[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-dualentry-accounting-benchmark-argon":3,"topics-all":38,"news-related-9cba46ac-2667-4ba5-8d9a-41c59fb5da47":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},"9cba46ac-2667-4ba5-8d9a-41c59fb5da47","会计基准首轮:Gemini 4 Argon 91.1%居首,开源GLM-5.3 Flash并列第二","DualEntry 首轮会计 AI 基准：18 个模型跑 101 个案例各三轮。Gemini 4 Argon 准确率 91.1%、一致性 86.1% 双第一；开源 GLM-5.3 Flash 并列第二（88.1%），单任务 0.6 美分约为 Argon 十九分之一。应收账款与月末结账平均仅 64.4%。","会计被认为是 LLM 落地最难的垂直领域之一：数字要精确到分，借贷方向不能反，而且同一个问题问三遍得给同一个答案。纽约的 AI-native ERP 公司 DualEntry 刚刚交出了第一份完整答卷——他们把 18 个前沿模型放进同一套会计环境里各跑三轮，用的是同一个测试公司账本、同一套工具和同一份指令，评分不靠 AI 评委，全部对着预先写好的参考答案做确定性校验，80% 及格、三次里过两次才算通过。\n\n## 首轮榜单：头部挤成一团\n\n结果头部非常挤。Google 的 Gemini 4 Argon 以 91.1% 准确率排第一；并列第二是智谱的 GLM-5.3 Flash 和 SpaceXAI 的 Grok 4.7，同为 88.1%，只差榜首 3 个百分点。Anthropic 目前最好的 Claude Fable 5.1 拿了 85.1%，OpenAI 最好的 GPT-6 Astra \u002F GPT-6.1 Sol 是 81.2%。再往后，DeepSeek V4.1 Flash 与 Kimi K3 同为 84.2%，Meta 的 Muse Wonder 83.2%，共有十二个模型挤在 80.2%–85.1% 的五分区间里。\n\n值得单独划重点的是：并列第二的 GLM-5.3 Flash 是开放权重模型。开源阵营在通用榜单上追闭源追了两年，在\"记账\"这种具体活儿上已经能跟 Grok 4.7 打平，离 Argon 只差 3 个点。\n\n## 真正拉开差距的是一致性\n\nDualEntry 单独报告了\"一致性\"——同一个任务三次运行全部通过的比例。这一项把准确率相同的模型拉开了：Grok 4.7 准确率并列第二，一致性只有 74.3%；Claude Fable 5.1 准确率排第四，一致性却有 80.2%。Argon 是唯一两项都第一的模型（86.1%）。DualEntry 联合创始人 Benedict Dohmen 的原话是：\"一个分录这次做对、下次做错的模型，还上不了账本。Argon 是我们测过最准也是最稳的，而财务团队两样都要。\"\n\n对做财务自动化的团队，这个指标的分量可能超过准确率本身——账是不能掷骰子的。\n\n## 成本能差 400 倍\n\n榜单另一个刺眼的维度是价格。DeepSeek V4.1 Flash 84.2% 的准确率，单任务成本 0.3 美分；Claude Fable 5.1 只高它不到 1 个点，单任务 1.18 美元，约 400 倍。GLM-5.3 Flash 更极端：88.1% 的并列第二，单任务只要 0.6 美分，约为 Argon（11.5 美分）的十九分之一。\n\n当然，便宜有代价：GLM-5.3 Flash 平均一道题要 57 秒、输出 4.9k token；Argon 只要 18 秒、0.7k token。快模型用\"少想直接答\"换速度，在延迟敏感的交互场景里是真实差异。\n\n## 集体软肋：应收账款和月末结账\n\n八个测试领域里，模型在交易分类上平均 94.4%，会计知识多选题全体满分，但在应收账款和月末结账上平均只有 64.4%，应收账款单项分数从 41.7% 一路散到 91.7%。单项冠军各有归属：银行对账是 Muse Wonder（91.7%），月末结账是 Kimi K3（91.7%），而 Argon 在银行对账上只有 75.0%。难度分层上，Argon、Grok 4.7、Gemini 3.8 Flash 在 hard 案例上并列最好（82.1%）；Claude Opus 5.5 和 Mistral Large 4 从 easy 到 hard 掉了 30.3 个点，是全场跌幅最大的两家。\n\n## 所以呢\n\n这份基准的价值不在\"谁第一\"，而在它测的东西离真实业务足够近：101 个合成案例全在会计系统里用工具完成，不靠 AI 评 AI。两个信号值得记下：一是头部差距已经缩到 3 个点，选型逻辑正从\"谁最准\"切换到\"谁最稳、谁最便宜\"；二是开源模型在垂直任务上正式挤进第一梯队——0.6 美分一题的并列第二，对任何要规模化跑分类、对账、审单的管线来说都不是小数。\n\n基准全文与 DualBench100 方法论在 [dualentry.com\u002Fbenchmark](https:\u002F\u002Fwww.dualentry.com\u002Fbenchmark) 公开（2026 年 10 月 13 日版次，Version 1.0），含全部 18 个模型的成本、时长与一致性明细。","https:\u002F\u002Fwww.dualentry.com\u002Fbenchmark","70056da9-f119-49d3-a85e-2515861a2ff9",[11,15,18,21],{"id":12,"name":13,"slug":13,"description":14,"color":14},"120fa59a-ff6f-4537-9bf5-f818df636a0e","benchmark",null,{"id":16,"name":17,"slug":17,"description":14,"color":14},"a9524a82-a7c5-4daa-bb4b-a7ee77bb0b94","gemini",{"id":19,"name":20,"slug":20,"description":14,"color":14},"01598627-1ea6-4b27-a5d8-874971571a71","llm",{"id":22,"name":23,"slug":23,"description":14,"color":14},"b9bd9039-fcdb-41a8-b85b-fc1587def2b9","open-source",[25],{"id":26,"lang":27,"title":28,"summary":29,"content":30},"fc118226-0d70-47f4-a437-34289038056d","en","Gemini 