community did not stop at bigger number, better model. It asked how quickly the model becomes operational.

• Does it keep going instead of stopping halfway through a multi-step task?
• Does the tool use look reliable enough for engineering work?
• Do rollout timing and API access delay real adoption?

That is why the HN response felt different from a normal launch thread. People were not mainly arguing about whether GPT-5.5 is impressive. They were testing whether it raises the trust ceiling for delegated computer work. If that answer keeps coming back yes, GPT-5.5 will matter less as a headline model and more as a new baseline for how much work users expect an AI system to carry on its own.