Anthropic 于 9 月 28 日推出 Claude Sonnet 5.5，OpenAI 于 9 月 29 日推出 GPT 6.1 Sol。巧合的是，两款模型在 OpenRouter 上的 API 标价完全相同：每百万输入 Token 2 美元，输出 10 美元。 雷峰网(公众号：雷峰网)两家公司给出的方向高度相似：把接近最高档模型的能力，压进更快、更便宜、适合 Agent 与专业工作的产品层。Sonnet 5.5 强调以更少步骤完成边界清晰的日常任务；GPT 6.1 Sol 则瞄准复杂编码、计算机操作和专业工作。我们把两款模型放进同一个工作台，通过同一个渠道商调用，要求模型连续完成 3D 密室开发、最新政策调查和诺奖化学科普三项任务。最终，GPT 6.1 Sol 花了 38.5 美元，Claude Sonnet 5.5 只用了 10.9 美元，相差约 3.5 倍。速度上的差异更有意思。开发 3D 密室时，Sol 用了 33 分钟，Sonnet 却花了 74 分钟；到了政策调查，情况完全反转，Sonnet 仅用 5 分钟就完成报告，Sol 则花了将近一小时。除了最终交付成果，我们还对比了端到端耗时、工具调用、原件归档、运行验证和实际账单。结果没有形成一条整齐的胜负线，GPT 6.1 Sol 在密室开发中更快，并在两项研究任务里建立了更深的证据链和发布闭环；Claude Sonnet 5.5 则以更短路径交付可用成果，两项研究任务均快出数倍，三项任务总成本也只有 Sol 的约 28%。这是一种典型的“锯齿状智能”：Sol 是一个愿意继续核验的深度工作狂，Sonnet 5.5 更像高效的日常先锋官。通过详细分析发现，模型只能决定能力上限，工具链、验证方式和停止策略才是决定我们最终得到什么，又为此付出多少的核心。这两款产品也折射出中段模型的行业位置正在上升：开始成为企业 AI 的主力执行层，承接大部分复杂而高频的生产任务。当模型能力满足使用门槛后，稳定交付和单位任务成本将比单项榜单领先更能决定模型的产业价值。本文选择近期发布的两款顶级模型来进行对比测试：分别来自 OpenAI 和 Anthropic 的 GPT 6.1 Sol 和 Claude Sonnet 5.5，我们使用 Claude Code Cli 作为测试平台，调用来自 OpenRouter 渠道商的两款模型分别完成三个复杂任务。两款模型均可搜索网页、读写文件、运行命令和调用浏览器。
用 Three.js 做一个第一人称 3D 密室逃脱：3 个房间、6 个可交互物件。
硬约束：
1. 物件之间必须有【连锁依赖链】，不能是六个平行谜题：拿到 A 才能使用 B，B 与 C 组合才能开 D，D 打开才能进第 3 个房间…… 请把这条依赖链明确画出来给我看；
2. 每个物件点击都必须有明确反馈（有反馈，但还没到能用的时候，也要给出“为什么现在用不了”的提示，而不是静默无响应）；
3. 有一个“当前目标”提示，保证玩家不会被卡死；
4. 全流程 5 分钟内可通关；
5. 全部文件和中间文件都保存在当前文件夹下。
做完后：自己从头到尾走一遍通关流程，把每一步的操作和反馈写出来。
如果你自己都没通到关，直接告诉我卡在哪一步。
模型既要实现 3 个房间和 6 个物件，还要把它们串成依赖图；玩家乱点、提前点击、距离过远时都要获得反馈；最后模型还得亲自从头走到尾。首先是 6.1 Sol，做出的《余光 / AFTERLIGHT》有一套完整的复古值班室视觉：绿色墙裙、格纹地板、三维标牌、物品栏、声音开关和常驻目标提示都已经进入成品。六个物件被安排成两路汇合的依赖链：A 保险丝恢复 B 配电箱供电，C 门禁卡与供电状态共同解锁 D 终端，随后才能进入第三个房间取得 E 钥匙并打开 F 出口。GPT 6.1 Sol 构建的 3D 密室逃脱游戏体验过程不同于 Sol，Sonnet 5.5 采用灰色牢房和配电室风格，画面装饰较少，谜题链条依然完整。玩家需要先从床垫下取出螺丝刀，再拆通风口后拿到管理员钥匙；钥匙同时打开牢房门和书桌抽屉，抽屉里的门禁卡解锁工具柜并取得电池，最后由“门禁卡 + 电池”共同打开出口电子门。第三个房间尽头的光门只负责触发结算，没有被偷算进 6 个交互物件。Claude Sonnet 5.5 构建的 3D 密室逃脱游戏体验过程通过评测人员的感受来看，Sol 更像是交作业的流水账日记，Sonnet 则明显设置了玩密室的悬念感。Sol 的每一个步骤都被标记出来，同时，门禁卡的设计更像违背物理规律的悬空设计。反观 Sonnet，钥匙的质感和门上的钥匙孔都展现出一种真实感。本题专门考验时效信息与证据纪律，行政令、国际联合声明、出口规则和科研安全政策混在一起，很容易把政治措辞写成技术突破，也容易把一条搜索摘要扩写成既成事实。
请基于美国行政令 14434、《京都愿景》及相关一手资料，完成一篇面向中国技术从业者的调查报告：
1. 说明“AI 改名 SI”在法律、技术和产业层面分别改变了什么、没有改变什么；
2. 建立“说法—原始证据—可信度—可能误读”表格；
3. 分析它对中国 AI 企业、开源社区和科研合作可能产生的三项实际影响；
4. 列出五个可能出现但不准确的中文标题，并逐一纠正；
5. 所有时效性事实必须附来源，不得把政策措辞当成技术突破。
请阅读诺贝尔奖官方 2026 年化学奖的公告及科普材料（材料需要自行获取），完成以下交付：
1. 用普通读者能理解的方式解释手性、对映异构体、非线性效应和不对称自催化
2. 解释 Kagan 与 Soai 的贡献分别解决了哪一步问题，不能混为一谈
3. 制作一个单文件 HTML 互动演示，展示微小初始偏差如何在自催化过程中被放大
4. 明确说明互动演示是教学模型，而非对 Soai 反应的精确动力学复现
5. 输出一份五页演示文稿提纲，并核查所有数字、年份和药物相关表述。
雷峰网原创文章，未经授权禁止转载。详情见转载须知。
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
