Copilot Fusion 快 40%、成本降 77%，听着是挺猛。
我拿 Copilot 写脚本也快半年了，真省心的不是快，是改 bug 那点碎活儿它给接了。
工具再强，方向还得自己定。
xitter-newx.xyz-mikolabs.workers.dev/satyanadella/status/21…
We're redefining the SDLC for the agentic era.
Lots of GitHub updates this morning at Universe:
• GitHub Copilot Fusion delivers frontier-level quality, 40% faster and at up to 77% lower cost
• Our unified Copilot harness is now in Foundry, a managed service for building long-running agents that reason, use tools, and get work done
• Copilot Projects gives you an orchestration loop to build with multiple agents across repos
And we’re expanding capacity and strengthening GitHub’s architecture to make the home of agentic development more resilient.
1
51
Agent-written code is now the majority path to production for many teams. Review is the bottleneck. Who owns it?"
We're redefining the SDLC for the agentic era.
Lots of GitHub updates this morning at Universe:
• GitHub Copilot Fusion delivers frontier-level quality, 40% faster and at up to 77% lower cost
• Our unified Copilot harness is now in Foundry, a managed service for building long-running agents that reason, use tools, and get work done
• Copilot Projects gives you an orchestration loop to build with multiple agents across repos
And we’re expanding capacity and strengthening GitHub’s architecture to make the home of agentic development more resilient.
4
7
2,150
Agents now open more than a third of GitHub pull requests, by Satya Nadella's count. That makes code review the new control point. Every board should ask management two things: what share of merged code is agent-written, and what gate did it pass? If nobody knows, you have unreviewed production risk. What does your gate look like?
We're redefining the SDLC for the agentic era.
Lots of GitHub updates this morning at Universe:
• GitHub Copilot Fusion delivers frontier-level quality, 40% faster and at up to 77% lower cost
• Our unified Copilot harness is now in Foundry, a managed service for building long-running agents that reason, use tools, and get work done
• Copilot Projects gives you an orchestration loop to build with multiple agents across repos
And we’re expanding capacity and strengthening GitHub’s architecture to make the home of agentic development more resilient.
1
1
1,167
この3つが Universeの発表。
We're redefining the SDLC for the agentic era.
Lots of GitHub updates this morning at Universe:
• GitHub Copilot Fusion delivers frontier-level quality, 40% faster and at up to 77% lower cost
• Our unified Copilot harness is now in Foundry, a managed service for building long-running agents that reason, use tools, and get work done
• Copilot Projects gives you an orchestration loop to build with multiple agents across repos
And we’re expanding capacity and strengthening GitHub’s architecture to make the home of agentic development more resilient.
8
808
ついに、AIに「一人で長く働かせる」段階まで来た。
GitHubのCopilot Fusionは40%速くて最大77%安い。Copilot Projectsは、何体ものAIをまとめて動かす仕組みだ。
これから働く人に足りなくなるのは、手を動かす時間ではない。AIの仕事を見て、どれを通すか決める目のほうだ👀
We're redefining the SDLC for the agentic era.
Lots of GitHub updates this morning at Universe:
• GitHub Copilot Fusion delivers frontier-level quality, 40% faster and at up to 77% lower cost
• Our unified Copilot harness is now in Foundry, a managed service for building long-running agents that reason, use tools, and get work done
• Copilot Projects gives you an orchestration loop to build with multiple agents across repos
And we’re expanding capacity and strengthening GitHub’s architecture to make the home of agentic development more resilient.
2
102
GitHub Copilot 这次换了个卖法：不让你选模型，让它自己按任务选。
Universe 2026 上线的 Copilot Fusion，本质是个编排层，在 Anthropic 和 OpenAI 的模型之间挑：简单活给便宜模型，难活再升级，档位分 Standard、High、Max。
先看实测数字，Artificial Analysis 的 Coding Agent Index v2.0：
综合分：Fusion Max 60，对 Opus 5.5 的 Claude Code Max 62，只差 2 分
单任务成本：5.59 美元，对 11.42 美元，差不多一半
每任务 token：720 万，约是后者的一半
但别只看综合分。分项里 Fusion Max 在 Terminal-Bench 4.0 拿了 71 分领先，FrontierCode 却只有 51 分，落后于 Opus 5.5 的 56 分。
所以这个产品的真实卖点不是“全面超过最强模型”，而是“接近最强表现，但成本砍半”。对每天跑大量常规任务的团队，这个账比榜单第一更有用；对最难的那类仓库级任务，还是得看分项，别被综合分带偏。
你会把这种自动挑模型交给它吗？还是宁可自己手动锁定一个模型？
We're redefining the SDLC for the agentic era.
