[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-microsoft-copilot-gpt-5-6-sol-default-token-budget-0806":3,"news-related-b9e635eb-ac5e-412d-8904-f113ad3fd5ec":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},"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","微软执行副总裁 Jay Parikh 8月4日内部邮件披露两大调整:自7月起各部门设AI token预算目标,工程师月token支出在数百到数千美元之间;GitHub Copilot 内部默认从Anthropic Claude 切到OpenAI GPT-5.6 Sol。直接对标tokenmaxxing转向valuemaxxing,与Meta \u002F Amazon \u002F Atlassian \u002F Citi 同步。","## 微软把\"最大化使用 AI\"换成\"最大化每次 token 的产出\"\n\n8月4日,微软执行副总裁、CoreAI 工程负责人 Jay Parikh 给员工发了一封内部邮件,核心一句话:\"Tokenmaxxing is not what we are optimizing for.\"翻译过来就是——别再把 token 烧得越多当 KPI 了。\n\n这是一个明确的姿态切换。配合这封邮件,微软同时做了两件具体的事:\n\n- **GitHub Copilot 内部默认模型从 Anthropic Claude 切到 OpenAI GPT-5.6 Sol**(2026年7月发布,定位是性价比导向的旗舰版),原因是 OpenAI 拿的是微软早年的投资,在知识产权定价上有优势\n- **从 2026 年 7 月起,微软各部门设\"AI token 预算目标\"(AI token budget target)**,员工可以追踪个人 token 支出,各业务线按月汇报\n\nCNBC 报道(8月5日)引用 Parikh 的原文:\"As we accelerate our use of GitHub Copilot to deliver on our goals, we all need to be aware of how we consume tokens.\"印度快报同日跟进,披露数字范围:\"许多工程师每月 token 支出在数百到数千美元之间\",目前没有公开的硬性美元上限,但\"未来可能进一步收紧\"。\n\n## 微软不是孤例——tokenmaxxing 时代在退潮\n\n把视角拉远一点,这其实是 2026 年下半年整个企业 AI 采购转向的标志性事件。Indian Express(8月5日)在综合 404 Media 原始报道的同时,梳理了同行清单:\n\n- **Meta**——已经下架了内部的\"Claudeonomics\"排行榜(原本鼓励员工比赛烧 token)\n- **Google**——5月I\u002FO上 Pichai 公开承认\"很多公司已经把年 token 预算烧光了\",并把 Gemini 3.5 Flash 定位为 off-ramp\n- **Amazon \u002F Adobe \u002F Atlassian \u002F Citi**——均在重新收紧内部 AI 支出\n- **Uber**——COO Andrew Macdonald 5月明确表示,公司花了 4 个月烧完 2026 年 AI 预算,\"高 token 用量没有带来成比例的消费者功能增长\"\n\nIBM 在 6月就提前喊出了\"valuemaxxing\"这个词,提倡把指标从\"用了多少 token\"换到\"完成了多少任务 \u002F 节省了多少时间 \u002F 修复了多少漏洞\"。404 Media 8月4日的报道总结得很直接:\"This makes Microsoft one of the last major companies to rein in its employees' expensive AI use.\"\n\n## 为什么会发生——三股力量同时转向\n\n**第一,AI 使用率被错误等同于生产力。** 多家公司的内部 leaderboard 直接造成\"烧 token = 努力工作\"的扭曲信号,工具被过度用于\"本来不需要 frontier 模型\"的简单任务。Parikh 写的话很直白:\"shifting more workloads to OpenAI models helps us get greater value from our token investment.\"这是承认:对相当一部分 GitHub Copilot 任务来说,Anthropic Claude 的旗舰价是过度的。\n\n**第二,GPT-5.6 Sol 这类\"次旗舰\"模型已经够用。** 微软从 Anthropic 默认切到 OpenAI,不是站队,是采购成本优化。GPT-5.6 Sol 在 Artificial Analysis Intelligence Index 上仍能拿到 50 分上下,而单任务成本只有 Claude 旗舰的约三分之一(华泰证券 8月4日研报数据)。当模型能力梯度被价格梯度拉开后,默认模型的选型就回到了一个古老命题——什么任务用多少钱的算力最划算。\n\n**第三,token 经济学的二阶效应。** 印度快报引述了 tech newsletter Sources 的关键背景:GitHub Copilot 在转为按 token 计费之前,毛利率长期为负。也就是说,微软自家卖 Copilot 给别人时已经发现 token 经济学撑不住内部规模化使用。这是把\"产品定价学到的东西\"反哺到\"内部消费\"的典型例子。\n\n## 这次切换给行业留下的 3 个开放问题\n\n1. **模型厂商的客户结构会重新洗牌。** OpenAI 在微软这种深度绑定客户里被\"钦点为默认\",是因为它有 IP 定价让步;对于没有这种关系的厂商,默认设置变化意味着销量直接下降。Anthropic、xAI、Mistral 在其他大客户里会不会也吃到同样的挤压?值得观察 8月内的财报口径。\n\n2. **FinOps 工具链会从\"成本中心\"变成\"生产力基础设施\"。** AI 时代之前的 FinOps(云成本管理)主要是给 CFO 看的,这一波 token 预算从财务口径被翻译成工程师个人 KPI,意味着计费面板、用量提示、prompt 优化建议都会被前移进 IDE。IBM 把这条线包装成\"agentic development platform\"(IBM Bob),核心卖点是\"intelligent model routing\"。\n\n3. **\"valuemaxxing\"的指标体系还没人做出标准答案。** IBM 提了一组候选指标(完成任务数 \u002F 节省时间 \u002F 修复漏洞数 \u002F 避免返工),但缺少跨公司的可比基线。IDC 预测 2028 年 70% 的领先 AI 驱动企业会\"动态管理多模型路由\"——前提是先有可比的 outcome 指标,这件事目前各家还在自家实验。\n\n参考资料:\n- CNBC(8月5日):Microsoft AI exec tells developers to default to OpenAI top model\n- 404 Media(8月4日):Microsoft Tells Engineers Tokenmaxxing Is Not What We Are Optimizing For\n- Indian Express(8月5日):Why Microsoft is cracking down on tokenmaxxing despite strong earnings\n- IBM Think(6月25日):Tokenmaxxing is dead, long live valuemaxxing\n- AI Weekly(8月5日):Microsoft caps engineer AI token spend, names GPT-5.6 default","https:\u002F\u002Fwww.cnbc.com\u002F2026\u002F08\u002F05\u002Fmicrosoft-makes-openai-gpt-5point6-sol-default-in-github-copilot-for-staff.html","6493f82a-2404-45ca-b616-e0a265514cb6",[11,15,18,21,24],{"id":12,"name":13,"slug":13,"description":14,"color":14},"5e628969-6d2a-437f-998a-104e4b16cfb1","ai-progress",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},"c1240582-7c06-4c52-b325-2a7db1e232f3","en","Microsoft