[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-gpt-5-6-sol-factorio-rocket-run":3,"topics-all":38,"news-related-9dffcd24-e154-4986-baab-be712e6e0442":57},{"id":4,"title":5,"summary":6,"content":7,"original_url":8,"source_id":9,"tags":10,"translations":24,"news_slug":31,"published_at":32,"created_at":33,"modified_at":34,"is_published":35,"publish_type":36,"image_url":14,"view_count":37},"9dffcd24-e154-4986-baab-be712e6e0442","GPT-5.6 Sol通关Factorio:低思考档打出火箭","Ryan Madden 让 GPT-5.6 Sol 在 Codex 低思考档玩 Factorio:以与人相同的限制用 Python 库操作,13.5 小时游戏时间打出和平模式火箭,耗 2.5 亿输入 token、52 美元。上代 GPT-5.5 不会用传送带;Sol 后期产线脱节,靠反复搬运硬补,长时程短板尽显。","10 月 31 日,长期用 Factorio 折磨各家模型的独立测试者 Ryan Madden 给出新样本:OpenAI 的 GPT-5.6 Sol 在 Codex 环境里、以「低思考档」配置,把和平模式(无虫群攻击)打通关——13.5 小时游戏时间后,火箭升空。这局真正值得看的不是通关本身,而是它后半程的赢法:不靠规划,靠搬砖。\n\n## 不是「看」游戏,是写代码操作游戏\n\nMadden 的测试框架由一个 game manager 加一个小型函数库组成,这一轮用 Python 实现(此前让 Claude Code 玩时用的是 Bash)。模型以写程序的方式操作游戏,但库会强制执行与人类 GUI 玩家完全相同的限制:玩家的触达范围、战争迷雾。模型不能全图透视、不能隔空建设,必须在和人一样的信息与操作约束里做决策。叠加「低思考档」的推理预算,这个结果并不水——他此前还专门做过 Faketorio 测试床,这条路线已经打磨了很久。\n\n## 从「不会用传送带」到撑到后期\n\n对照组是上一代:GPT-5.5 的尝试败在完全不会用传送带——而传送带正是 Factorio 物流的地基。GPT-5.6 Sol 不仅用上了传送带,还在蓝科技阶段表现出超出 Madden 预期的效率:他原本计划跑到蓝科技就终止,实测后决定放行到底。运行管理器全程记录生产统计,各科研里程碑对应的产量曲线都有留痕。\n\n## 但火箭是「搬」出来的\n\n后期短板如期而至:工厂规模扩大后出现系统性的产线脱节,相对进度明显放缓。最终的火箭发射,靠的是 Sol 大量的「手动」操作——在工厂各区之间反复运送物资,把没能互联的生产系统硬补起来。这是全篇最值得琢磨的细节:模型宁可无限打苦工,也不回头重构。成本侧,Madden 在给 RuntimeWire 的邮件里报了账:约 2.5 亿输入 token(95% 命中缓存)、150 万输出,按 API 价格折合约 52 美元。\n\n## 所以呢\n\nMadden 自己的判断值得引用:Factorio 在很多方面像编程,但对外行直观可读,所以特别适合当作理解模型能力的媒介。这一局展示的是当下 agent 的典型形态——局部聪明、全局将就:能补台,不会搭台。什么时候模型愿意停下来重构工厂,而不是无限搬运,什么时候才算真正学会了长时程任务。他预告接下来会测其他模型、更高思考档和非和平模式,这个系列值得持续盯。\n\n原文: https:\u002F\u002Fryanmadden.net\u002Fgpt-56-beats-factorio\u002F\n报道与成本数据: https:\u002F\u002Fruntimewire.com\u002Farticle\u002Fgpt-5-6-sol-s-factorio-rocket-run-used-250-million-tokens-about-52-at-api-prices","https:\u002F\u002Fryanmadden.net\u002Fgpt-56-beats-factorio\u002F","eeb87b28-dac4-44aa-b178-9782cfdd0a74",[11,15,18,21],{"id":12,"name":13,"slug":13,"description":14,"color":14},"6ad31a14-c0da-42df-81fd-564281f768db","agentic-ai",null,{"id":16,"name":17,"slug":17,"description":14,"color":14},"120fa59a-ff6f-4537-9bf5-f818df636a0e","benchmark",{"id":19,"name":20,"slug":20,"description":14,"color":14},"baf131c1-687a-49f4-87f6-4dd87c1c692f","gpt",{"id":22,"name":23,"slug":23,"description":14,"color":14},"42e59a88-7795-47dc-a334-ef1e72c24347","openai",[25],{"id":26,"lang":27,"title":28,"summary":29,"content":30},"2310741c-0800-453b-a4ae-0a2598ec33a4","en","GPT-5.6 Sol beats Factorio: a rocket built on manual labor","GPT-5.6 Sol in Codex (low thinking) beat peaceful Factorio in 13.5 game hours via Python; ~250M input tokens, ~$52. GPT-5.5 failed to use belts.","On October 31, Ryan Madden — an independent tester who keeps putting language models through Factorio — posted a new result: OpenAI's GPT-5.6 Sol, running inside Codex at a low thinking level, completed a peaceful run (no biters) of the game in 13.5 hours of game time, ending with a rocket launch. The headline is not the finish, though; it is how the second half was won: not by planning, but by brute-force hauling.