Anthropic launched Claude Sonet 5.5, Open AI launched GPT 6.1 Sol on September 29. Coincidentally, the API price of the two models on OpenRouter is exactly the same: Token 2 dollars per million and output 10. The Ray Peaks Network (WeChat Official Account:Ray Peaks Network) gives a high degree of direction: to push the ability to approach the highest-end models into a faster, cheaper product layer, suitable for Agent and professional work. Sonnet 5.5 emphasizes that fewer steps are taken to complete the day-to-day tasks of clear borders; GPT 6.1 Sol targets complex coding, computer operations and professional work. We put the two models into the same workstation, which we call through the same channel and require the model to perform the three successive tasks of 3D secret room development, up-to-date policy surveys and awards of chemistry. Finally, GPT 6.1 Sol spent $38.5 and Claude Sonet 5.5 spent $10.9, a difference of about 3.5 times. The difference in speed is more interesting. When developing the 3D secret room, Sol spent 33 minutes, and Sonnet spent 74 minutes; when the policy investigation was completed, the situation was completely reversed, Sonet completed the report in only 5 minutes, and Sol took almost an hour. In addition to the final deliverables, we compared end-to-end time, tool call, original filing, running validation and actual billing. As a result, there was no coherent win-and-win line, GPT 6.1 Sol developed faster in the chamber and built deeper evidentiary chains and publish closed loops in two research missions; Claude Sonnet 5.5 delivered available results on shorter routes, with both studies coming several times faster and the total cost of the three missions being about 28 per cent of Sol ' s. It's a classic "saw-nose intelligence": Sol is a deep-working workaholic willing to continue testing, Sonet 5.5 more like an efficient day-to-day pioneer. A detailed analysis shows that the model can only determine the capacity ceiling, and that the tool chain, the means of validation and the cessation strategy determine what we ultimately get and how much of it is at the core. The two products also reflect the industry position of the medium-term model, which is on the rise: it is beginning to become the main enterprise AI force implementation layer, taking on most of the complex and high-frequency production tasks. When model capabilities meet the threshold for use, stable delivery and unit mission costs will determine the model ' s industry value more than if the individual list were ahead. This paper selects two top-level models of the recent publish for comparison testing: GPT 6.1 Sol and Claude Sonet 5.5 from OpenAi and Anthropic, respectively. We use Claude Code Cli as a testing platform to call two models from OpenRouter channel to perform three complex tasks. Both models can search web pages, read and write files, run commands and call browsers.
