[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-microsoft-gpt5-6-default-token-budget":3,"news-related-c2ee2a09-d001-4740-9820-21fb672eee8b":41},{"id":4,"title":5,"summary":6,"content":7,"original_url":8,"source_id":9,"tags":10,"translations":27,"news_slug":34,"published_at":35,"created_at":36,"modified_at":37,"is_published":38,"publish_type":39,"image_url":14,"view_count":40},"c2ee2a09-d001-4740-9820-21fb672eee8b","Copilot 默认模型切到 GPT-5.6 Sol：tokenmaxxing 终结","微软执行副总裁 Jay Parikh 8 月 4 日内部邮件确认,GitHub Copilot 内部默认模型已从 Anthropic Claude 切换到 OpenAI GPT-5.6 Sol,7 月起各部门启用 AI token 预算目标。404 Media 首报,CNBC 当日跟进。这不是一家公司的财务动作,而是大模型竞争从'拼能力'转向'拼性价比'的标志性节点——一线工程师的 token 账单开始反向倒逼前沿模型选择。","## 微软叫停 tokenmaxxing:GitHub Copilot 默认模型从 Claude 切到 GPT-5.6 Sol,各部门戴上 AI token 预算紧箍咒\n\n8 月 4 日,微软执行副总裁、CoreAI 工程负责人 Jay Parikh 在一封内部邮件里给微软 11 万工程师画了一条线:**「tokenmaxxing 不是我们优化的目标」**(原始报道见 404 Media,CNBC 当日跟进)。邮件落地的两项硬动作,直接把大厂 AI 军备竞赛拉回到成本纪律上:\n\n1. **GitHub Copilot 内部默认模型从 Anthropic Claude 切换到 OpenAI GPT-5.6 Sol**。微软允许员工继续选用 Anthropic、Google、Moonshot AI、xAI 等其它模型,但 Parikh 要求「大部分时间默认用 GPT-5.6 Sol」。\n2. **自 2026 年 7 月起,各业务部门拿到 AI token 预算目标**(AI token budget target),员工可以在个人账单里看到月度 token 支出。CoreAI 暂未给单个团队或员工设独立预算,但 Parikh 在邮件里写明:「像管理其他所有关键资源一样管理 token 花费。」\n\n## 为什么是 GPT-5.6 Sol,而不是更便宜的中国开源模型\n\n表面看,微软选择 GPT-5.6 Sol 是因为「比其他模型更便宜」,但底层逻辑是 IP 套利。CNBC 在报道里点出:这是 9 个月前 OpenAI 完成公司重组、把对微软的知识产权授权延长到 2032 年的实际兑现——既然持有 OpenAI 的 IP 权益,自然把内部负载向 OpenAI 模型倾斜。\n\n微软发言人对 CNBC 的官方表态是:「我们把 OpenAI GPT-5.6 Sol 设为微软内部 GitHub Copilot 的默认模型,同时继续提供多种可选模型,工程师随时可以切换。」这等于把 Anthropic Claude 从「默认位置」降级为「可选位置」——尽管微软去年 11 月还宣布向 Anthropic 投资最高 50 亿美元、Anthropic 同时承诺在 Azure 上花 300 亿美元,关系没有破裂,但优先级已经悄悄重排。\n\n值得注意的是,微软并没有直接换成 DeepSeek V4 Flash 这类更激进的低成本选项。CNBC 的报道强调,微软的选择是「便宜的前沿旗舰」,不是「最便宜的模型」——这反映大厂内部的一个微妙平衡:**能力门槛不能掉,只能在能力相当的池子里挑成本最优**。\n\n## Tokenmaxxing 是怎么失控的:一份硅谷账单样本\n\n把视野放大,这只是 token 通胀崩盘前最后一波企业级反应。量子位整理了同期各家公司的账单细节(均为公开报道口径):\n\n- **Atlassian**:月度 AI 支出在不到一年内增长到原来的三倍,超过 1500 万美元,公司随后设置支出限制,要求员工控制模型使用成本。\n- **Uber**:整个公司到 4 月就烧光了 2026 全年的 AI 编程预算,Cursor 给 Uber 的新一轮报价上涨到此前 4 到 5 倍。Uber 随后规定每名员工使用每种 Agent 编程工具每月不得超过 1500 美元。\n- **亚马逊**:内部搞了 KiroRank 排行榜,按员工在 Kiro 平台上的 AI 使用量打分,引发刷榜;月底排行榜被关停,高级副总裁 Dave Treadwell 提醒:「请不要为了使用 AI 而使用 AI。」\n- **OpenAI 内部**:一名员工曾一周处理 2100 亿 tokens,「足够把整个维基百科填满 33 遍」。\n- **Meta**:内部出现名为 Claudeonomics 的员工自建排行榜,统计 8.5 万多名员工的 AI 使用情况,颁发「Token Legend」「Cache Wizard」虚拟称号;Meta 同期把 AI 使用纳入绩效考核,Shopify 也跟进。\n\n黄仁勋在 3 月的 All-In Podcast 公开站台:「如果一个年薪 50 万美元的工程师每年消耗的 token 不到 25 万美元,我会深感担忧。」GTC 2026 上他更进一步,愿意在工程师几十万年薪之外再提供相当于一半年薪的 token 预算。这套言论被大厂内部消化成 KPI,反过来把 token 烧穿。\n\n## 模型选择经济学:从「token 等于智能」到「智能边际 = token 边际」\n\nParikh 在邮件里没有否定 AI 价值,但给出了一个清晰的经济学判断:**生产率提升的边际收益,必须匹配 token 的边际成本**。不是每个问题都值得调用最强、最贵的前沿模型。模型跑得更久、上下文塞得更满、Agent 开得更多,也不等于最终成果更好。\n\nCNBC 引述的内部数据显示,许多工程师每月 token 花费在「几百到几千美元之间」,高消费工程师一个月烧掉数千美元已成常态。Wall Street 对四家 hyperscaler(微软、亚马逊、Alphabet、Meta)2026 年合计 7000 亿美元资本支出开始要回报,最新季度 Microsoft 现金生成同比下降 23%,Amazon 和 Alphabet 自由现金流甚至转负——这才是 token 紧箍咒的真正上层驱动。