Nadella promises 77% lower cost – independent measurement of Copilot Fusion shows 51%
At GitHub Universe on Wednesday, October 28, 2026, GitHub chose to bet on routing between models rather than building its own frontier model. Copilot Fusion selects in real time which model – or blend of models – will handle each step of a coding task. Independent measurements show roughly halved costs with modestly lower quality, but the company's own savings claims are far more optimistic than independent testing supports.
The news: a router, not a model
Copilot Fusion was launched at GitHub's developer conference GitHub Universe on Wednesday, October 28. The system decides which AI model, or which blend of models, will handle each step of a coding task – weighing what the task requires across reasoning, code generation, debugging and tool use, according to TNW.
GitHub's Chief Product Officer Mario Rodriguez was explicit about what the product is not, in an interview with VentureBeat: "It's not a model. It's an orchestration engine at runtime." He explained that Fusion can both route individual requests and orchestrate multiple model calls to complete a task.
How it works
Fusion has three reasoning effort levels – standard, high and max – and three execution patterns: single, cascade and critique. The model is thus not chosen once for all, but per request.
Artificial Analysis has identified the models in the pool: Claude Opus 5.5 and Claude Sonnet 5.5 from Anthropic, plus GPT-6 Astra and GPT-6 Sol from OpenAI. Which models are actually used can vary from request to request and depends on the policies an organization sets, according to VentureBeat.
The numbers: two sources, two stories
The interesting part lies in the gap between the company's own claims and the independent measurement.
Artificial Analysis published its testing on October 28: Fusion Max cost an average of $5.59 per task, versus $11.42 for Claude Code with Claude Opus 5.5 on the Max setting – roughly 51 percent lower. On quality, Fusion Max scored 60 points versus about 62 on the firm's Coding Agent Index v2.0 – two points behind. The test also showed 7.2 million tokens per task.
The Devin Fusion comparison is instructive: the competing orchestration system, combining Opus 5.5 Max and GPT-6 Luna High, scored 62 points on the same index with an average task price of $8.00.
GitHub's own claims run higher. Jay Parikh wrote in a post on GitHub's blog that Fusion Max scored within two points of Claude Opus 5.5 Max at half the price and with 40 percent less time per task, according to TNW. Microsoft CEO Satya Nadella went further: in a LinkedIn post, he claimed Fusion delivers frontier-level quality 40 percent faster and at up to 77 percent lower cost. These are company statements – the independently measured figure is around 51 percent, and the quality was measurably lower, not equal.
The context: an entire platform gets updated
Fusion was part of a larger launch wave. From the same day, Enterprise customers can use all Copilot features without a separate subscription; they pay by consumption via AI credits across the Copilot app, code review, automations and other features. Access currently goes through GitHub's sales team, with self-serve "coming soon," according to TNW and LavX News.
In addition came Copilot projects, which run agent work across multiple repositories, automations with a learning engine, and code review Max in three tiers.
Beneath the product layer, GitHub is investing heavily in infrastructure: 4 million CPU cores, 160 petabytes of fast storage and 50 terabits per second of network capacity added this year. A migration to Azure with multi-region rollout is underway and is to be completed early next year.
GitHub's own Copilot harness, built by the GitHub team, is, according to Microsoft's blog, already in use across the entire company and is now being made available through Microsoft Foundry – a fully managed way for developers and enterprise customers to build agents that reason, use tools and complete complex, long-running tasks.
The usage is already there: agents start 40 percent of pull requests
GitHub's annual Octoverse report, also released on October 28, shows that over 40 percent of all pull requests in September began with a draft from an AI agent – up from 17 percent in March. September alone saw 110 million pull requests started with an agent draft. Activity rose sharply around June, when agents began working across multiple days and running in the cloud, according to the report.
GitHub stresses that the agent figure is a floor, not a ceiling: the company cannot detect all agents, so the actual number is likely higher. COO Kyle Daigle stated in an interview with The Indian Express that GitHub now has 225 million users, of whom 50 million use Copilot daily, and that developers pushed 13.5 billion contributions in 2026 – more than in 2024 and 2025 combined. (That figure covers contributions, not commits, and should not be confused with commit counts from other sources, which use different measurement methodologies.)
What remains to be clarified
Three things remain unresolved. First, the methodology behind Nadella's 77 percent figure, which has not been independently documented and sits considerably above Artificial Analysis' measurement of around 51 percent. Second, the VentureBeat article about the independent test is truncated, so the analysis of token consumption and cost is not fully available. Third, how price and quality actually develop in practice after launch – Fusion routes dynamically, so the launch-day figures are a snapshot, not a constant.
What is clear, however, is the direction: GitHub is betting that orchestration – not model choice – will be what determines the price and quality of coding agents. And according to the company's own data from the same day, the agents have already stopped merely assisting. They are writing the code.