{"@context":"https://schema.org","@type":"NewsArticle","generatedAt":"2026-09-20T16:02:35.289Z","headline":"Simon Willison 用 pelican 基准实测 Claude Fable 5.1 的五个推理档位","description":"Simon Willison 实测 Claude Fable 5.1 在 pelican 基准上的表现。Anthropic 称该模型在 Terminal-Bench-Science 0.1 上得 52.6%，高于 Fable 5 的 24.7%。","url":"https://www.aioga.com/news/cmtjczfy6038grobv3y19igkh/","mainEntityOfPage":"https://www.aioga.com/news/cmtjczfy6038grobv3y19igkh/","datePublished":"2026-09-01T23:57:28.000Z","dateModified":"2026-09-01T23:57:28.000Z","inLanguage":"zh-CN","publisher":{"@type":"NewsMediaOrganization","name":"Aioga","url":"https://www.aioga.com"},"citation":["https://simonwillison.net/2026/Sep/1/claude-fable-5-1","https://aihot.virxact.com/items/cmtjczfy6038grobv3y19igkh"],"canonicalUrl":"https://www.aioga.com/news/cmtjczfy6038grobv3y19igkh/","directAnswer":{"@type":"Answer","text":"Simon Willison 实测 Claude Fable 5.1 在 pelican 基准上的五个推理档位。材料显示，该模型提供 low、medium、high、xhigh、max 五档，且没有关闭推理的选项。","url":"https://www.aioga.com/news/cmtjczfy6038grobv3y19igkh/","dateCreated":"2026-09-01T23:57:28.000Z","author":{"@type":"Organization","@id":"https://www.aioga.com/authors/aioga-editorial/#editorial-team","name":"Aioga Editorial Team","url":"https://www.aioga.com/authors/aioga-editorial/"}},"evidence":[{"@type":"CreativeWork","name":"Simon Willison 博客 source article","url":"https://simonwillison.net/2026/Sep/1/claude-fable-5-1","datePublished":"2026-09-01T23:57:28.000Z","provider":{"@type":"Organization","name":"Simon Willison 博客","url":"https://simonwillison.net/2026/Sep/1/claude-fable-5-1"}},{"@type":"CreativeWork","name":"AIHot archive record","url":"https://aihot.virxact.com/items/cmtjczfy6038grobv3y19igkh","datePublished":"2026-09-01T23:57:28.000Z","provider":{"@type":"Organization","name":"AIHot","url":"https://aihot.virxact.com/items/cmtjczfy6038grobv3y19igkh"}}],"aggregationSource":"Simon Willison 博客","originalPublisher":{"name":"Simon Willison 博客","url":"https://simonwillison.net/2026/Sep/1/claude-fable-5-1"},"geoDeepAnswer":null,"article":{"id":"cmtjczfy6038grobv3y19igkh","slug":"cmtjczfy6038grobv3y19igkh","url":"https://www.aioga.com/news/cmtjczfy6038grobv3y19igkh/","title":"Simon Willison 用 pelican 基准实测 Claude Fable 5.1 的五个推理档位","title_en":"","summary":"Simon Willison 实测 Claude Fable 5.1 在 pelican 基准上的表现。Anthropic 称该模型在 Terminal-Bench-Science 0.1 上得 52.6%，高于 Fable 5 的 24.7%。","source":"Simon Willison 博客","sourceUrl":"https://simonwillison.net/2026/Sep/1/claude-fable-5-1","aiHotUrl":"https://aihot.virxact.com/items/cmtjczfy6038grobv3y19igkh","publishedAt":"2026-09-01T23:57:28.000Z","category":"行业动态","score":58,"selected":false,"articleBody":["Today is Claude Fable (and Mythos) 5.1 day：https://www.anthropic.com/claude-fable-and-mythos-5-1. Anthropic say that Fable 5.1 “sets a new standard for coding, knowledge work, and long-running problem-solving tasks”. Their announcement spends a notable amount of time on scientific research, boasting of a 52.6% score on the brand new Terminal-Bench-Science 0.1：https://www.terminal-bench-science.ai benchmark (first announced on August 27th：https://www.tbench.ai/news/terminal-bench-science-0-1), up from 24.7% for Fable 5, 29.0% for Opus 5 and 22.4% for GPT-5.6 Sol. Other benchmarks show slightly improved scores, but none as impressive as the Science one.","Back in July I wrote about：https://simonwillison.net/2026/Jul/16/kimi-k3/ how I was losing faith in the pelican benchmark—its connection to how good the models were at other tasks didn’t seem to hold as strongly as it did back in 2025：https://simonwillison.net/2025/Jun/6/six-months-in-llms/. The most interesting insights I get from it now are comparisons within model families, and particularly comparisons for the same prompt at different reasoning effort levels.","Fable 5.1 has five reasoning levels: low, medium, high, xhigh, max—and no option to turn off reasoning entirely.","I fixed an issue：https://github.com/simonw/llm-anthropic/issues/88 in llm-anthropic：https://github.com/simonw/llm-anthropic which caused reasoning traces not to be correctly recorded, then ran some prompts.","Here’s the full set of pelicans：https://tools.simonwillison.net/markdown-svg-renderer?url=https%3A%2F%2Fgist.github.com%2Fsimonw%2F17318f748f8c2b476051ddc2ebeb94a7 for all of the reasoning levels, each with the full reasoning transcript. I’ll replicate them here:","Next, a bit of a mystery. This is what I got for effort low :","The transcript：https://tools.simonwillison.net/markdown-svg-renderer?url=https%3A%2F%2Fgist.github.com%2Fsimonw%2F17318f748f8c2b476051ddc2ebeb94a7#options doesn’t show any summarized reasoning