[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-mai-code-1-1-flash-copilot-moe-vision":3,"news-related-f8a3ead9-b1aa-4f6a-9146-19c01f7f1375":38},{"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},"f8a3ead9-b1aa-4f6a-9146-19c01f7f1375","微软把编码模型价格砍到四分之一：138B\u002F5B 稀疏 MoE 加原生视觉全量进入 Copilot","微软 8 月 11 日把 MAI-Code-1.1-Flash 推进 GitHub Copilot 全量通道。这款 138B 总参、5B 激活的稀疏 MoE 编码模型带 256K 上下文和原生视觉输入，列表价比前代低 73%（0.20\u002F0.02\u002F1.20 美元每百万 token），厂商自测 SWE-Bench Verified 从 71.6% 提到 72.6%，同时把单题 token 从 10.8K 压到 8.6K——这条曲线比分数本身更值得看。","# 微软把编码模型价格砍到四分之一：138B\u002F5B 稀疏 MoE 加原生视觉全量进入 Copilot\n\n## 背景：小尺寸编码模型的第二轮价格战\n\n2026 年 8 月 11 日，微软把 **MAI-Code-1.1-Flash** 推进 GitHub Copilot 的正式发布通道。这是继 MAI-Code-1-Flash（Build 2026 亮相）之后，微软 MAI 团队在「小尺寸编码模型」这条线上的迭代版本。按照 [GitHub 官方 changelog](https:\u002F\u002Fgithub.blog\u002Fchangelog\u002F2026-08-11-mai-code-1-1-flash-available-in-github-copilot\u002F) 的说法，新模型在编码质量、指令遵循、工具调用和性能四个维度都有改进，同时新增了原生视觉支持。\n\n值得注意的是同一天 GitHub 还发了另一条公告：MAI-Code-1-Flash 进入弃用倒计时，退役日期定在 2026 年 9 月 10 日。也就是说这不是并行两档，而是一次直接的代际替换。\n\n## 核心内容：参数、价格、以及那条 token 曲线\n\n按 LLM Stats 整理的模型卡数据，MAI-Code-1.1-Flash 的规格是：\n\n- **架构**：Transformer + 稀疏 MoE，**138B 总参数 \u002F 5B 激活参数**\n- **上下文**：256K tokens\n- **模态**：文本 + 图像输入，文本输出\n- **预训练截止**：2025 年 12 月；训练窗口为 2026 年 3 月至 8 月\n\n价格是这次发布最外显的变化。GitHub Copilot 列表价为输入 **0.20 美元\u002F百万 token**、缓存输入 **0.02 美元**、输出 **1.20 美元**；年付订阅按 **0.25× premium request** 计费。官方明确表述是「比 MAI-Code-1-Flash 低 73%」，归因于模型侧和服务侧的双重效率提升。\n\n厂商自测的 benchmark（在同一套 VS Code \u002F Copilot harness 下跑，属自报数据）更有意思：\n\n| 基准 | 1.1-Flash | token 消耗 | 1-Flash | token 消耗 |\n|---|---|---|---|---|\n| SWE-Bench Verified | 72.6% | 8.6K | 71.6% | 10.8K |\n| Terminal Bench 2.1 | 62.9% | 17.0K | 51.7% | 14.2K |\n\nSWE-Bench 上分数只涨了 1 个百分点，但**单题 token 从 10.8K 降到 8.6K**，降幅约 20%。发布文还引用了另一组数字：Copilot CLI 场景 Terminal-Bench 2.1 提升 22%、.NET 任务提升 15%、代码存活率提升 4%、回访率提升 9%，token 流式输出快 25%、每任务 token 少 25%。\n\n视觉能力是新增项，内部测试给出：Text2WebApp 通过率 74.1%、ScreenShot2WebApp 42.1%、Vision2Web Level3 仅 11.5%。\n\n## 评论：这次卖的不是分数，是单位成本\n\n如果只盯 SWE-Bench 的 72.6%，这次升级平淡无奇——一个百分点在当前榜单密度下几乎是噪声。但把「分数 \u002F token」这个比值拿出来看，故事就完全不同了：同样的题，少烧 20% 的 token，还便宜 73%，叠起来单位任务成本的下降是量级性的。\n\n这跟今年上半年整个行业的叙事转向是一致的。竞争焦点已经从「谁的能力排名更高」，滑向了「同等智能水平下谁更便宜」。微软给出的技术归因是训练侧的 RL 环境规模（模型卡称超过 15 万个 RL 环境，发布文说法是「数十万个 Copilot 内 RL 环境」）加上服务侧优化——本质上是用后训练的工程量去换推理时的经济性。\n\n**自适应解答长度控制**这个设计尤其值得留意：简单问题短输出、难题多给预算。这看起来像个小工程 trick，但它直接指向了 agentic coding 的核心痛点——大量 token 浪费在对简单任务的过度推理上。\n\n几个需要打折看待的地方：所有 benchmark 都是厂商自报、自家 harness 内跑的，未经第三方复现；模型卡上的定价栏写的是「待最终确定」，实际以 Copilot 定价页为准；Business \u002F Enterprise 用户需要管理员手动打开策略开关，默认关闭。Vision2Web Level3 只有 11.5% 也说明视觉编码在复杂场景下离可用还有距离。\n\n## 所以呢\n\n对于日常在 Copilot 里跑轻量编码循环、仓库问答、重构、工具调用 agent 的开发者，这是一次几乎无痛的降本升级——尤其是 0.25× 的 premium request 倍率，意味着同样的配额能跑四倍的请求量。\n\n更值得琢磨的是背后的趋势信号：当头部厂商开始把「每任务少花 25% token」写进发布文的核心卖点，说明模型能力的边际收益正在让位于推理经济性。下一代编码模型的竞争，可能不再是谁在 SWE-Bench 上多拿两分，而是谁能在同样的分数线上把成本再砍一半。","https:\u002F\u002Fgithub.blog\u002Fchangelog\u002F2026-08-11-mai-code-1-1-flash-available-in-github-copilot\u002F","998df6db-96e6-4b8e-8be1-cfa00a6cd177",[11,15,18,21],{"id":12,"name":13,"slug":13,"description":14,"color":14},"e82b2d09-81b2-43d1-977e-e018443b3c14","coding-agent",null,{"id":16,"name":17,"slug":17,"description":14,"color":14},"01598627-1ea6-4b27-a5d8-874971571a71","llm",{"id":19,"name":20,"slug":20,"description":14,"color":14},"7e89b5cc-57db-4f37-bc6d-28919a73931c","model-release",{"id":22,"name":23,"slug":23,"description":14,"color":14},"045c011e-e2bb-45ce-bdd6-0c927f8a3b87","token-efficiency",[25],{"id":26,"lang":27,"title":28,"summary":29,"content":30},"ea2ae4ec-b915-4b98-855d-07ca48b9743e","en","Microsoft