用户分享多款 AI 模型的使用取舍，但型号与结论尚待核实

There is no verified data establishing that GPT-5.6 is more token efficient than Claude in general. The Artificial Analysis cost per task figures above are the checkable version of that claim, and they are the ones we would put in a budget. Verdict The honest question is not only which model is smarter, but what a finished task costs. Sol is cheaper per task and one index point behind. For high volume work that maths favours Sol. For the narrow band of jobs where the last point of capability decides whether the output is usable at all, it does not, and our recommendation below weighs that against availability and compliance, where the gap is still wide. [...] Claude Opus 528,703 Claude Fable 533,127 With one caveat that matters. Altman said 54% on agentic coding, and this figure is the general index. On Artificial Analysis’s coding agent runs the gap is 25%, not 54%, so the number that matches his claim turns up on the wrong benchmark. The cost per task above has moved too: a Sol t

But this does not mean Fable is obsolete. Claude Fable 5 still looks like the safer model-plus-harness combination when reliability matters most. In our Golden Eval, Fable completed all 47 tasks, while Sol missed two. If the workflow involves customer-facing updates, production issues, CRM changes, billing, or anything where a wrong action is expensive, I would still prefer Fable. For everything else, especially high-volume coding agents, research agents, data extraction, and parallel subagent workflows, Sol high is the better default. The small reliability gap is easier to absorb when retries are cheap, and the cost/token savings compound fast. So my recommendation is simple: [...] So, the trade-off is straightforward: Fable was more dependable, while Sol was faster and more token-efficient. For workflows where a single wrong action can be expensive, Fable’s reliability may justify the higher price. For high-volume workloads where occasional retries are acceptable, Sol’s efficiency

The cost strategy is simple: do not use Fable 5 for everything. Use it when the task is hard enough that fewer iterations, better planning, or a higher first-pass success rate pays for the higher token price. ### The Fable 5 data-retention caveat# Try it in practiceMake this section actionablePractice the workflow instead of only comparing tools.Practice this This is the biggest enterprise consideration. Claude Fable 5 is a covered model with a 30-day data-retention requirement. It is not available under zero data retention. If your organization requires ZDR, Opus 4.8 or Sonnet 5 may remain the better choice until policies catch up. For individual users and teams without strict retention constraints, this is usually acceptable. [...] Use another tool if You are looking for a model literally named Claude Opus 5 — it does not exist You need a confirmed future Opus 5 release date — Anthropic has not announced one Key takeaways Anthropic has not released or announced a model n

Cost depends heavily on usage volume and task complexity. GPT-5.6 Sol tends to be faster and can be cheaper per task at high throughput. Claude Fable 5’s more verbose responses can increase token costs, but for tasks where it catches issues that would otherwise require human review, the effective cost may be lower. Running both models through a platform that gives you visibility into token usage per step makes cost optimization much easier. ### Are these models reliable enough for autonomous agentic systems? [...] ## Other agents ship a demo. Remy ships an app. React + Tailwind ✓ LIVE API REST · typed contracts ✓ LIVE DATABASE real SQL, not mocked ✓ LIVE AUTH roles · sessions · tokens ✓ LIVE DEPLOY git-backed, live URL ✓ LIVE Real backend. Real database. Real auth. Real plumbing. Remy has it all. RemyThe world's most powerful product manager agentTry Remy today GPT-5.6 Sol handles ambiguity by resolving it quickly — sometimes too quickly, making confident assumptions where

# Thread 4.1K views mengqiutu's profile picture mengqiutu GPT 5.6 真的蠻香的 5.6 的分群變得和 Claude 很像 Haiku → Luna，Sonnet → Terra，Opus → Sol Sol MAX官方定位聽說比較接近 Fable，而不是 Opus 以價格來說，一般Sol比較像 Opus Sol 用起來有 Opus 那種穩定感 以最近 Opus 降智又降速的情況，我自己的體感反而比較接近 Fable，App 常常卡住，CLI 穩定多了 Fable 我只敢拿來派工，但 GPT-5.6 已經可以拿來開發 Luna 在快問快答的表現蠻穩定，拿來討論很好用 但開發就算了，交代的任務失敗率慘不忍睹，只適合做不需要思考的呼叫型任務 Web 介面我到現在還搞不懂預設到底是 5.5 還是 5.6，一整個很謎 Terra 不上不下 速度不如 Luna，表現又不如 Sol 就是 Luna 不夠用，花一點錢換速度和表現 Sol 的話，開發使用，有加錢有差 比起 5.5 容易過度思考，5.6 穩定多了 圖只是這次任務的單一結果，沒有要做模型評比，只是拿來建立一個大概的使用感覺，方便之後知道不同情境該選哪個模型 Translate 37 12 Log in or sign up for ThreadsSee what people are talking about and join the conversation. Log in with username instead

## Post ## Post user avatar user avatar user avatar user avatar user avatar Avatar See this post in the app Use the app to view all comments and discover more posts.