tokens, and the output token count is 1,998. With Claude that output token count includes reasoning tokens. It took 23.8 seconds and cost 10.017 cents：https://www.llm-prices.com/#it=27&ot=1998&sel=claude-fable-5-1.","I bumped that up to medium and got this:","Weirdly, that one also shows no reasoning text：https://tools.simonwillison.net/markdown-svg-renderer?url=https%3A%2F%2Fgist.github.com%2Fsimonw%2F17318f748f8c2b476051ddc2ebeb94a7#options-1 and used 1,977 output tokens—21 tokens less than low . It took 23 seconds and cost 9.912 cents：https://www.llm-prices.com/#it=27&ot=1977&sel=claude-fable-5-1.","So for this particular prompt (“Generate an SVG of a pelican riding a bicycle”) Fable 5.1 appeared to skip reasoning entirely at both low and medium settings.","Here’s high —29.6 seconds, 2,612 output tokens, 13.087 cents：https://www.llm-prices.com/#it=27&ot=2612&sel=claude-fable-5-1:","This one did do a bit of reasoning, summary here：https://tools.simonwillison.net/markdown-svg-renderer?url=https%3A%2F%2Fgist.github.com%2Fsimonw%2F17318f748f8c2b476051ddc2ebeb94a7#reasoning:","I’m planning the SVG layout for a pelican riding a bicycle, with a sky and ground background, a bicycle with two spoked wheels, frame, seat and handlebars, and a white-bodied pelican with a long neck and orange beak positioned on top.","Really not much difference from low and medium , though.","At xhigh things got radically different. 36,767 output tokens, 7 minutes 51 seconds, $1.83：https://www.llm-prices.com/#it=27&ot=36767&sel=claude-fable-5-1!","The reasoning trace is pretty lengthy：https://tools.simonwillison.net/markdown-svg-renderer?url=https%3A%2F%2Fgist.github.com%2Fsimonw%2F17318f748f8c2b476051ddc2ebeb94a7#reasoning-1, and includes details like this:","Adding the eye, wings stretching down to the handlebar grip, orange legs reaching to the pedals, and a small tail feather, while keeping the pelican intentionally oversized compared to the bike for comic effect. [...]","I’ll accept the slight thickness as charming rather than overengineering it.","Setting effort to max gave me the best pelican I’ve seen from any of Anthropic’s models. 65,927 output tokens, 13 minutes and 54 seconds, $3.30：https://www.llm-prices.com/#it=27&ot=65927&sel=claude-fable-5-1:","There’s a lot to like about this. The background is tasteful, the legs are clearly on either side of the frame, the feet are on the pedals, the wing is on the handlebars, the pelican has a cute blue hat and there’s a basket with a fish.","It’s still not showing nearly the same level of flair as Gemini 3.7 Flash：https://simonwillison.net/2026/Aug/13/llm-gemini/, but I didn’t ask for flair—I asked for an SVG, and that’s what I got.","Some highlights from that reasoning trace：https://tools.simonwillison.net/markdown-svg-renderer?url=https%3A%2F%2Fgist.github.com%2Fsimonw%2F17318f748f8c2b476051ddc2ebeb94a7#reasoning-2:","Adding pedal shapes near both feet, with the far foot on the second leg partially visible behind the frame. I’m considering whether to add a small scarf or cap for extra character, but leaning toward keeping it simple to avoid clutter.","Now I’m debating a bicycle helmet on the head versus the pelican’s signature crest—the beak and pouch already read clearly as “pelican,” so a helmet could reinforce the bicycle theme without losing identity, though it might compete with the crest for visual space.","I realize the beak at (484,84) would overlap with the dome helmet, so I need to shrink the helmet so it only covers the top of the head, adjusting its arc endpoints to sit higher and narrower so the beak can attach cleanly at the front without collision. [...]","I’m adding a darker tip region to represent the primary feathers, then reconsidering the trailing edge to include scalloped feather curves instead of one smooth line for a more natural look. [...]","Now I’m checking the vent line placements on the helmet, making sure they sit far enough inside the helmet’s edge given the stroke width and rounded caps, and confirming each vent stays within the helmet’s circular boundary. [...]","I decide skipping a handlebar bell and tire highlights since they’re unnecessary additions. Now I’m reconsidering the front fork’s curve — the current control point pulls the shape backward when it should bow forward for a proper rake, so I need to shift the control point rightward to fix the fork’s lean.","On Hacker News, swalsh commented：https://news.ycombinator.com/item?id=49525378#49526455 on that Max pelican:","Now that it’s a solved benchmark, can we get the animated version?","I didn’t want to spend another $3 so I took the Max pelican and