GPT-6 Sol 与 Luna 上手评估：更便宜但没更聪明
Sol 事实错误率从 8.4% 降到 4.5%、价格减半，但智力指数仅 47→48；Luna 同样降价而能力持平，作者给出该换与该跳过的情况。
‣ out of the three models that dropped yesterday, you only need one. 1 — GPT-6 Sol - what OpenAI said it’s for everyday coding, agents, professional work, it's not the flagship, Astra is. - price vs last Sol cheaper by half - better than last Sol? yes on the jobs OpenAI said it's for, it's not smarter. - how i know 1 - factual errors : 8.4% → 4.5%. almost half 2 - coding and agent scores went up a bit. 3 - independent intelligence is mehh - 47 → 48. - rank in its field mid-tier, it is the cheapest coder in OpenAI’s stack but not the best. - worth switching from 5.6 Sol? yes . same intelligence and fewer mistakes for half the price, gud tek. - when to skip it hard coding, long reports, or “get this right first time.” (one shot ting complexing stuff), just use Opus 5.5 or Astra for those ones instead. - models still above it - GPT-6 Astra - OpenAI’s real top model and it beats Sol on almost every hard test and costs 5x. - Claude Opus 5.5 - actually smarter. AA Index ranks it 58 vs Sol’s 48. - Claude Fable 5.1 - slight edge on a few coding tests (DeepSWE ranks it 69.9 vs Sol's 68.8). not that superior 2 — GPT-6 Luna - what OpenAI said it’s for high volume work. summaries, extraction, quick questions. not for hard coding. - price vs last Luna cheaper by half. $0.20 / $1.20 → $0.10 / $0.50 - better than last Luna? no. same level of intelligence but cheaper. - how i know 1 - OpenAI says AutomationBench moved it up 5.4 points and the task costs less 2 - independent intelligence is flat. 37.26 → 37.32. 3 - knowledge-work scores went down . coding index 43 → 41 - rank in its field cheapest seat in the OpenAI stack. good for volume. not a coding model. - worth switching from 5.6 Luna? yes, only because it's cheaper, it doesn't have better output. - when to skip it coding, reports, or for one-shotting things. use Sol or Opus 5.5. - models still above it - GPT-6 Sol — same family, built for coding and agents - Claude Opus 5.5 — a totally different league - GPT-6 Astra / Claude Fable 5.1 — a different league too, don’t compare Luna to these unless you like getting embarrassed 3 — Claude Opus 5.5 - what Anthropic said it’s for long agentic coding, computer use, knowledge work. Fable-level quality at a lower price. - price vs last Opus cheaper. $5 / $25 → $4 / $20. - better than last Opus? yes. this is the only real upgrade out of these three i'm reviewing - how i know 1 - Terminal-Bench 4.0 - 52.3% → 66.4% 2 - Cursor Bench - 46.6% → 57.8%. Automation Bench - 26.9% → 40.0% 3 - independent AA Index has it at 58. number one right now 4 - Sonar - almost the same pass rate as Opus 5, but 27% less code and 40% fewer tokens - rank in its field number 1 for coding and long work but not the cheapest - worth switching from Opus 5? yes. better and cheaper. no reason to stay on 5. - when to skip it cheap volume work, just use Luna. if you have a tight budget, use Sol. but if you have stupid money and want to go all at anything, use Fable 5.1 or Astra. - models still above it - GPT-6 Astra — wins in some areas. science terminal bench ranks Astra at 64.6 vs 58.7 of Opus 5.5. - Claude Fable 5.1 — much more ahead. - GPT-6 Sol — not smarter, but it's faster and cheaper.