\n\n## 一线影响:GitHub Copilot 的隐性「代码风格」会怎么变\n\n对开发者来说,微软把默认模型从 Claude 切到 GPT-5.6 Sol,意味着两件事:\n\n第一,代码补全和 Copilot Chat 的「默认味道」会迁移。Claude 系模型在长上下文推理、复杂重构上有口碑,换成 GPT-5.6 Sol 后,日常补全体验可能会有可感知的差异——尤其是涉及多文件编辑、test generation、PR review 这类任务。\n\n第二,Anthropic Claude 不再是「打开就能用」的路径。CNBC 报道里写明,微软同时收紧了对 Anthropic Claude 的内部访问——尽管没有完全禁用,但默认位置让位意味着工程师要主动切换才会用到。这与微软去年宣布的 50 亿美元 Anthropic 投资在叙事上形成张力:投资归投资,内部预算归内部预算。\n\n## 行业含义:大模型竞争从「能力排行」转向「同等智能成本」\n\n把这条新闻放回大模型竞争主线上看,微软的默认模型切换不是孤立事件——它呼应了 7 月底 OpenAI 对 GPT-5.6 系列降价 20%–80%、DeepSeek V4 Flash 0731 用 50 分能力把单任务成本压到 GPT-5.6 Luna 的 35% 这一连串动作。华泰证券在 8 月初的研报里把这一轮概括为:大模型竞争由能力排名进一步转向「同等智能成本」。\n\n对模型厂商而言,这是一道新的考题:能不能在能力相当的前提下,把单 token 成本压到客户愿意「默认选用」的水位。对企业 IT 而言,这意味着 LLM 选型从「哪个最强」开始转向「哪个最值」。对一线工程师而言,token 紧箍咒才刚戴上——以后跑模型前得多想一秒,「这笔 token 真的能换回等价的产出吗」。\n\n微软这一动作落地,某种程度上是 tokenmaxxing 时代的正式句号。下一步看的不是模型会不会更强,而是 Wall Street 愿不愿意继续为「更强」付溢价。\n\n---\n\n**参考资料**\n- 404 Media:Microsoft Tells Engineers 'Tokenmaxxing Is Not What We Are Optimizing For' — https:\u002F\u002Fwww.404media.co\u002Fmicrosoft-tells-engineers-tokenmaxxing-is-not-what-we-are-optimizing-for\u002F\n- CNBC:Microsoft makes OpenAI GPT-5.6 Sol default in GitHub Copilot for staff — https:\u002F\u002Fwww.cnbc.com\u002F2026\u002F08\u002F05\u002Fmicrosoft-makes-openai-gpt-5point6-sol-default-in-github-copilot-for-staff.html\n- 量子位:微软叫停 Tokenmaxxing!预算卡死,超限自负 — https:\u002F\u002Fwww.qbitai.com\u002F2026\u002F08\u002F466739.html","https:\u002F\u002Fwww.404media.co\u002Fmicrosoft-tells-engineers-tokenmaxxing-is-not-what-we-are-optimizing-for\u002F","e06537c4-1c62-46c4-a4ac-d28107bbca86",[11,15,18,21,24],{"id":12,"name":13,"slug":13,"description":14,"color":14},"7ac06d8e-b074-4147-abfc-ffaa4c6b8744","ai-efficiency",null,{"id":16,"name":17,"slug":17,"description":14,"color":14},"baf131c1-687a-49f4-87f6-4dd87c1c692f","gpt",{"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},"42e59a88-7795-47dc-a334-ef1e72c24347","openai",{"id":25,"name":26,"slug":26,"description":14,"color":14},"95d7995a-fddb-47ba-b8e6-e976ac65414b","strategy",[28],{"id":29,"lang":30,"title":31,"summary":32,"content":33},"7ad77d89-66ec-4f7e-a31e-504613de647c","en","Copilot defaults to GPT-5.6 Sol: tokenmaxxing ends","In an August 4 internal memo, Microsoft EVP Jay Parikh told engineers 'tokenmaxxing is not what we are optimizing for.' GitHub Copilot's default model moves from Anthropic Claude to OpenAI GPT-5.6 Sol, and every Microsoft division now operates under an AI token budget target first introduced in July 2026. 404 Media broke the memo, CNBC confirmed with additional context the next day. This isn't a corporate cost-control footnote — it's the clearest signal yet that frontier-model selection is being driven by per-token economics rather than pure capability rankings.","## Microsoft pulls the plug on tokenmaxxing: GitHub Copilot's default moves to GPT-5.6 Sol, every division gets an AI token budget\n\nOn August 4, Microsoft EVP and CoreAI engineering lead Jay Parikh drew a line in an internal memo to Microsoft's 110,000 engineers: **'tokenmaxxing is not what we are optimizing for.'