GPT-6 Sol & Luna

HackerNews discussion (top-voted comments) centers on four themes. (1) Naming confusion: 'what the hell do Terra, Sol, and Astra really mean? Which one of them is the effective successor to gpt-5.4-mini?' — tier names read as marketing, not signal. (2) Practical tiering: one user compares model sizes to delegating chores — smaller models (Nano/Luna-class) 'won't reliably handle complex open-ended tasks' and need 'very explicit instruction,' so cheap models only pay off when the decision tree is clear. A 5.6 Sol user reports a reasoning 'dead zone': medium 'under engineered and took short cuts,' high 'over engineered and ignored instructions.' (3) Plan economics: OpenAI's metering gets rare praise — 'statistically verified that OpenAI Pro 20x = 4 * Pro 5x = 20 * Plus, exactly as advertised,' with no Anthropic-style separate meter ('five Fable sessions per week') — but a rebuttal claims 'since around July... Anthropic is actually giving more on $200' measured in raw API costs, and local-model alternatives require $6k–12k hardware (Mac Ultra 5, dual DGX Sparks). (4) Benchmark posture: users appreciate that the announcement admits GPT-6 Sol only *matches* Claude Fable 5.1 xhigh on FrontierCode at lower cost (a sign of 'skilled engineers' who 'recognize slop'), yet others distrust benchmarks outright: 'I only trust what a model does with MY code.' Additional friction: one commenter frames Luna's pricing as 'about suppressing inexpensive Chinese model competition.' A ProductHunt comment relays the official pitch: Sol 'beats Claude Opus 5 on Automation Bench at roughly 9% of the cost per task' and matches Fable 5.1 on DeepSWE at ~80% lower cost, carrying over Astra's alignment gains including 'fewer misleading claims about their own coding work.' Company-ethics sentiment is mixed (copyright training data, differential-privacy debate), but several users say delivered value 'far outweighed' the concerns.

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