Three.js as a first person called 3D secret chamber escaped: 3 rooms, 6 interactive objects.
Hard constraints:
There must be a [chain of dependencies] between objects, not six parallel puzzles: get A for use with B, B and C to open D, D for access to room 3... Please clearly paint this chain of dependence for me;
Each object must have a clear feedback (with feedback, but not yet available, a hint of “why not” rather than silence;
There is a “current target” warning that players will not be stuck;
4. Full process clearance within 5 minutes;
All documents and intermediate files are kept under the current folder.
After completion: go through the customs clearance process and write down the operation and feedback of each step.
If you don't get through it yourself, just tell me where it is.
The model is designed to achieve three rooms and six objects, and to provide them with a dependencies map; to obtain feedback from players when they are in disarray, early clicks, and too far away; and to have the final model go from the beginning to the end. First of all, 6.1 Sol, the Sunlight / AFTERLIGHT has a complete set of retrospects: green wall skirts, grafted floors, 3-D signboards, object bars, sound switches and permanent target alerts have all entered the product. Six items were arranged into a two-way chain of dependence: A fuse restores B distribution box power, C-barrel card unlocks the D terminal together with the power supply, before entering the third room to obtain E key and open F exit. GPT 6.1 Sol ' s built 3D secret chamber escape experience is different from Sol, Sonet 5.5, which uses grey cell and electrical room styles, with fewer graphics and a complete puzzle chain. Players need to remove screwdrivers from their mattresses before tearing down vents before getting the administrator's key; the key opens both the cell door and desk drawers, the door-barlock lockers in the drawers and the battery, and finally the electronic door is opened jointly by the “door-barrel + battery”. The light door at the end of the third room was only responsible for triggering the settlement and was not counted into six interactive objects. Claude Sonet 5.5's built 3D secret room escape experience is felt by the evaluators, and Sol looks more like a running water journal, and Sonet clearly sets up a hangover in the chamber. Each step of Sol was marked, while the door-bar card was designed more like a suspension design that ran counter to physical laws. Retrospect, Sonnet, the quality of the key and the keyhole on the door show a sense of truth. This topic is designed to test time-barred information and evidence discipline, combining executive orders, international joint statements, export rules and scientific security policies, and it is easy to write political language into technological breakthroughs and to expand a search for summary to a fait accompli.
Please complete a study for Chinese technology practitioners based on US Executive Order 14434, Kyoto Vision and related firsthand information:
Description of what has changed and what has not changed at the legal, technical and industrial levels, respectively, by the “AI name change SI”;
(b) Form “Statement—original evidence—credibility—possible misinterpretation”;
3. To analyse its possible three practical implications for Chinese AI enterprises, open-source communities and scientific cooperation;
4. List five possible but inaccurate Chinese titles and correct each;
5. All time-bound facts must be attached to the source and policy language must not be used as a technological breakthrough.
Please read the announcement of the Nobel Prize Official 2026 Chemical Prize and the materials of the Science and Technology Programme (the material is self-accessible) and complete the following:
1. Explanation of manuality, enantiomers, non-linear effects and asymmetric self-catalytication in a manner understood by ordinary readers
2. Explain which steps Kagan and Soai contributions have resolved separately and cannot be confused
3. Production of a single document HTML interactive presentation showing how small initial deviations are magnified during self-catalyzing
4. Clarify that the interactive presentation is a pedagogical model and not an accurate replica of the dynamics of the Soai response
5. Output of a five-page presentation outline and verification of all numbers, years and drug-related representations.
It's an original Ray Peak article, which is forbidden to reproduce without authorization. For further information, see the information below.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.