** The memo — first reported by 404 Media and followed up the next day by CNBC with additional sourcing — lands two hard policy moves that drag the hyperscaler AI arms race back to cost discipline:\n\n1. **GitHub Copilot's internal default model switches from Anthropic Claude to OpenAI GPT-5.6 Sol.** Engineers can still pick Anthropic, Google, Moonshot AI, xAI, or Microsoft's own models, but Parikh asks them to 'default to GPT-5.6 Sol most of the time.'\n2. **From July 2026, every Microsoft division operates under an AI token budget target.** Individual employees can see their monthly token spend in personal billing dashboards. CoreAI has not yet set per-team or per-engineer budgets, but Parikh writes: 'Manage token spend the way you manage every other critical resource.'\n\n## Why GPT-5.6 Sol, and not a cheaper Chinese open-weight model\n\nOn the surface Microsoft picked GPT-5.6 Sol because it is 'cheaper to use than other models.' The deeper logic is IP arbitrage. CNBC points out: this is the cash-in on the IP-rights extension Microsoft secured nine months ago when OpenAI completed its corporate restructuring and pushed Microsoft's intellectual property rights through 2032. Holding those rights naturally biases internal workloads toward OpenAI models.\n\nMicrosoft's official line to CNBC: 'We have set OpenAI's GPT-5.6 Sol as the default for Microsoft's internal use of GitHub Copilot while continuing to offer a range of model options that can be selected at any time by our engineers.' This quietly downgrades Anthropic Claude from 'the default' to 'an option' — even though Microsoft announced up to a  billion investment in Anthropic last November, and Anthropic committed 0 billion in Azure spend in return. The relationship isn't broken, but the priority order has shifted.\n\nCrucially, Microsoft did not pick the cheapest possible model. CNBC's reporting emphasizes the choice is 'cheap frontier flagship,' not 'cheapest available.' That's a subtle but real enterprise balance: **the capability floor cannot drop — cost optimization happens inside a comparable-quality pool.**\n\n## How tokenmaxxing went off the rails: a Silicon Valley invoice sampler\n\nZoom out and this is just the last wave of corporate reaction before the token-inflation bubble pops. QbitAI's Chinese coverage compiles the concurrent bills from public reporting — Atlassian, Uber, Amazon, OpenAI and Meta all surfaced in the same window:\n\n- **Atlassian** saw monthly AI spend triple in under a year to over 5 million, then imposed usage caps and asked engineers to control model costs.\n- **Uber** burned through its entire 2026 AI coding budget by April. Cursor's renewal quote to Uber rose to 4–5x the previous level. Uber now caps each engineer's spend per Agent coding tool at ,500 per month.\n- **Amazon** ran an internal leaderboard called KiroRank scoring engineers on Kiro platform usage. It triggered rank-chasing and was shut down by month-end, with SVP Dave Treadwell reminding staff: 'Don't use AI just to use AI.'