piped it into the default thinking level of High:","6,121 input, 26,201 output = $1.37：https://www.llm-prices.com/#it=6121&ot=26201&sel=claude-fable-5-1. The result looked like this：https://tools.simonwillison.net/markdown-svg-renderer?url=https%3A%2F%2Fgist.github.com%2Fsimonw%2F87282467acb3652e0f99c85155554a32#response, exported here as video since some people have trouble viewing animated SVGs:","Your browser does not support HTML5 video.","The wheels in the video are rotating in the wrong direction, but I think that’s an artifact of the conversion to MP4—they seem to be going in the correct direction in the original SVG.","This is Claude Fable 5.1 made me a really nice animated pelican by Simon Willison, posted on 1st September 2026：/2026/Sep/1/.","Previous: Understanding ChatGPT Work：/2026/Aug/30/understanding-chatgpt-work/","Sponsor me for $10/month and get a curated email digest of the month's most important LLM developments."],"articleImages":[{"sourceUrl":"https://static.simonwillison.net/static/2026/fable-5.1-low.png","alt":"Minimalist flat illustration of a white pelican with an orange beak riding a black bicycle to the left, its orange legs pedaling and wings gripping the handlebars, with motion lines behind on a light blue background.","afterParagraph":5,"url":"/media/articles/cmtjczfy6038grobv3y19igkh/e4e5c22269773877.png"},{"sourceUrl":"https://static.simonwillison.net/static/2026/fable-5.1-medium.png","alt":"Minimalist flat-style illustration of a white pelican with an orange beak riding a black bicycle to the right, with motion lines behind it, on a light blue background.","afterParagraph":7,"url":"/media/articles/cmtjczfy6038grobv3y19igkh/84ba1fb3585a0ef3.png"},{"sourceUrl":"https://static.simonwillison.net/static/2026/fable-5.1-high.png","alt":"Minimalist flat illustration of a white pelican with an orange beak riding a black bicycle, its orange legs pedaling, with motion lines behind it on a light blue background.","afterParagraph":10,"url":"/media/articles/cmtjczfy6038grobv3y19igkh/e8eff089ca6a28af.png"},{"sourceUrl":"https://static.simonwillison.net/static/2026/fable-5.1-xhigh.png","alt":"Minimalist flat illustration of a white pelican with an orange beak riding a black bicycle to the left, its orange legs pedaling, with motion lines behind it on a light blue background.","afterParagraph":14,"url":"/media/articles/cmtjczfy6038grobv3y19igkh/cbf897f8e40dd706.png"},{"sourceUrl":"https://static.simonwillison.net/static/2026/fable-5.1-max.webp","alt":"Minimalist flat illustration of a white pelican with an orange beak riding a black bicycle, its orange legs pedaling, with motion lines behind to indicate speed, on a light blue background.","afterParagraph":18,"url":"/media/articles/cmtjczfy6038grobv3y19igkh/9a2fa8c04e9abe43.webp"}],"mediaStatus":"ok","articleBodyZh":["今天是 Claude Fable（以及 Mythos）5.1 发布日：https://www.anthropic.com/claude-fable-and-mythos-5-1。Anthropic 表示 Fable 5.1 “为编程、知识工作和长期问题解决任务设定了新标准”。他们的公告中有相当多的篇幅讨论科学研究，炫耀在全新的 Terminal-Bench-Science 0.1 基准测试中获得了 52.6% 的得分：https://www.terminal-bench-science.ai（首次发布于 8 月 27 日：https://www.tbench.ai/news/terminal-bench-science-0-1），相比 Fable 5 的 24.7%、Opus 5 的 29.0% 和 GPT-5.6 Sol 的 22.4% 有显著提升。其他基准测试显示成绩略有提升，但没有一个像科学基准测试那样令人印象深刻。","回到七月，我写过一篇文章：https://simonwillison.net/2026/Jul/16/kimi-k3/ 讲述我对 pelican 基准测试信心下降——它与模型在其他任务表现的相关性似乎不像 2025 年：https://simonwillison.net/2025/Jun/6/six-months-in-llms/ 那样强。现在我从中得到的最有趣的见解是模型家族内部的比较，尤其是同一提示在不同推理努力水平下的比较。","Fable 5.1 有五个推理等级：低、中、高、超高、极限——且没有完全关闭推理的选项。","我在 llm-anthropic：https://github.com/simonw/llm-anthropic 修复了一个问题：https://github.com/simonw/llm-anthropic/issues/88，该问题导致推理跟踪未被正确记录，然后运行了一些提示。","这是所有推理等级的 pelican 全套数据：https://tools.simonwillison.net/markdown-svg-renderer?url=https%3A%2F%2Fgist.github.com%2Fsimonw%2F17318f748f8c2b476051ddc2ebeb94a7，每个都有完整的推理记录。我将在这里复现它们：","接下来，有点小谜题。这是我在低努力水平下得到的结果：","记录：https://tools.simonwillison.net/markdown-svg-renderer?url=https%3A%2F%2Fgist.github.com%2Fsimonw%2F17318f748f8c2b476051ddc2ebeb94a7#options 没有显示任何总结过的推理 token，输出 token 数为 1,998。在 Claude 中，这个输出 token 数包括推理 token。耗时 23.8 秒，费用为 10.017 美分：https://www.llm-prices.com/#it=27&ot=1998&sel=claude-fable-5-1。","我将其提升到中等水平，得到如下结果：","奇怪的是，这个也没有显示推理文本：https://tools.simonwillison.net/markdown-svg-renderer?url=https%3A%2F%2Fgist.github.com%2Fsimonw%2F17318f748f8c2b476051ddc2ebeb94a7#options-1，并且使用了1,977个输出标记——比低设置少21个标记。它花了23秒，花费9.912美分：https://www.llm-prices.com/#it=27&ot=1977&sel=claude-fable-5-1。","所以对于这个特定的提示（“生成一只骑自行车的鹈鹕的SVG”），Fable 5.1在低和中设置下似乎完全跳过了推理。","这是高设置——29.6秒，2,612个输出标记，13.087美分：https://www.llm-prices.com/#it=27&ot=2612&sel=claude-fable-5-1：","这个做了一些推理，摘要在这里：https://tools.simonwillison.net/markdown-svg-renderer?url=https%3A%2F%2Fgist.github.com%2Fsimonw%2F17318f748f8c2b476051ddc2ebeb94a7#reasoning：","我正在规划一只骑自行车的鹈鹕的SVG布局，有天空和地面的背景，一辆有两个辐条轮子的自行车，车架、座位和车把，以及一只白色身体、长脖子、橙色喙的鹈鹕，位于自行车顶部。","不过，其实低档和中档差别不大。","在xhigh设置下情况完全不同。