caps engineer AI tokens, makes GPT-5.6 Sol the default","Microsoft EVP Jay Parikh told employees in an August 4 memo that tokenmaxxing is not what we are optimizing for, then made two concrete moves: from July 2026 every Microsoft division gets an AI token budget target, and the internal default model in GitHub Copilot switches from Anthropic Claude to OpenAI GPT-5.6 Sol. Engineers' individual token bills reportedly run from hundreds to a few thousand dollars a month.","## Microsoft trades \"use as much AI as you can\" for \"maximize the value of every token\"\n\nOn August 4, Microsoft EVP of CoreAI engineering Jay Parikh sent an internal memo with one blunt sentence: \"Tokenmaxxing is not what we are optimizing for.\" The signal is unmistakable — the era of using raw token consumption as a productivity proxy at one of the world's largest software companies is over.\n\nTwo concrete moves ship with the memo:\n\n- **GitHub Copilot's internal default model switches from Anthropic Claude to OpenAI GPT-5.6 Sol.** GPT-5.6 Sol launched in July 2026 and is positioned as a value-oriented flagship. Parikh's stated reason: Microsoft has early-stage IP rights from its investment in OpenAI, so the per-token economics work in Microsoft's favor.\n- **Every Microsoft division gets an AI token budget target starting July 2026.** Individual employees can track their own spend, and each business line reports monthly. According to Indian Express, engineer-level spend is in the hundreds to a few thousand dollars a month today; no hard dollar cap has been published, but Parikh signaled that further restrictions may come.\n\nCNBC (August 5) quotes Parikh directly: \"As we accelerate our use of GitHub Copilot to deliver on our goals, we all need to be aware of how we consume tokens.\" The framing is deliberate — Parikh is explicit that Microsoft still wants to be \"AI-first,\" just not at any token price.\n\n## Microsoft is not the first — tokenmaxxing is being walked back across the industry\n\nStep back and this is the canonical 2026-H2 corporate AI procurement story. Indian Express (August 5) compiled the industry map on top of 404 Media's original reporting:\n\n- **Meta** has already shut down the internal \"Claudeonomics\" leaderboard, which previously gamified engineer token burn.\n- **Google**, in May's I\u002FO keynote, Sundar Pichai publicly conceded that many companies had burned through their annual token budgets, and positioned Gemini 3.5 Flash as the off-ramp.\n- **Amazon, Adobe, Atlassian and Citi** are all tightening internal AI spend.\n- **Uber** — COO Andrew Macdonald said in May that the company burned through its 2026 AI budget in four months, and that higher token usage has not produced proportional consumer feature gains.\n\nIBM flagged this shift back in June with the term **valuemaxxing**: instead of counting tokens used, measure tasks completed, developer time saved, vulnerabilities resolved, rework avoided. 404 Media's August 4 piece is unambiguous: \"This makes Microsoft one of the last major companies to rein in its employees' expensive AI use.\"\n\n## Why now — three forces converged\n\n**1. AI usage was mistakenly equated with productivity.** Internal leaderboards at several companies made token burn a proxy for effort. Parikh's line lands on that head-on: \"shifting more workloads to OpenAI models helps us get greater value from our token investment.\" Translation — for a meaningful share of GitHub Copilot tasks, Anthropic's flagship pricing was overkill.\n\n**2. The second-tier flagships are good enough.** Microsoft switching defaults from Anthropic to OpenAI is not a vendor politics move, it is procurement math. GPT-5.6 Sol still scores around 50 on the Artificial Analysis Intelligence Index, and per Huatai Securities' August 4 research note, its per-task cost is roughly a third of Claude's flagship. When model capability and price diverge, the default-model question reverts to the oldest question in infrastructure — which task deserves which tier of compute.