\n- **OpenAI internally**: one employee reportedly processed 210 billion tokens in a single week — 'enough to fill Wikipedia 33 times over.'\n- **Meta**: an employee-built leaderboard called 'Claudeonomics' tracked AI usage across 85,000+ employees, complete with 'Token Legend' and 'Cache Wizard' virtual titles. Meta added AI usage to performance reviews; Shopify followed.\n\nJensen Huang poured fuel on the fire in March on the All-In Podcast: 'If a 00K-a-year engineer isn't consuming 50K of tokens a year, I'd be very worried.' At GTC 2026 he went further, saying he'd happily fund token budgets equal to half an engineer's salary on top of their base pay — because those tokens could 10x an engineer's leverage. Big-tech internalization of those comments turned into KPIs, and the spend blew through.\n\n## Model-selection economics: from 'tokens equal intelligence' to 'intelligence marginal = token marginal'\n\nParikh's memo doesn't reject AI value. It states a clean economic judgment: **the marginal return on productivity must match the marginal cost of tokens.** Not every problem deserves the strongest, most expensive frontier model. Longer runs, fuller context windows, more Agents running in parallel — none of that guarantees better outcomes.\n\nCNBC cites internal Microsoft data: many engineers spend anywhere from 'hundreds to a few thousand dollars a month' on tokens. Spending several thousand dollars per month per heavy user is now normal. Wall Street is starting to demand returns on the roughly 00 billion in collective 2026 capex from Microsoft, Amazon, Alphabet and Meta. Microsoft's free cash flow fell 23% year over year in the most recent quarter; Amazon's and Alphabet's went outright negative. That's the real upper-level driver behind the token caps.\n\n## Day-one impact for engineers: how the 'taste' of Copilot changes\n\nFor developers, switching GitHub Copilot's default from Claude to GPT-5.6 Sol changes two things in practice:\n\nFirst, the 'default flavor' of completions and Copilot Chat shifts. Claude-family models have a reputation for long-context reasoning and complex refactors; switching to GPT-5.6 Sol produces noticeable differences in daily completion experience, especially on multi-file edits, test generation and PR review.\n\nSecond, Anthropic Claude is no longer the zero-friction path. CNBC reports that Microsoft also tightened internal access to Anthropic Claude — not a full block, but the default slot is gone, so engineers have to actively switch to use it. That creates narrative tension with Microsoft's  billion Anthropic investment announced last year: investments stay, but internal budgets shift.\n\n## Industry signal: LLM competition moves from 'capability rank' to 'intelligence-per-dollar'\n\nIn the wider LLM competition, Microsoft's default switch isn't an isolated event. It echoes the late-July GPT-5.6 family price cut (20%–80%), and DeepSeek V4 Flash 0731 hitting 50 on the Artificial Analysis Intelligence Index at 35% of GPT-5.6 Luna's per-task cost. Huatai Securities' early-August research note summarized the cycle: LLM competition is moving from capability ranking to 'intelligence at equivalent cost.'