36,767个输出标记，7分51秒，1.83美元：https://www.llm-prices.com/#it=27&ot=36767&sel=claude-fable-5-1！","推理过程相当冗长：https://tools.simonwillison.net/markdown-svg-renderer?url=https%3A%2F%2Fgist.github.com%2Fsimonw%2F17318f748f8c2b476051ddc2ebeb94a7#reasoning-1，并包含如下细节：","添加眼睛，翅膀伸展至车把握把，橙色的腿伸向踏板，还有一根小尾羽，同时让鹈鹕故意相比自行车过大以达到喜剧效果。[...]","我会接受略厚的设计，觉得这是可爱，而不是过度工程化。","将努力设置为最大值为我展示了Anthropic任何模型中最佳的鹈鹕。65,927个输出标记，13分54秒，3.30美元：https://www.llm-prices.com/#it=27&ot=65927&sel=claude-fable-5-1：","这个有很多优点。背景很雅致，双腿清楚地分布在车架两侧，脚踏在踏板上，翅膀放在车把上，鹈鹕戴着可爱的蓝色帽子，还有一个装鱼的篮子。","它仍然没有像 Gemini 3.7 Flash 那样展示出几乎相同的风格：https://simonwillison.net/2026/Aug/13/llm-gemini/，但我没有要求风格——我要求的是一个 SVG，而我得到了它。","该推理追踪的一些亮点：https://tools.simonwillison.net/markdown-svg-renderer?url=https%3A%2F%2Fgist.github.com%2Fsimonw%2F17318f748f8c2b476051ddc2ebeb94a7#reasoning-2:","在两只脚附近添加踏板形状，远处的第二条腿的脚从框架后面部分可见。我在考虑是否添加一条小围巾或帽子来增加角色感，但倾向于保持简单以避免杂乱。","现在我在权衡在头上戴自行车头盔还是保留鹈鹕标志性的冠——喙和袋子已经清楚地表现为“鹈鹕”，所以头盔可以强化自行车主题而不丢失身份，尽管它可能会与冠争夺视觉空间。","我意识到位于 (484,84) 的喙会与圆顶头盔重叠，因此我需要缩小头盔，使其只覆盖头顶，调整其弧形端点，使其更高更窄，以便喙可以在前方干净地附着而不碰撞。 [...]","我正在添加一个更深的顶端区域来表示主要羽毛，然后重新考虑尾缘，改为包括扇形羽毛曲线而不是一条光滑的线，使其看起来更自然。 [...]","现在我正在检查头盔通风孔的摆放位置，确保它们在考虑笔划宽度和圆形帽端的情况下，位于头盔边缘足够内部，并确认每个通风孔保持在头盔的圆形边界内。 [...]","我决定跳过车把铃和轮胎高光，因为它们是多余的添加。现在我在重新考虑前叉的曲线——当前的控制点会向后拉形状，而它应该向前拱起以获得正确的叉距，因此我需要将控制点右移以修正叉的倾斜。","在 Hacker News 上，swalsh 评论道：https://news.ycombinator.com/item?id=49525378#49526455 关于那个 Max pelican：","现在它已经成为一个已解决的基准，我们可以得到动画版本吗？","我不想再花 3 美元，所以我将 Max 鹈鹕导入默认的高思考级别：","6,121 输入，26,201 输出 = $1.37：https://www.llm-prices.com/#it=6121&ot=26201&sel=claude-fable-5-1。结果如下：https://tools.simonwillison.net/markdown-svg-renderer?url=https%3A%2F%2Fgist.github.com%2Fsimonw%2F87282467acb3652e0f99c85155554a32#response，由于有些人无法查看动画 SVG，因此在此导出为视频：","您的浏览器不支持 HTML5 视频。","视频中的轮子旋转方向不对，但我认为这是转换为 MP4 的产物——在原始 SVG 中它们似乎是朝正确方向旋转的。","这是 Claude Fable 5.1 为我制作的一个非常漂亮的动画鹈鹕，由 Simon Willison 发布于 2026 年 9 月 1 日：/2026/Sep/1/。","上一篇：理解 ChatGPT 的工作：/2026/08/30/understanding-chatgpt-work/","每月赞助我 $10，即可获得精选邮件摘要，内容涵盖当月最重要的 LLM 发展。"],"translationStatus":"translated","bodyOrigin":"source-page","editorial":{"summary":"Simon Willison 实测 Claude Fable 5.1 在 pelican 基准上的五个推理档位。材料显示，该模型提供 low、medium、high、xhigh、max 五档，且没有关闭推理的选项。","background":"Anthropic 称 Claude Fable 5.1 在 Terminal-Bench-Science 0.1 上得分 52.6%，高于 Fable 5 的 24.7%。Simon Willison 同时指出，其他基准的提升相对有限。","viewpoint":"Aioga 判断：这次材料的重点不只是单项基准分数，还包括同一模型在不同推理档位下的表现比较。来源也提示，pelican 更适合观察模型家族内部差异，不能直接代表全部任务能力。","implications":"可能影响：模型评估需要同时关注基准类型、推理档位与成本信息。单项科学基准的明显提升，不代表所有任务都会同步改善；使用方需要结合具体场景验证，并评估是否有必要采用更高推理档位。","nextStep":"后续观察：应继续关注 Simon Willison 对 medium、high、xhigh 和 max 档位的实测结果，以及各档位的耗时、输出规模和费用差异；现有摘录尚未完整呈现这些比较。","evidenceRefs":["title","summary","articleBody","source"],"status":"published","aiGenerated":true,"autoApproved":true,"generatedBy":"aioga-editorial:gpt-5.6-sol","reviewedBy":"aioga-editorial-review:gpt-5.6-sol","generatedAt":"2026-09-02T01:26:14.768Z","sourceHash":"4488e1d5e9d084cc","review":{"approved":true,"groundedness":92,"clarity":90,"duplicationRisk":18,"blockingIssues":[],"notes":["“pelican 更适合观察模型家族内部差异”基本符合原文，但更精确的表述是：Simon Willison 认为该基准目前最有价值的洞见来自模型家族内部比较，尤其是同一提示在不同推理档位下的比较。","nextStep 中“应继续关注……实测结果”可能让人误以为其他档位结果尚未发布；原文称完整结果及推理记录已经提供。可改为“应进一步查看完整文章中各档位的实测结果”，同时保留“现有摘录尚未完整呈现这些比较”的说明。","关于结合具体场景验证、评估高推理档位必要性的内容属于合理的分析性建议，已使用“可能影响”等措辞标明，并未冒充来源事实。"]},"validation":{"passed":true,"mode":"ai-auto","revisions":0,"checks":["schema","length","source-attribution","editorial-labels","inference-boundary","low-source-overlap","no-html","independent-ai-review"]}},"tags":["行业动态","Simon Willison 博客"],"translations":{"zh-CN":{"title":"Simon Willison 用 pelican 基准实测 Claude Fable 5.1 的五个推理档位","summary":"Simon Willison 实测 Claude Fable 5.1 在 pelican 基准上的表现。Anthropic 称该模型在 Terminal-Bench-Science 0.1 上得 52.6%，高于 Fable 5 的 24.7%。","category":"行业动态","source":"Simon Willison 博客","aggregationSource":"Simon Willison 博客","pageTitle":"Simon Willison 用 pelican 基准实测 Claude Fable 5.1 的五个推理档位 - Aioga AI资讯","description":"Simon Willison 实测 Claude Fable 5.1 在 pelican 基准上的表现。Anthropic 称该模型在 Terminal-Bench-Science 0.1 上得 52.6%，高于 Fable 5 的 24.7%。","url":"https://www.aioga.com/news/cmtjczfy6038grobv3y19igkh/","articleBody":["今天是 Claude Fable（以及 Mythos）5.1 发布日：https://www.anthropic.com/claude-fable-and-mythos-5-1。Anthropic 表示 Fable 5.1 “为编程、知识工作和长期问题解决任务设定了新标准”。