#001 大模型进入“斩杀线”大战，Token 就是新货币

过去一周，新模型密集到几乎没有人能全部试完。GLM 5.3 Flash、Qwen 3.8 Flash Next、MiniMax H3 等新发布，把竞争从“谁更强”推向新的斩杀线：能力是否够用、成本是否足够低，以及普通人能否真正获得体验。 当模型差距越来越难靠榜单感知，免费试用、真实任务与原生 Agent 组合正在成为新的选择逻辑。我们也从 DeepSeek Harness、OpenClaw、WorkBuddy、Grok Bot 和 Omarchy 出发，讨论 Agent 为什么需要一台永不下线的电脑，以及 MCP、Skill、Connector 和 API Key 是否还应该暴露给用户。 本期录制于 2026 年 9 月 3 日，新闻窗口为 8 月 27 日至 9 月 3 日。节目中的观点与预测属于参与者个人判断；产品能力、价格与数据以录制时的公开信息为准。 本期人物：杨攀、歸藏、橘子（冯雷）、向阳乔木（乔向阳） Over the past week, so many new models arrived that almost no one could try them all. Releases including GLM 5.3 Flash, Qwen 3.8 Flash Next, and MiniMax H3 are pushing the competition beyond “who is strongest” toward a new viability threshold: whether a model is capable enough, affordable enough, and genuinely accessible to ordinary users. As benchmark differences become harder to feel, free access, real-world tasks, and native model-agent combinations are becoming a new way to choose. Starting with DeepSeek Harness, OpenClaw, WorkBuddy, Grok Bot, and Omarchy, we also ask why an agent needs an always-on computer and whether MCP, Skills, Connectors, and API keys should still be exposed to users. This episode was recorded on September 3, 2026, covering news from August 27 through September 3. The views and predictions expressed are the participants’ own. Product capabilities, pricing, and data reflect publicly available information at the time of recording. Featuring: Yang Pan, Guizang, Orange (Feng Lei), and Xiangyang Qiaomu (Qiao Xiangyang) 00:00 节目片头 / Intro 00:48 开场：一周密集发布的模型清单 / Opening: a crowded week of model releases 04:27 GLM 5.3 Flash 与 MiniMax H3，便宜模型为何更容易进入真实使用 / GLM 5.3 Flash and MiniMax H3: why inexpensive models reach real use sooner 16:49 Harness 评测：完成率、Token 成本与“斩杀线” / Harness evaluation: completion rate, token cost, and the kill line 28:28 当所有模型都差不多，用户该如何选择 / How users choose when every model feels similar 36:54 原生模型加原生 Agent，会不会成为默认组合 / Will first-party models plus first-party agents become the default? 50:43 WorkBuddy、Qoder、OpenClaw 与 DeepSeek Harness / WorkBuddy, Qoder, OpenClaw, and DeepSeek Harness 58:24 Grok Bot：为什么 Agent 需要一台永不下线的云电脑 / Grok Bot and why an agent needs an always-on cloud computer 01:06:43 Microduck 为什么一夜走红 / Why Microduck became popular overnight 01:12:59 云端 Agent 与本地 Memory，能否两者兼得 / Can cloud agents and local memory coexist? 01:18:11 收购、ARR 与 Claude 套餐争议：本周商业新闻快扫 / Acquisitions, ARR, and Claude plan disputes: a rapid business-news round 01:23:27 Token Is New Money：买 Mac Studio 像在家里“铸币” / Token Is New Money: a Mac Studio as a mint at home 01:31:48 AI 应用市场回暖了吗？利润空间比需求更现实 / Is the AI application market recovering? Margins are more concrete than demand 01:35:29 为什么 AI 应用团队仍然跑不过模型厂商 / Why AI application teams still struggle to outrun model companies 01:41:08 设计工程师、全栈闭环与“每个人都是 OPC” / Design engineers, full-stack ownership, and “everyone as an OPC” 01:44:10 片尾 / Outro 相关资料 / Related links: GLM-5.3-Flash：https://autoclaw.z.ai/blog/model/glm-... Qwen3.8-Flash-Next：https://qwen.ai/blog?id=qwen3.8-flash... MiniMax H3：https://www.minimax.io/news/minimax-h... DeepSeek Harness：https://github.com/deepseek-ai/deepse... WorkBuddy：https://open.workbuddy.cn/ Grok Bot：https://x.ai/news/introducing-grok-bot Next Token | 词元之外是一档关注 AI 技术、产品、创业与创作的视频播客。 Next Token | 词元之外 is a video podcast about AI technology, products, startups, and creative work. 节目主页：https://nexttoken.tv Show website: https://nexttoken.tv 社区：  / discord   Community:   / discord   录制场地合作：象外空间（北京朝阳·来广营福熙大道）。感谢象外空间为本期节目提供录制场地与支持。 Venue partner: Xiangwai Space (Fuxi Avenue, Laiguangying, Chaoyang, Beijing). Thanks to Xiangwai Space for providing the recording venue and support for this episode. 本期片头与片尾配乐经剪辑和混音后使用： “Electric Dreams” by Scott Buckley — released under CC BY 4.0. www.scottbuckley.com.au “Origami” by Scott Buckley — released under CC BY 4.0. www.scottbuckley.com.au Edited and mixed for Next Token Weekly. #人工智能 #大模型 #ArtificialIntelligence #LargeLanguageModels #Agent #NextToken