4 Argon tops accounting benchmark, open GLM ties 2nd","Gemini 4 Argon tops DualEntry's accounting AI benchmark at 91.1%; open-weight GLM-5.3 Flash ties Grok 4.7 for second at 88.1%, at $0.006 per task.","Accounting has long been considered one of the hardest verticals for LLMs: numbers must be exact to the cent, debit and credit directions cannot flip, and asking the same question three times should yield the same answer. DualEntry, a New York-based AI-native ERP company, has just published its first full answer — it ran 18 frontier models through the same accounting environment three times each, using an identical test-company ledger, the same tools and the same instructions. Grading is deterministic against pre-written reference answers, with no AI judge: an 80% score passes, and a task counts only if passed in at least two of three runs.\n\n## A tightly packed leaderboard\n\nThe top of the table is crowded. Google's Gemini 4 Argon ranks first at 91.1% accuracy; tied for second are Zhipu AI's GLM-5.3 Flash and SpaceXAI's Grok 4.7 at 88.1%, just three points behind. Anthropic's best, Claude Fable 5.1, scored 85.1%; OpenAI's best, GPT-6 Astra \u002F GPT-6.1 Sol, came in at 81.2%. Behind them, DeepSeek V4.1 Flash and Kimi K3 both sit at 84.2%, Meta's Muse Wonder at 83.2% — twelve models in total packed into the 80.2%–85.1% band.\n\nOne detail deserves emphasis: GLM-5.3 Flash, tied for second, is an open-weight model. After two years of open models chasing closed ones on general leaderboards, on the concrete job of bookkeeping an open model now ties Grok 4.7 and sits three points from Argon.\n\n## Consistency is the real divider\n\nDualEntry reports a separate \"consistency\" metric — the share of tasks a model passes in all three runs. This is where models with equal accuracy separate: Grok 4.7, tied for second on accuracy, managed only 74.3% consistency; Claude Fable 5.1, fourth on accuracy at 85.1%, reached 80.2%. Argon is the only model leading both at 86.1%. As DualEntry co-founder Benedict Dohmen put it: \"A model that gets a journal entry right once and wrong the next time isn't ready for the books. Argon was the most accurate model we tested and the most consistent, and finance teams need both.\"\n\nFor teams building finance automation, this metric may matter more than accuracy itself — books cannot be a dice roll.\n\n## A 400x cost spread\n\nThe other glaring dimension is price. DeepSeek V4.1 Flash scored 84.2% at $0.003 per task; Claude Fable 5.1 scored under one point higher at $1.18 per task — roughly 400 times the cost. GLM-5.3 Flash is more extreme still: 88.1% accuracy, tied second, at $0.006 per task, about one-nineteenth of Argon's $0.115.\n\nCheap has a price, though: GLM-5.3 Flash averages 57 seconds and 4.9k output tokens per task, versus 18 seconds and 0.7k for Argon. The fast model trades \"think less, answer now\" for speed — a real difference in latency-sensitive interactive settings.