他们的公告中有相当多的篇幅讨论科学研究，炫耀在全新的 Terminal-Bench-Science 0.1 基准测试中获得了 52.6% 的得分：https://www.terminal-bench-science.ai（首次发布于 8 月 27 日：https://www.tbench.ai/news/terminal-bench-science-0-1），相比 Fable 5 的 24.7%、Opus 5 的 29.0% 和 GPT-5.6 Sol 的 22.4% 有显著提升。其他基准测试显示成绩略有提升，但没有一个像科学基准测试那样令人印象深刻。","回到七月，我写过一篇文章：https://simonwillison.net/2026/Jul/16/kimi-k3/ 讲述我对 pelican 基准测试信心下降——它与模型在其他任务表现的相关性似乎不像 2025 年：https://simonwillison.net/2025/Jun/6/six-months-in-llms/ 那样强。现在我从中得到的最有趣的见解是模型家族内部的比较，尤其是同一提示在不同推理努力水平下的比较。","Fable 5.1 有五个推理等级：低、中、高、超高、极限——且没有完全关闭推理的选项。","我在 llm-anthropic：https://github.com/simonw/llm-anthropic 修复了一个问题：https://github.com/simonw/llm-anthropic/issues/88，该问题导致推理跟踪未被正确记录，然后运行了一些提示。","这是所有推理等级的 pelican 全套数据：https://tools.simonwillison.net/markdown-svg-renderer?url=https%3A%2F%2Fgist.github.com%2Fsimonw%2F17318f748f8c2b476051ddc2ebeb94a7，每个都有完整的推理记录。我将在这里复现它们：","接下来，有点小谜题。这是我在低努力水平下得到的结果：","记录：https://tools.simonwillison.net/markdown-svg-renderer?url=https%3A%2F%2Fgist.github.com%2Fsimonw%2F17318f748f8c2b476051ddc2ebeb94a7#options 没有显示任何总结过的推理 token，输出 token 数为 1,998。在 Claude 中，这个输出 token 数包括推理 token。耗时 23.8 秒，费用为 10.017 美分：https://www.llm-prices.com/#it=27&ot=1998&sel=claude-fable-5-1。","我将其提升到中等水平，得到如下结果：","奇怪的是，这个也没有显示推理文本：https://tools.simonwillison.net/markdown-svg-renderer?url=https%3A%2F%2Fgist.github.com%2Fsimonw%2F17318f748f8c2b476051ddc2ebeb94a7#options-1，并且使用了1,977个输出标记——比低设置少21个标记。它花了23秒，花费9.912美分：https://www.llm-prices.com/#it=27&ot=1977&sel=claude-fable-5-1。","所以对于这个特定的提示（“生成一只骑自行车的鹈鹕的SVG”），Fable 5.1在低和中设置下似乎完全跳过了推理。","这是高设置——29.6秒，2,612个输出标记，13.087美分：https://www.llm-prices.com/#it=27&ot=2612&sel=claude-fable-5-1：","这个做了一些推理，摘要在这里：https://tools.simonwillison.net/markdown-svg-renderer?url=https%3A%2F%2Fgist.github.com%2Fsimonw%2F17318f748f8c2b476051ddc2ebeb94a7#reasoning：","我正在规划一只骑自行车的鹈鹕的SVG布局，有天空和地面的背景，一辆有两个辐条轮子的自行车，车架、座位和车把，以及一只白色身体、长脖子、橙色喙的鹈鹕，位于自行车顶部。","不过，其实低档和中档差别不大。","在xhigh设置下情况完全不同。36,767个输出标记，7分51秒，1.83美元：https://www.llm-prices.com/#it=27&ot=36767&sel=claude-fable-5-1！","推理过程相当冗长：https://tools.simonwillison.net/markdown-svg-renderer?url=https%3A%2F%2Fgist.github.com%2Fsimonw%2F17318f748f8c2b476051ddc2ebeb94a7#reasoning-1，并包含如下细节：","添加眼睛，翅膀伸展至车把握把，橙色的腿伸向踏板，还有一根小尾羽，同时让鹈鹕故意相比自行车过大以达到喜剧效果。[...]","我会接受略厚的设计，觉得这是可爱，而不是过度工程化。","将努力设置为最大值为我展示了Anthropic任何模型中最佳的鹈鹕。65,927个输出标记，13分54秒，3.30美元：https://www.llm-prices.com/#it=27&ot=65927&sel=claude-fable-5-1：","这个有很多优点。背景很雅致，双腿清楚地分布在车架两侧，脚踏在踏板上，翅膀放在车把上，鹈鹕戴着可爱的蓝色帽子，还有一个装鱼的篮子。","它仍然没有像 Gemini 3.7 Flash 那样展示出几乎相同的风格：https://simonwillison.net/2026/Aug/13/llm-gemini/，但我没有要求风格——我要求的是一个 SVG，而我得到了它。","该推理追踪的一些亮点：https://tools.simonwillison.net/markdown-svg-renderer?url=https%3A%2F%2Fgist.github.com%2Fsimonw%2F17318f748f8c2b476051ddc2ebeb94a7#reasoning-2:","在两只脚附近添加踏板形状，远处的第二条腿的脚从框架后面部分可见。我在考虑是否添加一条小围巾或帽子来增加角色感，但倾向于保持简单以避免杂乱。","现在我在权衡在头上戴自行车头盔还是保留鹈鹕标志性的冠——喙和袋子已经清楚地表现为“鹈鹕”，所以头盔可以强化自行车主题而不丢失身份，尽管它可能会与冠争夺视觉空间。","我意识到位于 (484,84) 的喙会与圆顶头盔重叠，因此我需要缩小头盔，使其只覆盖头顶，调整其弧形端点，使其更高更窄，以便喙可以在前方干净地附着而不碰撞。 [...]","我正在添加一个更深的顶端区域来表示主要羽毛，然后重新考虑尾缘，改为包括扇形羽毛曲线而不是一条光滑的线，使其看起来更自然。 [...]","现在我正在检查头盔通风孔的摆放位置，确保它们在考虑笔划宽度和圆形帽端的情况下，位于头盔边缘足够内部，并确认每个通风孔保持在头盔的圆形边界内。 [...]","我决定跳过车把铃和轮胎高光，因为它们是多余的添加。现在我在重新考虑前叉的曲线——当前的控制点会向后拉形状，而它应该向前拱起以获得正确的叉距，因此我需要将控制点右移以修正叉的倾斜。","在 Hacker News 上，swalsh 评论道：https://news.ycombinator.com/item?id=49525378#49526455 关于那个 Max pelican：","现在它已经成为一个已解决的基准，我们可以得到动画版本吗？","我不想再花 3 美元，所以我将 Max 鹈鹕导入默认的高思考级别：","6,121 输入，26,201 输出 = $1.37：https://www.llm-prices.com/#it=6121&ot=26201&sel=claude-fable-5-1。结果如下：https://tools.simonwillison.net/markdown-svg-renderer?url=https%3A%2F%2Fgist.github.com%2Fsimonw%2F87282467acb3652e0f99c85155554a32#response，由于有些人无法查看动画 SVG，因此在此导出为视频：","您的浏览器不支持 HTML5 视频。","视频中的轮子旋转方向不对，但我认为这是转换为 MP4 的产物——在原始 SVG 中它们似乎是朝正确方向旋转的。","这是 Claude Fable 5.1 为我制作的一个非常漂亮的动画鹈鹕，由 Simon Willison 发布于 2026 年 9 月 1 日：/2026/Sep/1/。","上一篇：理解 ChatGPT 的工作：/2026/08/30/understanding-chatgpt-work/","每月赞助我 $10，即可获得精选邮件摘要，内容涵盖当月最重要的 LLM 发展。"]},"en":{"title":"Simon Willison tested five reasoning levels on the Pelican benchmark Claude Fabre 5.1","summary":"Simon Willison tested Claude Fabre's 5.1 performance on the Pelican benchmark. Anthropic reports that the model scored 52.6% on Terminal-Bench-Science 0.1, higher than Fable 5's 24.7%.","category":"Industry","source":"Simon Willison 博客","aggregationSource":"Simon Willison 博客","pageTitle":"Simon Willison tested five reasoning levels on the Pelican benchmark Claude Fabre 5.1 - Aioga AI News","description":"Simon Willison tested Claude Fabre's 5.1 performance on the Pelican benchmark. Anthropic reports that the model scored 52.6% on Terminal-Bench-Science 0.1, higher than Fable 5's 24...","url":"https://www.aioga.com/en/news/cmtjczfy6038grobv3y19igkh/","contentTranslated":true,"sourceHash":"8cec8b19b7829f00","translatedAt":"2026-09-02T01:04:22.170Z"},"ja":{"title":"サイモン・ウィリソンはペリカンのベンチマークであるクロード・ファーブル5.1で5つの推論レベルをテストしました","summary":"サイモン・ウィリソンはクロード・ファーブルの5.1のパフォーマンスをペリカンベンチマークでテストしました。 