quarters coding prices: 138B\u002F5B MoE into Copilot","On August 11, 2026, Microsoft rolled MAI-Code-1.1-Flash into general availability on GitHub Copilot. The 138B-total \u002F 5B-active sparse MoE coding model brings 256K context and native vision input at a list price 73% below its predecessor ($0.20\u002F$0.02\u002F$1.20 per million tokens). Vendor-reported SWE-Bench Verified rose from 71.6% to 72.6% while per-task tokens dropped from 10.8K to 8.6K — that curve matters more than the score.","# Microsoft Cuts Coding Model Price to a Quarter: 138B\u002F5B Sparse MoE with Native Vision Ships to Copilot\n\n## Background: The Second Price War in Small Coding Models\n\nOn August 11, 2026, Microsoft moved **MAI-Code-1.1-Flash** into general availability on GitHub Copilot. It succeeds MAI-Code-1-Flash (introduced at Build 2026) as the MAI team's iteration on the small-tier coding model line. According to the [official GitHub changelog](https:\u002F\u002Fgithub.blog\u002Fchangelog\u002F2026-08-11-mai-code-1-1-flash-available-in-github-copilot\u002F), the new model improves across coding quality, instruction following, tool use, and performance, while adding native vision support.\n\nNotably, GitHub published a second announcement the same day: MAI-Code-1-Flash enters its deprecation countdown, with retirement set for September 10, 2026. This is not a two-tier parallel offering — it is a direct generational replacement.\n\n## Core Details: Parameters, Pricing, and That Token Curve\n\nPer the model card data compiled by LLM Stats, MAI-Code-1.1-Flash specifications are:\n\n- **Architecture**: Transformer + sparse MoE, **138B total parameters \u002F 5B active**\n- **Context**: 256K tokens\n- **Modalities**: Text + image in, text out\n- **Pretrain cutoff**: December 2025; training window March–August 2026\n\nPricing is the most visible change. GitHub Copilot list rates are **$0.20 per million input tokens**, **$0.02 cached input**, and **$1.20 output**; annual subscribers are billed at a **0.25× premium request multiplier**. The official framing is «73% lower than MAI-Code-1-Flash», attributed to combined model-side and serving-side efficiency gains.\n\nThe vendor-reported benchmarks (run in the same VS Code \u002F Copilot harness, self-reported) are more interesting:\n\n| Benchmark | 1.1-Flash | Tokens | 1-Flash | Tokens |\n|---|---|---|---|---|\n| SWE-Bench Verified | 72.6% | 8.6K | 71.6% | 10.8K |\n| Terminal Bench 2.1 | 62.9% | 17.0K | 51.7% | 14.2K |\n\nSWE-Bench score rose by only one percentage point, but **per-task tokens fell from 10.8K to 8.6K**, roughly a 20% reduction. The launch post cites additional figures: +22% on Terminal-Bench 2.1 in Copilot CLI, +15% on .NET tasks, +4% code survival, +9% return visits, with tokens streaming 25% faster and 25% fewer tokens consumed per task.\n\nVision is the new capability. Internal tests report: Text2WebApp 74.1% pass rate, ScreenShot2WebApp 42.1%, and Vision2Web Level3 at just 11.5%.\n\n## Commentary: The Product Being Sold Is Unit Cost, Not Score\n\nFixating on the 72.6% SWE-Bench figure makes this look like a dull upgrade — one percentage point is close to noise at current leaderboard density. But look at the score-per-token ratio and the story inverts: the same problems, 20% fewer tokens burned, at 73% lower list price. Stacked together, the drop in per-task unit cost is order-of-magnitude.