GPT-6 Astra

User discussion (112 HackerNews comments; top 25 by upvotes shown) centers less on Astra itself — access is still limited at launch — and more on whether the GPT-6 line justifies OpenAI's 'world's most intelligent' framing quoted in coverage. Several high-upvote threads reject the AGI label outright: one user writing a 30-page paper with a rival model reports passages that are 'plain wrong or weirdly out of place' with no self-awareness of the flaws, arguing a human researcher 'can easily outclass it' in writing and problem understanding. A detailed critique calls the ARC-AGI-3 scorecard 'extremely misleading': GPT-5.6 Sol is shown at 7.8% while the same source estimates ~30% had the responses-API harness (used for Astra) been applied. Practical coder feedback on OpenAI models: Codex 'follows direction slavishly,' is 'too literal,' restricts context-window sizes so compaction forces repeated grounding work, and still lacks Claude Code features like defined plugin subagents. Over-engineering is a repeated complaint — an overnight run on a ~1000-LOC Python script produced over-architected output — with users sharing mitigation phrases ('do not overengineer and do not gold plate,' 'avoid bike shedding') and noting a training bias toward provenance programming in CI workflows. Efficiency reports favor OpenAI: one user burned ~2.5B tokens across 74 subagent tasks within a single $200 20x Pro week on 5.6 Sol Ultra, versus $1000 in credit plus an exhausted $200 monthly plan on Fable 5 Max for four smaller projects. Supervision is described as mandatory: 'like Tesla FSD; it kind of works but you have to be very vigilant' watching thinking traces to catch derailment. Several users say even Sol's capability ceiling is 'not enough for many tasks' and are eager for Astra. Net sentiment: respectful of raw capability, skeptical of launch marketing, frustrated by literalness and context limits.

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