返回资讯中心
外部精选
软件工程
#开发者实时流

社区质疑 Fable 5.2 是否只是 Opus 5.5 的高 effort 设置版本

来源标题:Is Fable 5.2 just Opus 5.5 with a higher effort setting?

有讨论质疑 Fable 5.2 可能并非独立新模型,而只是 Opus 5.5 调高 effort 设置后的产物。目前尚无官方说明或基准对比证实这一猜测,相关讨论仍在进行中。

Hacker News 最新ianberdin2026/11/10 · 7 天前12 分钟阅读

以下正文同步自 Hacker News 最新,版权归原站所有,已转换为易读排版。

34 models249 runs$104.93 measuredupdated 2026-11-09

One API call per cell, first answer kept, providers' own billing. The MacBook is the test - everyone knows what one looks like, so a 5% error shows. The pelican is the control every model passes. Order is our quality rank, judged by eye; sort by newest or price instead.

The MacBook from medium to max for the three Fable models and Claude Opus 5.5. Each cell is the first run; a dash means that effort was not run.

Batch API rows: billed at half price, and the time includes Anthropic's queue.

The best laptop on this page, by our call: its max on the Batch API. The most detailed, high quality, zero errors, for 255,957 tokens. Streamed, max stops at the 128,000-token ceiling and returns an empty file. Its xhigh draws a product shot for $1.72.

Spent all 128,000 output tokens - the whole ceiling the streaming API allows this model - on thinking, and returned an empty file. Nothing arrived until the stream closed 20 minutes in. $2.56 for no drawing. Re-running it on the same API cannot help: 128,001 is rejected by name. The row below is the same request given 300,000 tokens on the Batch API, and it needed 255,957 of them.

Rank 1 since 2026-10-30, and that is our call rather than a measurement. Its xhigh cell matches the screen and port furniture of GPT-6 Astra's best drawing and adds the wordmark and a real screen glow, for $1.72 against $2.29; its Batch API max is the most complete laptop in this catalog; and in our view it is the best model overall, clearly ahead of Fable 5 and 5.1. Astra still wins the two things this page counts against it: legible legends on every key, and a ladder where all four cells land. Released 2026-09-21 and benchmarked the next day, and on that day it rearranged the top of this page without taking it. Its xhigh cell is a product shot: an Apple menu bar with named menus, status icons and the 9:41 clock, a dock of individually recognisable apps, a notch, side ports, speaker grilles, the MacBook Pro wordmark on the chin, a keyboard that recedes with the deck, and one vanishing point shared by lid and base. That is detail parity with the rank-1 Astra drawing - it even has the chin wordmark Astra omits - and it costs $1.72 against Astra's $2.29 and a third of the $4.95 Fable 5.1 cell it matches. Astra kept the top seat then on two things: legible legends on every key, and a ladder where all four cells land. Cheap is where Opus 5.5 is most convincing: high is $0.40 for a correctly assembled laptop with grilles, a real key grid and a trackpad - the cell Opus 5 gets faulted for, drawn right - at 21% less cost and 32% less time than Opus 5 spent on its near-identical version, which is Anthropic's claimed 30% speed-up showing up in an independent measurement. The ladder then breaks at the top, and expensively. At max it spent all 128,000 output tokens - its own ceiling, not our sweep budget - on thinking, emitted nothing until the stream closed, and billed $2.56 for an empty file. Fable 5.1 failed this way against a 64K budget and finished when given its real ceiling; Opus 5.5 had its real ceiling from the start, so there is no larger number to re-run it with. The pelican dial is ordered and cheap: $0.05 buys a complete bird and bike with no saddle and a stub of a chain, $0.13 buys a finished illustration - chain wrapped over both sprockets, cranks, legs on the pedals, motion lines, sun. A postscript on that empty max cell: it was never a runaway. Re-run on the Batch API, the one surface that allows this model more than 128,000 output tokens, the identical request finished naturally at 255,957 - and drew the most complete laptop in this catalog, menu bar and clock and dock and side ports and labelled keys and glow on the desk. The model is not failing the task at max effort; the streaming API simply cannot hold the answer. That drawing costs $2.56 at the Batch API's half price, arrives asynchronously hours later, and is therefore not something you can ask for in a chat - which is why it sits beside the empty cell rather than replacing it. We re-ran it on 2026-11-09 to check whether the model had changed (see Re-runs below). Max did exactly the same, to the token and the cent: all 128,000 spent thinking, $2.56, an empty file. On the Batch API, the September way, max drew the same kind of laptop again, at 243,707 tokens - 5% under September - for $2.44. Xhigh drew the same kind of laptop with 25% more tokens, $2.15 against $1.72. The pelican's max, stopped after nine minutes on launch day, finished this time at 101,695 tokens and 15.7 minutes for $2.03 - a road under hills and clouds, a dotted chain, a wing gripping the handlebar and both feet on the pedals.

In Playcode since 2026-09-22 · runs 2026-09-22, 2026-09-23

正文由 探How 从来源站点 RSS 同步,内容未经改写;遇到排版缺失或需要图片、视频时请以原文为准。

打开原文
去社区讨论 →