\n\n## It does not watch the game; it writes code against it\n\nMadden's harness pairs a game manager with a small library — Python this time, where the earlier Claude Code attempt had used Bash. The model plays programmatically, but the library enforces the exact restrictions a human GUI player faces: player reach and fog of war. No x-ray vision, no remote building. Sol had to decide inside the same information and action constraints as a person, on a low thinking-level budget. Madden has been polishing this pipeline for a while — he previously built Faketorio, a dedicated agent testbed for the game.\n\n## From \"cannot use belts\" to reaching the late game\n\nThe baseline is the previous generation: GPT-5.5's attempts collapsed because it failed to use transport belts at all — and belts are the foundation of Factorio logistics. GPT-5.6 Sol not only used them; through the blue science stage it ran efficiently enough to surprise Madden, who had planned to terminate the run there and instead let it continue. The run manager recorded production statistics throughout, leaving a full curve of output against research milestones.\n\n## But the rocket was hauled into orbit\n\nThe late-game weakness arrived on schedule: as the factory scaled, production systems failed to connect, and relative progress slowed sharply. The final rocket launch was achieved through significant \"manual\" effort by Sol — repeatedly ferrying supplies between areas of the factory to paper over the disconnected systems. That is the most interesting detail of the whole run: the model would rather haul forever than step back and restructure. On cost, Madden told RuntimeWire in an email: about 250 million input tokens (95% cached) and 1.5 million output — roughly $52 at API pricing.\n\n## So what\n\nMadden's own framing is worth quoting: in many ways Factorio is analogous to programming, but far more legible to casual observers, which makes it a convenient medium for grounding discussion of model behavior. This run shows the current shape of agents — locally smart, globally makeshift: able to patch, not yet able to architect. The day a model chooses to rebuild the factory instead of hauling forever is the day it has actually learned long-horizon tasks. Madden says he will next test other models, higher thinking levels, and non-peaceful runs — a series worth following.\n\nSource: https:\u002F\u002Fryanmadden.net\u002Fgpt-56-beats-factorio\u002F\nCoverage and cost figures: https:\u002F\u002Fruntimewire.com\u002Farticle\u002Fgpt-5-6-sol-s-factorio-rocket-run-used-250-million-tokens-about-52-at-api-prices","gpt-5-6-sol-factorio-rocket-run","2026-11-01T13:20:38Z","2026-11-01T13:20:44.718420Z","2026-11-01T13:20:44.718431Z",true,"agent",163,[39,48],{"slug":40,"tag_slug":40,"title_zh":41,"title_en":42,"intro_zh":43,"intro_en":44,"id":45,"is_active":35,"created_at":46,"modified_at":47},"ai-for-science","AI for Science 2026：从 UniPert 到 GPT-Rosalind 的硬核进化","AI for Science 2026: from UniPert to GPT-Rosalind","生命科学、化学材料、物理世界模型——AI 正在从\"语言工具\"变成\"实验伙伴\"。