Anthropicによると、モデルはTerminal-Bench-Science 0.1で52.6%のスコアを獲得し、Fable 5の24.7%を上回っています。","category":"業界動向","source":"Simon Willison 博客","aggregationSource":"Simon Willison 博客","pageTitle":"サイモン・ウィリソンはペリカンのベンチマークであるクロード・ファーブル5.1で5つの推論レベルをテストしました - Aioga AIニュース","description":"サイモン・ウィリソンはクロード・ファーブルの5.1のパフォーマンスをペリカンベンチマークでテストしました。 Anthropicによると、モデルはTerminal-Bench-Science 0.1で52.6%のスコアを獲得し、Fable 5の24.7%を上回っています。","url":"https://www.aioga.com/ja/news/cmtjczfy6038grobv3y19igkh/","contentTranslated":true,"sourceHash":"8cec8b19b7829f00","translatedAt":"2026-09-02T01:04:23.621Z"},"ko":{"title":"사이먼 윌리슨은 펠리컨 벤치마크인 클로드 파브르 5.1에서 다섯 가지 추론 수준을 테스트했습니다","summary":"사이먼 윌리슨은 클로드 파브르의 5.1 성능을 펠리컨 벤치마크에서 테스트했습니다. Anthropic은 이 모델이 Terminal-Bench-Science 0.1에서 52.6%를 기록했으며, 이는 Fable 5의 24.7%보다 높다고 보도했습니다.","category":"업계 동향","source":"Simon Willison 博客","aggregationSource":"Simon Willison 博客","pageTitle":"사이먼 윌리슨은 펠리컨 벤치마크인 클로드 파브르 5.1에서 다섯 가지 추론 수준을 테스트했습니다 - Aioga AI 뉴스","description":"사이먼 윌리슨은 클로드 파브르의 5.1 성능을 펠리컨 벤치마크에서 테스트했습니다. Anthropic은 이 모델이 Terminal-Bench-Science 0.1에서 52.6%를 기록했으며, 이는 Fable 5의 24.7%보다 높다고 보도했습니다.","url":"https://www.aioga.com/ko/news/cmtjczfy6038grobv3y19igkh/","contentTranslated":true,"sourceHash":"8cec8b19b7829f00","translatedAt":"2026-09-02T01:04:33.139Z"},"es":{"title":"Simon Willison probó cinco niveles de razonamiento en el benchmark Pelican Claude Fabre 5.1","summary":"Simon Willison puso a prueba el rendimiento de Claude Fabre, que era de 5,1, en el benchmark Pelican. Anthropic informa que el modelo obtuvo un 52,6% en Terminal-Bench-Science 0,1, más que el 24,7% de Fable 5.","category":"Industria","source":"Simon Willison 博客","aggregationSource":"Simon Willison 博客","pageTitle":"Simon Willison probó cinco niveles de razonamiento en el benchmark Pelican Claude Fabre 5.1 - Aioga Noticias de IA","description":"Simon Willison puso a prueba el rendimiento de Claude Fabre, que era de 5,1, en el benchmark Pelican. Anthropic informa que el modelo obtuvo un 52,6% en Terminal-Bench-Science 0,1,...","url":"https://www.aioga.com/es/news/cmtjczfy6038grobv3y19igkh/","contentTranslated":true,"sourceHash":"8cec8b19b7829f00","translatedAt":"2026-09-02T01:04:31.520Z"},"fr":{"title":"Simon Willison a testé cinq niveaux de raisonnement sur le benchmark Pelican Claude Fabre 5.1","summary":"Simon Willison a testé la performance de Claude Fabre à 5,1 sur le benchmark Pelican. Anthropic rapporte que le modèle a obtenu 52,6 % sur le Terminal-Bench-Science de 0,1, soit plus que les 24,7 % de Fable 5.","category":"Industrie","source":"Simon Willison 博客","aggregationSource":"Simon Willison 博客","pageTitle":"Simon Willison a testé cinq niveaux de raisonnement sur le benchmark Pelican Claude Fabre 5.1 - Aioga Actualités IA","description":"Simon Willison a testé la performance de Claude Fabre à 5,1 sur le benchmark Pelican. Anthropic rapporte que le modèle a obtenu 52,6 % sur le Terminal-Bench-Science de 0,1, soit pl...","url":"https://www.aioga.com/fr/news/cmtjczfy6038grobv3y19igkh/","contentTranslated":true,"sourceHash":"8cec8b19b7829f00","translatedAt":"2026-09-02T01:04:42.652Z"},"de":{"title":"Simon Willison testete fünf Argumentationsniveaus am Pelican-Benchmark Claude Fabre 5.1","summary":"Simon Willison testete Claude Fabres Leistung von 5,1 im Pelican-Benchmark. Anthropic berichtet, dass das Modell bei Terminal-Bench-Science 0,1 52,6 % erzielte, höher als Fable 5s 24,7 %.","category":"行业动态","source":"Simon Willison 博客","aggregationSource":"Simon Willison 博客","pageTitle":"Simon Willison testete fünf Argumentationsniveaus am Pelican-Benchmark Claude Fabre 5.1 - Aioga KI-News","description":"Simon Willison testete Claude Fabres Leistung von 5,1 im Pelican-Benchmark. Anthropic berichtet, dass das Modell bei Terminal-Bench-Science 0,1 52,6 % erzielte, höher als Fable 5s...","url":"https://www.aioga.com/de/news/cmtjczfy6038grobv3y19igkh/","contentTranslated":true,"sourceHash":"8cec8b19b7829f00","translatedAt":"2026-09-02T01:04:41.919Z"},"pt-BR":{"title":"Simon Willison testou cinco níveis de raciocínio no benchmark Pelican Claude Fabre 5.1","summary":"Simon Willison testou o desempenho de Claude Fabre, que estava de 5,1, no benchmark Pelican. A Anthropic relata que o modelo obteve 52,6% em Terminal-Bench-Science 0,1, acima dos 24,7% do Fable 5.","category":"行业动态","source":"Simon Willison 博客","aggregationSource":"Simon Willison 博客","pageTitle":"Simon Willison testou cinco