GitHub Copilot Fusion Launches With 50% Cost Savings as Infrastructure Expands to Handle AI Agent Surge

GitHub Copilot Fusion Delivers Frontier Performance at Half the Cost

GitHub launched Copilot Fusion at GitHub Universe on October 28, introducing an AI orchestration engine that intelligently routes coding tasks across multiple AI models to balance performance and cost. According to independent testing by Artificial Analysis, Copilot Fusion Max scored within two points of Anthropic Claude Opus 5.5 Max on the Coding Agent Index while completing tasks at half the cost—$5.59 per task compared to $11.42—and 40% faster3

3

. Microsoft CEO Satya Nadella highlighted potential savings of up to 77% in company evaluations, positioning task-level costs rather than individual model pricing at the center of GitHub's value proposition for enterprise teams3

3

.

The AI model routing service evaluates tasks across reasoning, code generation, debugging, and tool use, then decides whether to dispatch work to a single model or coordinate multiple models. GitHub Chief Product Officer Mario Rodriguez explained that Copilot Fusion operates through three execution patterns: Single mode routes directly to the best model, Cascade starts with a less costly model and escalates if needed, and Critique uses one model to draft while another reviews before the original model revises its response3

3

. Developers select from three effort tiers—Standard for everyday coding, High for difficult work, and Max for complex tasks—without manually managing model selection3

3

.

Infrastructure Upgrades Address AI Agent Traffic Surge

GitHub recorded 7.4 billion code commits in September 2026, representing a fivefold increase over the previous year as AI agents generated unprecedented volumes of programming code1

1

. This surge in AI-driven code commits caught the company unprepared, leading to outages and availability issues throughout the year, with the platform's 90-day availability dropping to approximately 84% on May 71

1

. The Microsoft-owned code hosting service is now adding 4 million CPU cores and 160 petabytes of high-speed storage to its infrastructure, along with 50 Tb/s of network capacity4

4

.

Jay Parikh, EVP of CoreAI at Microsoft, emphasized that capacity alone won't solve the challenge. GitHub is redesigning its technical architecture to separate compute and storage, expecting this change to enable 35x more throughput1

1

. The company is also migrating to a multi-region deployment in Azure, with completion expected early next year to improve resilience2

2

. These infrastructure upgrades respond to what Parikh described as a fundamental shift: "I think all the software we've written to date will amount to less than 1% of the software we'll write in the next five years"2

2

.

AI Agents Now Drive 40% of Pull Requests

The annual Octoverse report revealed that developers pushed code 17.9 billion times in the 12 months ending September 30, exceeding the combined total from 2024 and 20252

2

. By September, more than 40% of pull requests started with a draft from an AI agent, up from 17% in March—representing 110 million agent-initiated pull requests in September alone2

2

. Activity rose sharply around June when AI agents began working across multiple days and running in the cloud2

2

. GitHub's platform now hosts more than 225 million developers, with the United States leading at 38.3 million and India at 32 million2

2

.

GitHub Enterprise Gets Full Copilot Access Without Additional Seats

GitHub announced that all GitHub Enterprise users can now access every Copilot feature without purchasing separate per-seat subscriptions, instead paying only for the AI credits they consume1

1

. This change aims to boost workplace adoption of AI-powered tools while giving organizations more flexibility in managing inference costs. Existing Copilot Business and Copilot Enterprise customers can maintain their current plans or switch when they choose2

2

.

The company also introduced the GitHub Copilot app, which connects AI agents and tools across projects. Copilot Projects enables developers to run agent work across multiple repositories in parallel managed sessions, while session continuity allows one cloud agent session to move between the app, VS Code, the command line, and the web2

2

. Copilot Automations can now learn from their results and propose changes for developer approval, while Copilot Connectors simplify connections to Model Context Protocol servers, integrating real-time production telemetry from services like Cloudflare, Vercel, and Sentry4

4

.

New Security Features Target Vulnerability Detection and Supply Chain Protection

GitHub rolled out AI-powered code scanning that checks every pull request and repository for vulnerabilities, extending beyond languages covered by traditional rule-based scanning to include already-merged code1

1

. The company introduced reachability analysis tools that help development teams prioritize alerts by showing whether their code calls a vulnerable function, complete with an inspectable call path4

4

. A new package firewall inspects npm packages at download time, checking enterprise policy settings before fetching packages and blocking known malicious packages, typosquatted names, and zero-day dependency attacks1

1

.

GitHub also launched Copilot Code Review Max, which the company says achieved the highest F1 score on ReviewBench for high-severity issues2

2

. These security enhancements address growing concerns about software development quality as teams accelerate their use of AI agents. Parikh noted that "moving faster only helps if we can maintain the quality of what ships," emphasizing the need for safeguards that catch meaningful, actionable bugs4

4

.