Lots of GitHub updates this morning at Universe:
• GitHub Copilot Fusion delivers frontier-level quality, 40% faster and at up to 77% lower cost
• Our unified Copilot harness is now in Foundry, a managed service for building long-running agents that reason, use tools, and get work done
• Copilot Projects gives you an orchestration loop to build with multiple agents across repos
And we’re expanding capacity and strengthening GitHub’s architecture to make the home of agentic development more resilient.
2
2
167
you're paying 2x for 2 points.
copilot fusion max: 60 on the artificial analysis coding agent index.
claude code with opus 5.5 max: 62.
github says fusion costs about half the api money per task.
fusion is a router. for every prompt it picks the models and one of three patterns:
→ single: one model does the job
→ cascade: a cheap model drafts, a judge accepts it or escalates to a stronger one
→ critique: one model drafts, several others review it in parallel, then it rewrites
satya's own numbers: frontier-level quality, 40% faster, up to 77% lower cost. there's no fusion fee, you pay each model it calls at its normal rate, and paid plans get 30% off until nov 30.
one line from github's docs worth knowing: if fusion throws away a draft, the file edits that draft already made stay in your repo. review before you commit.
my take: a team burning thousands a month on agent runs won't feel 2 points, but it will feel half the bill. for the one bug that eats your whole week, i'd still pay for the 62.
would you swap claude code for half the price and 2 points less, or is the last 2 points the whole point?
We're redefining the SDLC for the agentic era.
Lots of GitHub updates this morning at Universe:
• GitHub Copilot Fusion delivers frontier-level quality, 40% faster and at up to 77% lower cost
• Our unified Copilot harness is now in Foundry, a managed service for building long-running agents that reason, use tools, and get work done
• Copilot Projects gives you an orchestration loop to build with multiple agents across repos
And we’re expanding capacity and strengthening GitHub’s architecture to make the home of agentic development more resilient.
23
29
566
😎💫
We're redefining the SDLC for the agentic era.
Lots of GitHub updates this morning at Universe:
• GitHub Copilot Fusion delivers frontier-level quality, 40% faster and at up to 77% lower cost
• Our unified Copilot harness is now in Foundry, a managed service for building long-running agents that reason, use tools, and get work done
• Copilot Projects gives you an orchestration loop to build with multiple agents across repos
And we’re expanding capacity and strengthening GitHub’s architecture to make the home of agentic development more resilient.
10
533
GitHub Copilot 来刷存在感了，感觉微软做的 AI 产品很多功能都不太实用。“Copilot”已经是个过时的概念，AI 不再是副驾驶，我们希望 AI 去自动驾驶。
日常让 Copilot 审查代码，审查的结果有价值的不多，Office Copilot 也是个半成品，上次想加个定时任务都不行，整体功能和体验比起 Claude Code 和 Codex 差一个量级，即使它能用 Opus 5.5。
--- 下面是这次 GitHub Copilot 更新的总结 ---
GitHub Copilot 能自动挑模型了，免费版也能用；最仔细的新代码审查平均一次约 10 美元
GitHub 在 10 月 28 日的年度大会 Universe 上发布了一批 Copilot 新功能。开发者每天会碰到的变化有两个：Copilot 会自己挑 AI 模型；提交代码后，Copilot 能审得更细，还能自动修好卡住合并的问题。
先说计费：Copilot 按“AI 点数”扣费，1 点等于 1 美分，每个套餐每月送一笔，10 美元一个月的 Pro 送 1500 点。
【Copilot 替你挑模型】
以前要自己在菜单里选 GPT 还是 Claude，选强的贵，选便宜的怕搞不定。新功能 Fusion 按难度安排：简单活交给一个模型；难一点的先让便宜模型写草稿，不过关再换强模型；最难的让一个模型写、几个模型同时挑错。
Fusion 今天正式上线，VS Code、JetBrains、命令行和 Copilot 桌面应用都能用。所有套餐（含免费版）都有基础档，更深的 High、Max 档要 39 美元一个月的 Pro+ 或更高套餐。Fusion 不另收费，用了哪些模型就按原价扣点数，高档位一次可能叫好几个模型，扣得更多。11 月 30 日前付费用户打七折。
GitHub 说，在 Artificial Analysis 的编程评测里，Fusion Max 只比 Claude Code 搭配 Opus 5.5 Max 低 2 分，成本约一半。这是 GitHub 自己公布的成绩。
【代码审查和合并】
开发者改完代码要提交“合并请求”（PR），审过才能并进正式代码。Copilot 代码审查新加了最仔细的 Max 档：几家公司的模型各审一遍，再汇总、剔除误报，适合改登录逻辑这类出错代价大的代码。
Max 档还在公开测试，Pro 及以上套餐可用。它不便宜：官方估算平均每次约 10 美元，最多 25 美元，原有的快速档约 0.25 美元。按 Pro 每月 1500 点算，审一次 Max 就用掉三分之二。
Copilot 现在还能正式“批准”PR，算进仓库要求的审批人数，但默认关闭，要管理员打开。PR 页面也多了“Watch and fix”按钮：自动测试没过或代码有冲突时，Copilot 在云端自己去修，付费套餐可用。
【其他】
GitHub Enterprise 企业客户今天起不用单独买 Copilot 座位，按实际用量付费，目前要找销售开通。Copilot 桌面应用新增“项目”功能（公开测试），给一个目标，就能跨多个代码仓库同时派多个 AI agent 干活。
原文：github.blog/news-insights/co…
Fusion：github.blog/changelog/2026-1…
代码审查 Max 与批准：github.blog/changelog/2026-1…
审查价格：docs.github.com/en/copilot/c…
We're redefining the SDLC for the agentic era.