\n\n## Shared weakness: receivables and month-end close\n\nAcross the eight task areas, models averaged 94.4% on transaction classification and every model aced the multiple-choice accounting-knowledge tasks. But on accounts receivable and month-end close they averaged just 64.4%, with AR scores scattering from 41.7% to 91.7%. Category champions vary: Muse Wonder leads bank reconciliation (91.7%), Kimi K3 leads month-end close (91.7%), while Argon manages only 75.0% on reconciliation. On hard cases, Argon, Grok 4.7 and Gemini 3.8 Flash share the best result at 82.1%; Claude Opus 5.5 and Mistral Large 4 drop 30.3 points from easy to hard, the steepest falls in the field.\n\n## So what\n\nThe value of this benchmark is not \"who is number one\" but how close it stays to real work: all 101 synthetic cases are completed with tools inside an actual accounting system, with no AI grading AI. Two signals are worth noting. First, the gap at the top has shrunk to three points, so model selection is shifting from \"who is most accurate\" to \"who is most consistent and cheapest\". Second, open-weight models have formally entered the top tier on vertical tasks — a tied-second model at $0.006 per task is not a small number for any pipeline running classification, reconciliation or document review at scale.\n\nThe full benchmark and the DualBench100 methodology are public at [dualentry.com\u002Fbenchmark](https:\u002F\u002Fwww.dualentry.com\u002Fbenchmark) (October 13, 2026 release, Version 1.0), with cost, duration and consistency detail for all 18 models.","dualentry-accounting-benchmark-argon","2026-10-15T13:10:00Z","2026-10-15T13:12:51.825257Z","2026-10-15T13:12:51.825270Z",true,"agent",135,[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},"aa718f3c-14cf-4984-8762-794eaaee8c16","MAI-Gym:9B 智能体在 2 万个并行环境里做在线 RL","mai-gym-gui-agent-online-rl-9b","2026-10-21T09:30:00+00:00",{"id":65,"title":66,"news_slug":67,"published_at":68},"bb6e0f7b-3899-4e45-a267-bfa512d470b0","FDE-Gym:95个真实交付任务,最强Agent只过14个","fde-gym-agent-delivery-benchmark","2026-10-18T23:08:06+00:00",{"id":70,"title":71,"news_slug":72,"published_at":73},"d529114c-bfd7-4efb-9af4-444d3065b397","Step 5 Preview 开源,2.7 美元\u002F百万输出","stepfun-step-5-preview-600b-moe","2026-10-17T21:12:04+00:00",{"id":75,"title":76,"news_slug":77,"published_at":78},"26801a16-3f57-4bd8-adad-3b49a1d28f08","OmniCompass基准:开源模型音视频理解仍落后","omnicompass-omni-modal-video-grounding-benchmark","2026-10-16T15:10:00+00:00",{"id":80,"title":81,"news_slug":82,"published_at":83},"5a5efb41-7037-46cb-b93e-6842f9ff2d62","UltraBERT:91.0 GLUE+47.7 BEIR,BERT 配方重审后 encoder 双榜登顶","ultrabert-encoder-dual-lm-heads","2026-10-14T17:10:59+00:00",{"id":85,"title":86,"news_slug":87,"published_at":88},"206ea36a-1eea-463f-a241-1e3b32f5ec2d","USTC GraphForge:证据图把任务和 rubric 钉在一起,Qwen3.6-27B 涨三基准","graphforge-ustc-qwen36-27b-evidence-graph","2026-10-04T03:05:00+00:00"]