GitHub kodlash vazifalari uchun AI modellarini va bajarish naqshlarini avtomatik tanlaydigan runtime model routeri va orkestratori — Copilot Fusion’ni ishga tushirdi.
Har bir qadam uchun model tanlashni ishlab chiquvchilardan talab qilish o‘rniga, Fusion fikrlash, kod generatsiyasi, nosozliklarni tuzatish va ko‘rib chiqishni muvofiqlashtiradi va yakunda yagona, izchil javob taqdim etadi.
Bu funksiya barcha Copilot rejalarida mavjud. U Visual Studio Code, Copilot CLI, GitHub Copilot ilovasi va JetBrains IDE’larida umumiy foydalanish uchun tayyor. GitHub hujjatlariga ko‘ra, model tanlovi obuna huquqlari va tashkilot siyosatlariga bog‘liq qoladi.
Ishlab chiquvchilar Fusion’ni model tanlovchisi orqali faollashtiradi. Tizim har bir so‘rovni fikrlash, kodlash, nosozliklarni tuzatish va vosita ishlatish qobiliyat signallari yordamida baholaydi, so‘ngra tanlangan harakat darajasiga mos bajarish strategiyasini tanlaydi.
Fusion uchta bajarish naqshini qo‘llab-quvvatlaydi:
- Single (Yagona): Bitta model vazifani qo‘shimcha ko‘rib chiqish chaqiruvlarisiz bajaradi.
- Cascade (Kaskad): Samarali model yechim loyihasini tayyorlaydi. Shu bilan birga, hakam uni qabul qilish yoki kuchliroq modelga eskalatsiya qilishni hal qiladi.
- Critique (Tanqid): Bir nechta model mustaqil ravishda loyihani parallel ko‘rib chiqadi. Tizim ularning fikr-mulohazalarini jamlaydi va asl modelga javobini qayta ko‘rib chiqish imkonini beradi.
Bajarish naqshlari bir suhbat ichida so‘rovlar orasida o‘zgarishi mumkin. Oddiy so‘rovlar keraksiz ko‘rib chiqish chaqiruvlaridan qochadi, kaskad va tanqid ish oqimlari esa qo‘shimcha model ishlovi tufayli ko‘proq vaqt oladi.
Routing tizim salomatligi, mavjudligi va keshlash chegaralarini ham hisobga oladi. Fusion asosiy suhbatni iloji boricha bir xil modelda saqlaydi, yordamchi modellar esa faqat o‘zlariga tayinlangan ish uchun zarur kontekstni oladi.
Ishlab chiquvchilar bajarish jarayonini kuzatishi va qaysi modellar hissa qo‘shganini tekshirishi mumkin. Biroq oraliq loyihalar qayta ko‘rib chiqilishi yoki bekor qilinishi mumkin va faqat yakuniy javob ko‘rsatiladi.
Fusion Standard (Standart) har bir vazifa uchun bitta model tanlab, xarajat, sifat va kechikishni muvozanatlashtiradi. Fusion High (Yuqori) murakkab nosozliklarni tuzatish va bir nechta faylni qamrab olgan o‘zgarishlar kabi keng qamrovli, aniq belgilangan o‘zgarishlarga mo‘ljallangan. Fusion Max (Maksimal) chuqurroq fikrlash va to‘liq tahlilni ustuvor qiladi.
Alohida Fusion Fast (Tez) rejimi Visual Studio Code ichki chatida kechikishga sezgir ssenariylar uchun xizmat qiladi va uni qo‘lda tanlab bo‘lmaydi.
Mahalliy model routingi (yo‘naltirish) ommaviy sinovda. Fusion Standard qo‘llab-quvvatlanadigan Windows tizimlarida MAI Code 1.1 Flash’dan foydalanishi mumkin, yoki Copilot CLI orqali Windows, macOS va Linux’da Qwen3.8 Flash Next’dan foydalanishi mumkin. Zarur bo‘lgan joylarda eksperimental sozlamalar yoqilishi kerak.
Fusion administrator siyosatlari, obuna cheklovlari, ma’lumotlar rezidensiyasi talablari, FedRAMP bilan bog‘liq nazoratlar yoki baholash-model siyosatlari tomonidan bloklangan modellarni istisno qiladi. Agar hech qanday mos model qolmasa, Fusion model tanlovchisidan yo‘qoladi.