Lots of GitHub updates this morning at Universe:
• GitHub Copilot Fusion delivers frontier-level quality, 40% faster and at up to 77% lower cost
• Our unified Copilot harness is now in Foundry, a managed service for building long-running agents that reason, use tools, and get work done
• Copilot Projects gives you an orchestration loop to build with multiple agents across repos
And we’re expanding capacity and strengthening GitHub’s architecture to make the home of agentic development more resilient.
Had a lot of fun bringing Copilot Projects to life together with @david_dossett, @pierceboggan and others 🚀
We're redefining the SDLC for the agentic era.
Lots of GitHub updates this morning at Universe:
• GitHub Copilot Fusion delivers frontier-level quality, 40% faster and at up to 77% lower cost
• Our unified Copilot harness is now in Foundry, a managed service for building long-running agents that reason, use tools, and get work done
• Copilot Projects gives you an orchestration loop to build with multiple agents across repos
And we’re expanding capacity and strengthening GitHub’s architecture to make the home of agentic development more resilient.
Hire the robot, fire the human and don’t look back. 🙄
We're redefining the SDLC for the agentic era.
Lots of GitHub updates this morning at Universe:
• GitHub Copilot Fusion delivers frontier-level quality, 40% faster and at up to 77% lower cost
• Our unified Copilot harness is now in Foundry, a managed service for building long-running agents that reason, use tools, and get work done
• Copilot Projects gives you an orchestration loop to build with multiple agents across repos
And we’re expanding capacity and strengthening GitHub’s architecture to make the home of agentic development more resilient.
Agenetic economy is taking over.
GitHub recorded 7.38 B commits in Sept-2026, more than 5 times the figure a year earlier.
And up to now this year, GitHub has added 4M CPU cores, 160PB of high-speed storage and 50Tb/s of network capacity to absorb a flood of agent-written code.
That curve already outran GitHub’s hardware once, when a traffic surge knocked out authentication, Actions, pull requests and Copilot for 7 hours and 47 minutes on 17 August.
GitHub is now moving to a multi-region Azure deployment, due to finish early next year, and rebuilding Git to decouple storage from compute.
We're redefining the SDLC for the agentic era.
Lots of GitHub updates this morning at Universe:
• GitHub Copilot Fusion delivers frontier-level quality, 40% faster and at up to 77% lower cost
• Our unified Copilot harness is now in Foundry, a managed service for building long-running agents that reason, use tools, and get work done
• Copilot Projects gives you an orchestration loop to build with multiple agents across repos
And we’re expanding capacity and strengthening GitHub’s architecture to make the home of agentic development more resilient.
$MSFT just removed the separate copilot seat charge for licensed github enterprise cloud users on its new sales led plan. the included ai credits go too. chat and agents start billing from the first request. that's a bet on usage, with less money tied to assigning a seat. handing everyone access is easy. getting them to spend the ai budget is the part to watch.
We're redefining the SDLC for the agentic era.
Lots of GitHub updates this morning at Universe:
• GitHub Copilot Fusion delivers frontier-level quality, 40% faster and at up to 77% lower cost
• Our unified Copilot harness is now in Foundry, a managed service for building long-running agents that reason, use tools, and get work done
• Copilot Projects gives you an orchestration loop to build with multiple agents across repos
And we’re expanding capacity and strengthening GitHub’s architecture to make the home of agentic development more resilient.
3
1,197