\n\nThis tracks the industry's narrative shift through the first half of the year. Competition has slid from «who ranks higher on capability» toward «who is cheaper at equivalent intelligence». Microsoft's stated technical attribution is RL environment scale on the training side (the model card cites over 150,000 RL environments; the launch post says «hundreds of thousands of RL environments in Copilot») plus serving-side optimization — essentially trading post-training engineering volume for inference-time economics.\n\n**Adaptive solution-length control** deserves particular attention: short outputs for easy asks, more budget for hard ones. It reads like a minor engineering trick, but it targets the core pain point of agentic coding — enormous token waste from over-reasoning on simple tasks.\n\nSeveral caveats warrant discounting: all benchmarks are vendor self-reported and run in the vendor's own harness, without third-party reproduction. The model card's pricing field reads «to be finalized»; the live Copilot pricing page is authoritative. Business and Enterprise users need an administrator to flip the policy, which is off by default. And Vision2Web Level3 at 11.5% signals that vision-grounded coding remains far from usable in complex scenarios.\n\n## So What?\n\nFor developers running lightweight coding loops, repo Q&A, refactoring, and tool-using agents inside Copilot day to day, this is a nearly frictionless cost reduction — the 0.25× premium request multiplier in particular means the same quota now covers four times the request volume.\n\nThe more interesting signal is structural. When a frontier vendor writes «25% fewer tokens per task» into the headline of a launch post, it indicates that marginal returns on raw capability are yielding to inference economics. The next round of coding model competition may not be about gaining two more points on SWE-Bench, but about halving cost at the same score line.","mai-code-1-1-flash-copilot-moe-vision","2026-08-12T06:00:00Z","2026-08-12T18:07:48.570498Z","2026-08-12T18:07:48.570507Z",true,"agent",73,{"items":39},[40,45,50,55,60,65],{"id":41,"title":42,"news_slug":43,"published_at":44},"89a79f9a-bfd2-4ebe-8f03-92fa74a3a34f","Ornith-1.5 开源：模型自己出题、自己搭考场，397B 到 9B 三档齐发","ornith-1-5-self-improvement-open-models","2026-08-20T13:30:00+00:00",{"id":46,"title":47,"news_slug":48,"published_at":49},"e846f1c6-4644-4f84-a664-83ec82734210","Meta Muse Spark 1.2 与 Muse Code 把「1.2 → 编程」的推理效率推回前沿","meta-muse-spark-12-coding-agent-54-index","2026-08-05T08:00:00+00:00",{"id":51,"title":52,"news_slug":53,"published_at":54},"dea38861-2618-4468-9bad-a18eea96a818","Base44 Base 1：年入 1.5 亿美元的 vibe-coding 平台，终于把自己的 LLM 训出来","base44-base-1-vibe-coding-llm-launch","2026-07-29T06:00:00+00:00",{"id":56,"title":57,"news_slug":58,"published_at":59},"1e7d0673-aecc-42b5-8560-92a2b4d4daf6","快手 KAT-Coder-V2.5 把 Agentic Coding 训练改写成基础设施工程","kuaishou-kat-coder-v2-5","2026-07-27T06:00:00+00:00",{"id":61,"title":62,"news_slug":63,"published_at":64},"6ba58314-305f-4255-83c5-87bdd1123b49","字节 Seed 2.1 押注「Agent-first」：模型自己参与训练，多模态重夺 SOTA","bytedance-seed-2-1-agent-first","2026-06-27T15:30:00+00:00",{"id":66,"title":67,"news_slug":68,"published_at":69},"6ceaf229-f1a2-4231-b531-797a99faa194","Ornith-1.0：模型自写 RL harness，SWE-Bench 比肩 Opus 4.7","ornith-1-0-397b-moe-swe-bench-opus-4-7","2026-06-26T18:01:01+00:00"]