Skip to content
NewsProduct launch

GitHub Copilot Fusion reaches general availability, scoring within two points of Claude Opus 5.5 Max at half the cost

· by Pondero Newsdesk

The short version

GitHub's Copilot Fusion exited research preview and reached general availability on October 28, 2026, merging HydraFusion's multi-model orchestration with Copilot's auto model selection, and scored 60 on the Artificial Analysis Coding Agent Index, within two points of Claude Code running Opus 5.5 Max, at roughly half the API cost and token count per task, per GitHub's changelog.

GitHub Copilot Fusion reaches general availability, scoring within two points of Claude Opus 5.5 Max at half the cost

GitHub's Copilot Fusion left research preview and reached general availability on October 28, 2026, and for the first time the feature's performance claim came from an outside benchmark instead of GitHub's own test suite. Fusion's top effort tier scored 60 on the Artificial Analysis Coding Agent Index, within two points of Claude Code running Opus 5.5 Max, while using roughly half the average API cost per task, about half as many tokens, and close to 40% less time per task, per GitHub's changelog.

What

Fusion combines Project HydraFusion's multi-model orchestration with the fine-tuned picking logic from Copilot's existing auto model selection feature, folding two previously separate preview options into one default choice, GitHub said in its changelog. Selecting Copilot Fusion in the model picker does not select a specific model; it hands the task to a system that decides, for each prompt, which models to use and how to combine them, while the developer only sets an effort level: Fusion Standard for everyday coding, Fusion High for harder reasoning tasks, or Fusion Max for the most demanding work. Under those effort levels, Fusion routes each task through one of three execution patterns: Single, where one model answers directly; Cascade, where an efficient model drafts first and a check decides whether to accept the draft or escalate to a stronger model; and Critique, where a second model reviews a first model's draft and the original model revises it.

Enterprise and organization administrators keep the model policy controls that already govern Copilot. They can see which models Fusion draws from, disable a specific model for Fusion without disabling it for everyone, and set Fusion, at any effort level, as the default for users under Enterprise managed settings, per GitHub's changelog. Usage is billed per model at that model's standard rate, the same as picking a model manually, and GitHub is running a temporary 30% discount on model costs for paid Copilot plans that use Fusion. The feature is live now in VS Code, the GitHub Copilot app, Copilot CLI, and JetBrains IDEs, with support for more clients described as coming later. Anyone who had opted into the prior "auto," "HydraFusion," or "HydraFusion Max" research-preview settings is carried over automatically to the matching Fusion effort level.

The general-availability announcement landed on the opening day of GitHub Universe 2026, the company's two-day developer conference running October 28 and 29 at Fort Mason Center in San Francisco, per GitHub's event guide.

A temporary discount pressures admins to decide before the benchmark gap moves again

For a Copilot admin, the calculation around model selection changes because the evidence changed, not just the feature's status. A preview backed only by GitHub's own test numbers is a different decision than a generally available default that lands within two benchmark points of a named frontier competitor while cutting API cost and token count roughly in half. That combination weakens the case for leaving manual model selection on for routine coding work, since Fusion now inherits whatever model policy and spending guardrails already apply to an org instead of requiring a separate approval path. The 30% discount, which GitHub frames as temporary without naming an end date, adds a second, more time-sensitive pressure: an admin who waits a full billing cycle to compare Fusion against manual selection risks losing the discount before the comparison finishes. What the benchmark does not resolve is the choice between effort levels. GitHub published cost, token, and time figures only for the Max tier against one named competitor, leaving Standard and High without their own published comparison.

Context and reactions

GitHub's benchmark claims for this feature were self-reported until now. When Project HydraFusion shipped as a Copilot CLI-only research preview on September 4, 2026, the company cited its own offline evaluation across three agentic coding benchmarks measured against Claude Opus 5: 67% lower estimated cost with task quality 4.9 percentage points higher on TerminalBench 2.1, and 65% lower cost on CheckpointBench, per GitHub's original HydraFusion post. GitHub never refreshed those specific figures even as it expanded HydraFusion to VS Code and the Copilot app on October 1, moved the model pool to Claude Opus 5.5, GPT-6 Astra, GPT-6.1 Sol, and GPT-6 Luna, and added a Max tier on October 14. Today's Artificial Analysis Coding Agent Index score is the first third-party number attached to this feature, and the comparison target moved too, from Opus 5 to Opus 5.5 Max, tracking Anthropic's own model lineup over the same stretch.

GitHub also named a production customer for the first time. Hari Pokhrel, a principal software engineer at AT&T, said Fusion Standard had been "automatically selecting the right model for the task, reducing unnecessary spend and preserving capacity for higher-value work" across the company, per GitHub's changelog. That is a named customer quote GitHub chose to publish, not an independently measured efficiency result, and GitHub disclosed neither AT&T's deployment size nor a specific percentage of spend avoided.

What to watch next

Watch whether GitHub publishes separate cost and latency figures for the Standard and High effort levels, since the GA benchmark covers only Max against one competitor. Also watch the 30% discount's expiration and whether Artificial Analysis updates the Coding Agent Index as Anthropic, OpenAI, and other labs ship new flagship models that could move the two-point gap in either direction.

Sources