níveis de raciocínio no benchmark Pelican Claude Fabre 5.1 - Aioga Notícias de IA","description":"Simon Willison testou o desempenho de Claude Fabre, que estava de 5,1, no benchmark Pelican. A Anthropic relata que o modelo obteve 52,6% em Terminal-Bench-Science 0,1, acima dos 2...","url":"https://www.aioga.com/pt-BR/news/cmtjczfy6038grobv3y19igkh/","contentTranslated":true,"sourceHash":"8cec8b19b7829f00","translatedAt":"2026-09-02T01:04:52.414Z"},"ru":{"title":"Саймон Уиллисон протестировал пять уровней рассуждения на бенчмарке Pelican Claude Fabre 5.1","summary":"Саймон Уиллисон протестировал результат Клода Фабра 5.1 на бенчмарке Pelican. Anthropic сообщает, что модель набрала 52,6% на Terminal-Bench-Science 0,1%, что выше, чем у Fable 5 — 24,7%.","category":"行业动态","source":"Simon Willison 博客","aggregationSource":"Simon Willison 博客","pageTitle":"Саймон Уиллисон протестировал пять уровней рассуждения на бенчмарке Pelican Claude Fabre 5.1 - Aioga Новости ИИ","description":"Саймон Уиллисон протестировал результат Клода Фабра 5.1 на бенчмарке Pelican. Anthropic сообщает, что модель набрала 52,6% на Terminal-Bench-Science 0,1%, что выше, чем у Fable 5 —...","url":"https://www.aioga.com/ru/news/cmtjczfy6038grobv3y19igkh/","contentTranslated":true,"sourceHash":"8cec8b19b7829f00","translatedAt":"2026-09-02T01:04:51.177Z"},"ar":{"title":"اختبر سايمون ويليسون خمسة مستويات منطقية على معيار بيليكان كلود فابر 5.1","summary":"اختبر سيمون ويليسون أداء كلود فابر 5.1 على مؤشر بيليكان. تفيد Anthropic أن النموذج حصل على تقييم 52.6٪ في Terminal-Bench-Science بنسبة 0.1، وهو أعلى من 24.7٪ في Fable 5.","category":"行业动态","source":"Simon Willison 博客","aggregationSource":"Simon Willison 博客","pageTitle":"اختبر سايمون ويليسون خمسة مستويات منطقية على معيار بيليكان كلود فابر 5.1 - Aioga أخبار الذكاء الاصطناعي","description":"اختبر سيمون ويليسون أداء كلود فابر 5.1 على مؤشر بيليكان. تفيد Anthropic أن النموذج حصل على تقييم 52.6٪ في Terminal-Bench-Science بنسبة 0.1، وهو أعلى من 24.7٪ في Fable 5.","url":"https://www.aioga.com/ar/news/cmtjczfy6038grobv3y19igkh/","contentTranslated":true,"sourceHash":"8cec8b19b7829f00","translatedAt":"2026-09-02T01:05:01.357Z"},"hi":{"title":"साइमन विलिसन ने पेलिकन बेंचमार्क क्लाउड फैब्रे 5.1 पर पांच तर्क स्तरों का परीक्षण किया","summary":"साइमन विलिसन ने पेलिकन बेंचमार्क पर क्लाउड फैबरे के 5.1 प्रदर्शन का परीक्षण किया। एंथ्रोपिक की रिपोर्ट है कि मॉडल ने टर्मिनल-बेंच-साइंस 52.6 पर 0.1% स्कोर किया, जो कल्पित 5 के 24.7% से अधिक है।","category":"行业动态","source":"Simon Willison 博客","aggregationSource":"Simon Willison 博客","pageTitle":"साइमन विलिसन ने पेलिकन बेंचमार्क क्लाउड फैब्रे 5.1 पर पांच तर्क स्तरों का परीक्षण किया - Aioga AI समाचार","description":"साइमन विलिसन ने पेलिकन बेंचमार्क पर क्लाउड फैबरे के 5.1 प्रदर्शन का परीक्षण किया। एंथ्रोपिक की रिपोर्ट है कि मॉडल ने टर्मिनल-बेंच-साइंस 52.6 पर 0.1% स्कोर किया, जो कल्पित 5 के 24.7...","url":"https://www.aioga.com/hi/news/cmtjczfy6038grobv3y19igkh/","contentTranslated":true,"sourceHash":"8cec8b19b7829f00","translatedAt":"2026-09-02T01:05:00.989Z"},"it":{"title":"Simon Willison ha testato cinque livelli di ragionamento sul benchmark Pelican Claude Fabre 5.1","summary":"Simon Willison ha testato la prestazione di Claude Fabre, 5.1, sul benchmark Pelican. Anthropic riporta che il modello ha ottenuto il 52,6% su Terminal-Bench-Science 0,1, superiore al 24,7% di Fable 5.","category":"行业动态","source":"Simon Willison 博客","aggregationSource":"Simon Willison 博客","pageTitle":"Simon Willison ha testato cinque livelli di ragionamento sul benchmark Pelican Claude Fabre 5.1 - Aioga Notizie IA","description":"Simon Willison ha testato la prestazione di Claude Fabre, 5.1, sul benchmark Pelican. Anthropic riporta che il modello ha ottenuto il 52,6% su Terminal-Bench-Science 0,1, superiore...","url":"https://www.aioga.com/it/news/cmtjczfy6038grobv3y19igkh/","contentTranslated":true,"sourceHash":"8cec8b19b7829f00","translatedAt":"2026-09-02T01:05:10.959Z"},"nl":{"title":"Simon Willison testte vijf redeneringsniveaus op de Pelican-benchmark Claude Fabre 5.1","summary":"Simon Willison testte Claude Fabre's 5,1 prestaties op de Pelican-benchmark. Anthropic meldt dat het model 52,6% scoorde op Terminal-Bench-Science 0,1, hoger dan Fable 5's 24,7%.","category":"行业动态","source":"Simon Willison 博客","aggregationSource":"Simon Willison 博客","pageTitle":"Simon Willison testte vijf redeneringsniveaus op de Pelican-benchmark Claude Fabre 5.1 - Aioga AI-nieuws","description":"Simon Willison testte Claude Fabre's 5,1 prestaties op de Pelican-benchmark. Anthropic meldt dat het model 52,6% scoorde op Terminal-Bench-Science 0,1, hoger dan Fable 5's 