Bitta operatsion cheklov alohida e’tiborga loyiq: AI loyihasini bekor qilish, o‘sha loyihada allaqachon qilingan fayl o‘zgartirishlarini avtomatik ravishda bekor qilmaydi. Ishlab chiquvchilar o‘zgarishlarni saqlashdan oldin ish maydonidagi o‘zgarishlarni ko‘rib chiqishi kerak. Fusion shuningdek subagentlardan alohida ishlaydi va ularni ishga tushirmaydi.
Bu farq avvalroq xabar qilingan Copilot prompt-injection zaifliklari, ish oqimi darajasidagi kodlash-agent jailbreaklari va Agentjacking hujumlari bilan birga muhim ahamiyatga ega. Bu xabarlar alohida xavfsizlik muammolari haqida bo‘lib, Fusion’ning o‘zida isbotlangan zaifliklar emas.
GitHub har bir ishlatilgan model uchun uning standart tarifida haq oladi, alohida Fusion to‘lovi yo‘q. Yuqori va Maksimal rejimlarda bir nechta model ishtirok etishi mumkinligi sababli ko‘proq AI kreditlari sarflanishi mumkin. Pullik obunachilar 2026-yil 30-noyabrgacha model xarajatlarida 30% chegirma oladi.
Fusion’ning mustaqil kontekst oynasi yo‘q: har bir qadam uni bajaradigan modelning cheklovlarini meros qilib oladi, ko‘rsatilgan sig‘im esa ishtirok etuvchi modellar orasidagi eng kichik cheklovlarni konservativ tarzda aks ettiradi.
SOC’ga tezkor to‘xtatish uchun ogohlantirish kontekstini bering. 16K SOC’dan tahdid razvedkasini integratsiya qiling.
GitHub has launched Copilot Fusion, a runtime model router and orchestrator that automatically selects AI models and execution patterns for coding tasks.
Instead of requiring developers to choose a model for each step, Fusion coordinates reasoning, code generation, debugging, and review, ultimately providing a single cohesive response.
The feature is available across all Copilot plans. It is generally available in Visual Studio Code, Copilot CLI, the GitHub Copilot app, and JetBrains IDEs. According to GitHub’s documentation, model selection remains subject to subscription entitlements and organizational policies.
Developers activate Fusion through the model picker. The system evaluates each request using capability signals covering reasoning, coding, debugging, and tool use, then chooses an execution strategy appropriate to the selected effort level.
Fusion supports three execution patterns:
- Single: One model completes the task without additional review calls.
- Cascade: An efficient model drafts a solution. At the same time, a judge determines whether to accept it or escalate to a stronger model.
- Critique: Multiple models independently review a draft in parallel. The system aggregates their feedback, allowing the original model to revise its response.
Execution patterns can change between prompts within the same conversation. Straightforward requests avoid unnecessary review calls, while cascade and critique workflows take longer because they involve additional model processing.
Routing also considers system health, availability, and caching boundaries. Fusion keeps the main conversation on the same model when possible, while assisting models receive only the context they need for their assigned work.
Developers can follow execution progress and inspect which models contributed. However, intermediate drafts may be revised or discarded, and only the final response is displayed.
Fusion Standard balances cost, quality, and latency by selecting one model per task. Fusion High targets substantial, well-scoped changes, including complex debugging and modifications spanning several files. Fusion Max prioritizes deeper reasoning and thorough analysis.
A separate Fusion Fast mode serves latency-sensitive scenarios in Visual Studio Code inline chat and cannot be selected manually.
Local model routing is in public preview. Fusion Standard can use MAI Code 1.1 Flash on supported Windows systems, or Qwen3.8 Flash Next through Copilot CLI across Windows, macOS, and Linux. Experimental settings must be enabled where required.
Fusion excludes models blocked by administrator policies, subscription restrictions, data-residency requirements, FedRAMP-related controls, or evaluation-model policies. If no eligible models remain, Fusion disappears from the model picker.
One operational limitation deserves particular attention: discarding an AI draft does not automatically reverse file edits that draft already made. Developers must review workspace changes before committing them. Fusion also operates separately from subagents and does not launch them.
That distinction matters alongside previously reported Copilot prompt-injection vulnerabilities, workflow-level coding-agent jailbreaks, and Agentjacking attacks. These reports concern separate security issues, not demonstrated vulnerabilities in Fusion itself.
GitHub charges for each model used at its standard rate, without a separate Fusion fee. High and Max can consume more AI credits because multiple models may participate. Paid subscribers receive a 30% model-cost discount through November 30, 2026.
Fusion has no independent context window: each step inherits the limits of the model executing it, while the displayed capacity conservatively reflects the smallest limits among participating models.
Give your SOC the alert context to stops incidents quickly.Integrate Threat Intel from 16K SOCs.