\n\nFor model vendors this is a new exam question: at comparable capability, can you push per-token cost low enough that enterprise IT will default to you? For enterprise IT it means LLM selection now starts from 'which is the best value' rather than 'which is the strongest.' For engineers, the token budget has only just landed — every model call now needs a half-second of thought: 'will this token actually buy equivalent output?'\n\nMicrosoft's move lands as the formal closing bracket of the tokenmaxxing era. What comes next isn't whether models keep getting stronger — it's whether Wall Street will keep paying the premium for 'stronger.'\n\n---\n\n**References**\n- 404 Media: 'Microsoft Tells Engineers Tokenmaxxing Is Not What We Are Optimizing For' — https:\u002F\u002Fwww.404media.co\u002Fmicrosoft-tells-engineers-tokenmaxxing-is-not-what-we-are-optimizing-for\u002F\n- CNBC: 'Microsoft makes OpenAI GPT-5.6 Sol default in GitHub Copilot for staff' — https:\u002F\u002Fwww.cnbc.com\u002F2026\u002F08\u002F05\u002Fmicrosoft-makes-openai-gpt-5point6-sol-default-in-github-copilot-for-staff.html\n- QbitAI (量子位): '微软叫停Tokenmaxxing!预算卡死,超限自负' — https:\u002F\u002Fwww.qbitai.com\u002F2026\u002F08\u002F466739.html","microsoft-gpt5-6-default-token-budget","2026-08-08T08:00:00Z","2026-08-08T10:04:42.894098Z","2026-08-19T01:48:03.231362Z",true,"agent",119,{"items":42},[43,48,53,58,63,68],{"id":44,"title":45,"news_slug":46,"published_at":47},"b9e635eb-ac5e-412d-8904-f113ad3fd5ec","微软宣布工程师 AI token 预算上限并把 OpenAI GPT-5.6 Sol 设为 GitHub Copilot 内部默认模型","microsoft-copilot-gpt-5-6-sol-default-token-budget-0806","2026-08-05T16:00:00+00:00",{"id":49,"title":50,"news_slug":51,"published_at":52},"418a9ac0-18fd-49a4-b7a8-d29d1c1ba497","AI 承诺的四天工作制为什么没来：OpenAI \u002F Anthropic 内部工时真相","ai-four-day-work-week-myth-openai-90-hours","2026-08-16T03:30:00+00:00",{"id":54,"title":55,"news_slug":56,"published_at":57},"31f3215c-0892-419d-a610-fe815cc60bbe","GPT-5.6 降价 80% 把竞争拉进「同等智能成本」：DeepSeek V4 Flash 接招，国产模型卡出双线赛道","gpt-5-6-luna-price-cut-equal-intelligence-cost","2026-08-12T03:00:00+00:00",{"id":59,"title":60,"news_slug":61,"published_at":62},"d95940eb-69c1-467e-9d60-5886ab71d985","GPT-5.6-Cyber 上线、Daybreak 分层、Astra 推迟:OpenAI 把\"网络安全模型\"做成一个独立产品线","openai-gpt-5-6-cyber-daybreak-astra-2026","2026-08-11T04:00:00+00:00",{"id":64,"title":65,"news_slug":66,"published_at":67},"d1c7b405-fe4e-40f9-9249-a12e2bba6913","GPT-5.6 八月更新：把「推理强度滑块」下放给 Plus\u002FPro，同时把免费用户拉进 Luna 时代","openai-gpt-5-6-august-update-reasoning-slider","2026-08-10T20:00:00+00:00",{"id":69,"title":70,"news_slug":71,"published_at":72},"3967306f-062a-41a6-ab58-f99e70fc0e68","AISI 122 轮 cyber eval 越界：OpenAI 与 Anthropic 同日披露","aisi-mythos-5-gpt-5-6-cyber-eval-incident-2026","2026-08-08T04:00:00+00:00"]