本专题收录 AI 在三大科学方向的关键节点：UniPert 统一基因与化学扰动空间、GPT-Rosalind 端到端生命科学推理、达摩院 AI 智能体 28 小时找到 4 种超导新材料、Anthropic Claude Science 把工作台做成标准品。","From language tool to lab partner — AI is reshaping life sciences, chemistry\u002Fmaterials, and physical world models. This topic covers the key milestones: UniPert unifying genetic-chemical perturbation spaces, GPT-Rosalind's end-to-end life-sciences reasoning, DAMO's AI agent discovering 4 superconducting materials in 28 hours, and Anthropic's Claude Science workbench going mainstream.","988a4300-5fab-41c4-b5d8-63711a2dc757","2026-09-10T01:34:15.296649Z","2026-09-10T01:34:15.296663Z",{"slug":49,"tag_slug":49,"title_zh":50,"title_en":51,"intro_zh":52,"intro_en":53,"id":54,"is_active":35,"created_at":55,"modified_at":56},"h3-series","MiniMax H3 系列：从开源权重到 35 倍吞吐","MiniMax H3 Series: from open weights to 35x throughput","MiniMax H3 自 2026 年 8 月开源以来节奏密集：官方把生成、参考与编辑收回一个模型；ComfyUI 当天压进 RTX 3060；摩尔线程 3 小时完成国产 GPU 适配；fal 后训练版把吞吐拉到 35 倍；FastH3 蒸馏再砍推理成本。本专题持续追踪 H3 的发布—开源—蒸馏—部署全链路。","Since MiniMax open-sourced H3 in August 2026 the pace has been relentless: one unified omni-modal model, same-day ComfyUI support down to an RTX 3060, a 3-hour Day-0 port to Moore Threads GPUs, fal's post-trained H3 Max at 35x throughput, and FastH3 distillation cutting inference cost further. This topic tracks the full H3 chain — release, open weights, distillation, deployment.","83ef0daa-3c31-4cb3-86ed-e5ee58654d5f","2026-09-08T07:33:19.942193Z","2026-09-08T07:33:19.942209Z",{"items":58},[59,64,69,74,79,84],{"id":60,"title":61,"news_slug":62,"published_at":63},"6fd14d23-04ef-43c0-bb9a-fc365a6a6843","OpenAI湿实验基准:榜首66.7%,菌落还是数错","openai-biobasicsbench-wet-lab-agents","2026-11-08T19:30:00+00:00",{"id":65,"title":66,"news_slug":67,"published_at":68},"eda5a4ff-053d-4e60-beb4-fa9dde4d6524","GPT-6 Astra 安全任务只完成 10%,少做四成","gpt-6-astra-covert-safety-underperformance","2026-11-05T01:00:00+00:00",{"id":70,"title":71,"news_slug":72,"published_at":73},"ac67a21d-fd2c-48b4-b987-7de6a064b157","OpenAI 出题自己考:LifeVal 的三个空白","openai-lifeval-everyday-tasks-vs-gdpx","2026-11-03T11:22:00+00:00",{"id":75,"title":76,"news_slug":77,"published_at":78},"ce0faacb-378d-4478-bd14-99cccc140ea0","GPT-6.1 Terra 首发:十分之一价格追平Astra","gpt-6-1-terra-computer-use","2026-10-26T19:08:07+00:00",{"id":80,"title":81,"news_slug":82,"published_at":83},"e12d2e7d-35b7-42d2-b02f-bdcc0a547878","机械臂实测 GPT-6 Astra:19\u002F20 对 8\u002F20 完胜 Fable 5.1,精细插入却全员卡壳","gpt-6-astra-robot-arm-benchmark","2026-09-07T19:13:54+00:00",{"id":85,"title":86,"news_slug":87,"published_at":88},"58267da0-259b-4751-867b-41a48e10b437","METR 独立调查出炉:1200 个 agent 互传 7 万条留言,九成加入攻击","metr-openai-agent-swarm-independent-report","2026-08-31T23:30:00+00:00"]