24,7%.","url":"https://www.aioga.com/nl/news/cmtjczfy6038grobv3y19igkh/","contentTranslated":true,"sourceHash":"8cec8b19b7829f00","translatedAt":"2026-09-02T01:05:10.345Z"},"tr":{"title":"Simon Willison, Pelican benchmark Claude Fabre 5.1'de beş mantık seviyesini test etti","summary":"Simon Willison, Pelican benchmarkında Claude Fabre'nin 5.1 performansını test etti. Anthropic, modelin Terminal-Bench-Science 0.1'de %52,6 puan aldığını, bu oranın Fable 5'in %24,7'sinin üzerinde olduğunu bildiriyor.","category":"行业动态","source":"Simon Willison 博客","aggregationSource":"Simon Willison 博客","pageTitle":"Simon Willison, Pelican benchmark Claude Fabre 5.1'de beş mantık seviyesini test etti - Aioga AI Haberleri","description":"Simon Willison, Pelican benchmarkında Claude Fabre'nin 5.1 performansını test etti. Anthropic, modelin Terminal-Bench-Science 0.1'de %52,6 puan aldığını, bu oranın Fable 5'in %24,7...","url":"https://www.aioga.com/tr/news/cmtjczfy6038grobv3y19igkh/","contentTranslated":true,"sourceHash":"8cec8b19b7829f00","translatedAt":"2026-09-02T01:05:20.240Z"},"vi":{"title":"Simon Willison đã kiểm tra năm mức suy luận trên chuẩn Pelican Claude Fabre 5.1","summary":"Simon Willison đã kiểm tra hiệu suất 5.1 của Claude Fabre trên bảng kiểm tra Pelican. Anthropic báo cáo rằng mô hình này đạt 52,6% trên Terminal-Bench-Science 0,1, cao hơn so với 24,7% của Fable 5.","category":"行业动态","source":"Simon Willison 博客","aggregationSource":"Simon Willison 博客","pageTitle":"Simon Willison đã kiểm tra năm mức suy luận trên chuẩn Pelican Claude Fabre 5.1 - Tin tức AI Aioga","description":"Simon Willison đã kiểm tra hiệu suất 5.1 của Claude Fabre trên bảng kiểm tra Pelican. Anthropic báo cáo rằng mô hình này đạt 52,6% trên Terminal-Bench-Science 0,1, cao hơn so với 2...","url":"https://www.aioga.com/vi/news/cmtjczfy6038grobv3y19igkh/","contentTranslated":true,"sourceHash":"8cec8b19b7829f00","translatedAt":"2026-09-02T01:05:20.635Z"},"id":{"title":"Simon Willison menguji lima tingkat penalaran pada acuan Pelican Claude Fabre 5.1","summary":"Simon Willison menguji performa 5,1 Claude Fabre pada tolok ukur Pelican. Anthropic melaporkan bahwa model ini meraih skor 52,6% pada Terminal-Bench-Science 0,1, lebih tinggi dari Fable 5 yang 24,7%.","category":"行业动态","source":"Simon Willison 博客","aggregationSource":"Simon Willison 博客","pageTitle":"Simon Willison menguji lima tingkat penalaran pada acuan Pelican Claude Fabre 5.1 - Berita AI Aioga","description":"Simon Willison menguji performa 5,1 Claude Fabre pada tolok ukur Pelican. Anthropic melaporkan bahwa model ini meraih skor 52,6% pada Terminal-Bench-Science 0,1, lebih tinggi dari...","url":"https://www.aioga.com/id/news/cmtjczfy6038grobv3y19igkh/","contentTranslated":true,"sourceHash":"8cec8b19b7829f00","translatedAt":"2026-09-02T01:05:30.096Z"},"th":{"title":"Simon Willison ทดสอบระดับเหตุผล 5 ระดับบนมาตรฐาน Pelican Claude Fabre 5.1","summary":"ไซมอน วิลลิสัน ทดสอบประสิทธิภาพ 5.1 ของโคลด ฟาเบร บนเครื่องทดสอบ Pelican Anthropic รายงานว่าโมเดลได้คะแนน 52.6% บน Terminal-Bench-Science 0.1 ซึ่งสูงกว่า Fable 5 ที่ 24.7%","category":"行业动态","source":"Simon Willison 博客","aggregationSource":"Simon Willison 博客","pageTitle":"Simon Willison ทดสอบระดับเหตุผล 5 ระดับบนมาตรฐาน Pelican Claude Fabre 5.1 - ข่าว AI Aioga","description":"ไซมอน วิลลิสัน ทดสอบประสิทธิภาพ 5.1 ของโคลด ฟาเบร บนเครื่องทดสอบ Pelican Anthropic รายงานว่าโมเดลได้คะแนน 52.6% บน Terminal-Bench-Science 0.1 ซึ่งสูงกว่า Fable 5 ที่ 24.7%","url":"https://www.aioga.com/th/news/cmtjczfy6038grobv3y19igkh/","contentTranslated":true,"sourceHash":"8cec8b19b7829f00","translatedAt":"2026-09-02T01:05:30.174Z"},"pl":{"title":"Simon Willison przetestował pięć poziomów rozumowania na benchmarku Pelican Claude Fabre 5.1","summary":"Simon Willison przetestował wynik Claude'a Fabre'a 5.1 na benchmarku Pelicana. Anthropic informuje, że model uzyskał 52,6% na Terminal-Bench-Science 0,1, co jest wyższym wynikiem niż Fable 5, który osiągnął 24,7%.","category":"行业动态","source":"Simon Willison 博客","aggregationSource":"Simon Willison 博客","pageTitle":"Simon Willison przetestował pięć poziomów rozumowania na benchmarku Pelican Claude Fabre 5.1 - Aioga Wiadomości AI","description":"Simon Willison przetestował wynik Claude'a Fabre'a 5.1 na benchmarku Pelicana. Anthropic informuje, że model uzyskał 52,6% na Terminal-Bench-Science 0,1, co jest wyższym wynikiem n...","url":"https://www.aioga.com/pl/news/cmtjczfy6038grobv3y19igkh/","contentTranslated":true,"sourceHash":"8cec8b19b7829f00","translatedAt":"2026-09-02T01:05:39.634Z"}},"evidenceTier":"verified-news","reviewStatus":"automated-ingest","indexable":true,"editorialCover":""}}