\n\n**3. Token economics bit Microsoft in its own product.** Indian Express cites a tech newsletter called Sources for a critical fact: GitHub Copilot ran at strongly negative gross margins before it moved to usage-based billing earlier this year. Microsoft learned the lesson selling Copilot to outside customers, and is now applying it internally — the textbook case of product economics feeding back into internal consumption.\n\n## Three open questions the industry is now sitting with\n\n1. **Model vendor revenue mix will reshuffle.** OpenAI gets the \"default\" slot at a deep Microsoft account because of its IP pricing concession. Vendors without that kind of relationship will see real revenue impact when similar default changes land at other large customers. Watch Anthropic, xAI and Mistral in the August earnings cycle.\n\n2. **FinOps tools move from a cost line into the developer IDE.** Pre-AI FinOps was a CFO problem. Token budgets that show up as engineer-level KPIs mean usage meters, prompt suggestions and routing choices are pushed all the way into the editor. IBM is packaging exactly this bet as the \"agentic development platform\" IBM Bob, with intelligent model routing as the headline feature.\n\n3. **No one has the standard valuemaxxing metric set yet.** IBM floated a candidate list (tasks completed, time saved, vulnerabilities resolved, rework avoided) but cross-company comparability is missing. IDC's prediction that 70% of leading AI-driven enterprises will dynamically manage multi-model routing by 2028 is contingent on comparable outcome metrics first — and that piece is still being worked out inside each company.\n\n## Bottom line\n\nThis is the first time a hyperscaler with a deep OpenAI relationship publicly admits that frontier model pricing inside its own developer tooling was miscalibrated. The default switch and the budget target are not just an internal Microsoft story — they are a signal to every model vendor and every enterprise AI buyer that the post-tokenmaxxing era is starting with concrete procurement changes, not just opinion pieces.\n\nReferences:\n- CNBC (August 5): Microsoft AI exec tells developers to default to OpenAI top model as part of efficiency push\n- 404 Media (August 4): Microsoft Tells Engineers Tokenmaxxing Is Not What We Are Optimizing For\n- Indian Express (August 5): Why Microsoft is cracking down on tokenmaxxing despite strong earnings\n- IBM Think (June 25): Tokenmaxxing is dead, long live valuemaxxing\n- AI Weekly (August 5): Microsoft caps engineer AI token spend, names GPT-5.6 default","microsoft-copilot-gpt-5-6-sol-default-token-budget-0806","2026-08-05T16:00:00Z","2026-08-06T09:53:22.901299Z","2026-08-06T09:53:22.901306Z",true,"agent",174,{"items":42},[43,48,53,58,62,67],{"id":44,"title":45,"news_slug":46,"published_at":47},"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":49,"title":50,"news_slug":51,"published_at":52},"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":54,"title":55,"news_slug":56,"published_at":57},"c2ee2a09-d001-4740-9820-21fb672eee8b","Copilot 默认模型切到 GPT-5.6 Sol：tokenmaxxing 终结","microsoft-gpt5-6-default-token-budget","2026-08-08T08:00:00+00:00",{"id":59,"title":60,"news_slug":60,"published_at":61},"4afbd081-d5dd-4595-80f8-dd72c69a136a","GPT-5.5 让模型在发布前先改自己跑的引擎：这不是新模型,是 OpenAI 的 release 范式更新","2026-08-03T18:00:00+00:00",{"id":63,"title":64,"news_slug":65,"published_at":66},"0fd9ee7a-5b8f-49d2-9032-57f763de40e3","OpenAI 下一代模型 Astra 一口气破解 10 个数学难题:从 27 年未决的非 sofic 群到 46 年未动的高维球体堆积","openai-astra-ten-math-proofs-2026","2026-08-01T10:00:00+00:00",{"id":68,"title":69,"news_slug":70,"published_at":71},"c8220c45-cac5-468d-b89e-8d0ed088a2d8","Stripe 75 亿美元拿下 OpenRouter:LLM 接入层被支付平台收编了","stripe-openrouter-llm-routing-acquired","2026-08-21T00:00:00+00:00"]