GitHub Copilot Fusion Nears Claude Code at Half the Task Cost

GitHub stops asking developers to choose a model

At its Universe conference in San Francisco on October 28, GitHub made Copilot Fusion generally available. The system decides which AI models handle a coding task and how those models work together.2224 GitHub's own documentation calls Fusion a runtime router and orchestrator, not a model. A developer picks it in the model picker, and Fusion then chooses an execution pattern, a reasoning-effort level and the models that carry out each request.21 Mario Rodriguez, GitHub's chief product officer, put it this way to VentureBeat: "It's not a model. It's an orchestration engine at runtime."24

The launch's headline claim is that Fusion beats Anthropic's Claude Code at half the cost. Read the coverage closely and the accurate version is narrower, though it is still impressive. Fusion comes close to Claude Code's benchmark score for about half the money per task. It does not beat it.

How Fusion works

Fusion merges two earlier pieces of Copilot. One is the multi-model orchestration from Project HydraFusion, which GitHub released as a research preview in September. The other is Copilot's existing automatic model selection.227 Developers no longer pick a model. They pick an effort level: Standard for everyday work, High for well-scoped but substantial tasks such as multi-file changes, and Max for the hardest reasoning. A fourth level, Fusion Fast, is used automatically in VS Code inline chat and can't be selected.21

For each prompt, Fusion uses one of three patterns. In Single, one model does the whole job. In Cascade, a cheaper model writes a draft and a judge either accepts it or passes the work up to a stronger model. In Critique, one model drafts while several others review it in parallel, and the drafting model revises using their combined feedback.21 The pattern can change from one prompt to the next within the same session, and GitHub says choosing the pattern adds very little latency.21

GitHub also describes cost-saving details in how Fusion handles context. It tries to keep the main conversation on the same model so cached tokens can be reused, and it gives helper models only the context they need.26 According to Artificial Analysis, the model pool includes Anthropic's Claude Opus 5.5 and Sonnet 5.5 and OpenAI's GPT-6 Astra and GPT-6 Sol. GitHub has said the pool will change over time.2326

Fusion is available on every Copilot plan in VS Code, the Copilot app, Copilot CLI and JetBrains IDEs. Routing to local models is in public preview. On supported Windows machines, Fusion Standard can use MAI Code 1.1 Flash. Through the CLI, it can use Qwen3.8 Flash Next on Windows, macOS and Linux.

What the benchmark shows

The main evidence comes from Artificial Analysis, which tested Fusion's Max setting in Copilot CLI on version 2.0 of its Coding Agent Index.10 Fusion scored 60, two points behind Claude Code running Opus 5.5 at max effort. Its average cost per task was $5.59, against $11.42 for Claude Code.2324 It also used about 7.2 million tokens per task, roughly half as many as the Claude Code setup.23

So the cost result holds up. The "beats" framing does not. The New Stack's headline says Fusion beats Claude Code, but its own article says Fusion landed within two points of it.22 Daily.dev's summary went further and reported the scores the wrong way round, as 60 for Fusion and 58 for Claude Code.25 Artificial Analysis says plainly that Fusion scored 2 points below.10

The index's sub-scores are more mixed than either headline suggests. Fusion leads on Terminal-Bench 4.0 with 71. It ranks third on FrontierCode with 51, behind Claude Code at 56 and Cognition's Devin Fusion at 54. Its 57 on SWE-Atlas-QnA-AA trails the leaders.10 Devin Fusion, which pairs Opus 5.5 with GPT-6 Luna, scored 62 at $8.00 per task.24 That matters: GitHub isn't alone in orchestrating several models. A rival already matches Claude Code's score at a lower price, and Fusion's advantage is that it costs less still.

Other figures circulating need care. Microsoft CEO Satya Nadella described Fusion as delivering frontier-level quality 40% faster at up to 77% lower cost.7 VentureBeat traced the 77% figure to GitHub's own tests on the DeepSWE benchmark with Copilot's agent harness. There, Fusion High and Max beat comparable frontier-model setups by up to 8.7 points on task completion.24 That is a company test, and it should not be treated as equivalent to the third-party index. The September HydraFusion preview had claimed up to 67% lower cost on a terminal benchmark.27

The bill is more complicated than "half the cost"

Several outlets point out that "half the API cost" compares API rates per task. It is not a comparison of subscription bills.93 Fusion has no fee of its own. Each model it calls is charged at that model's standard rate from the user's GitHub AI Credits, and paid plans get a 30% discount on Fusion model costs through November 30.3 Xenospectrum notes that this discount is separate from the roughly 51% per-task gap in the benchmark.26

Access also varies by plan. According to one breakdown of GitHub's pricing, Copilot Free and the $10 Pro tier get only Fusion Standard. The High and Max levels, including the setup that was benchmarked, require Pro+ at $39 or Max at $100.3 By comparison, Claude Code is included in Claude's $20 Pro plan, and Claude Max starts at $100.3

The launch also changes how enterprises pay. GitHub Enterprise customers can now use Copilot features without a separate Copilot subscription and pay for the AI credits they use.7 Code completion keeps working when a budget runs out, but Chat, code review and cloud agents need budget remaining.4 Fusion fits that model well. If customers are billed per token, a router that cuts tokens per task becomes a strong selling point.

AI code review moves inside the agent

Fusion's Critique pattern is automated code review happening inside the agent's own loop. Before a developer sees the result, other models have already checked the draft.21 Rodriguez described Critique as a way to apply stronger models to the review step without paying for the most expensive model at every stage.24

Anthropic has moved in a similar direction from another angle. Claude Code's pull-request Code Review feature became generally available on October 15, with review levels that can be set for each repository.17 Read together, the two launches suggest review is shifting from a step after the code is written to something built into how the code is produced.

GitHub's documentation includes a caveat developers should not miss. If Fusion throws away a draft, any file edits that draft already made are not rolled back automatically, so changes need checking before they are committed.21 Fusion also does not start subagents. In other words, the review layer improves answers but doesn't make the workspace any safer.

What Anthropic has been shipping

Claude Code has kept up a fast release pace. It moved from v2.1.227 to v2.1.305 between August and late October, and v2.1.312 shipped on October 27.1714 The October releases focused on operations: forked subagents on by default in 2.1.302, a setting in 2.1.296 that pins every subagent in a workflow to one model, and a plugin system called Mods.131618

The two companies are taking different approaches. Anthropic is giving developers more precise control over which model runs where. GitHub is taking that decision away from developers and making it automatically. Both are reasonable for different buyers. Teams that have validated one model for a regulated codebase may not want routing they can't predict. GitHub's answer is to let admins remove individual models from Fusion's pool without disabling them across Copilot.2722

Our reading

GitHub's numbers suggest developers were already choosing models by hand at scale. Paid Copilot users used 47 different models in September, and no single model handled more than 25% of requests.4 Replit and Ramp have also made automatic routing their default.7 Fusion makes that practice official and adds a pricing argument to it.

GitHub's position here is unusual. Anthropic supplies many of Copilot's models, and GitHub is now marketing a product that matches Anthropic's own agent at half its cost.4 The supported claim is that Fusion comes close to Claude Code's score for about half the price. The claim that it beats Claude Code is not supported. As VentureBeat notes, the real test is whether teams see similar savings once review, governance and deployment costs are counted.24 Teams can measure that with the discount in effect until November 30.