- Book: Observability for LLM Applications
- Also by me: Thinking in Go (2-book series) — Complete Guide to Go Programming + Hexagonal Architecture in Go
- My project: Hermes IDE | GitHub — an IDE for developers who ship with Claude Code and other AI coding tools
- Me: xgabriel.com | GitHub
Two weeks ago, GitHub's multi-model router was a research preview called HydraFusion, with a new Max option and a warning label. On October 28 it came out the other side with a new name and a bigger claim. GitHub's changelog says Copilot Fusion Max scored 60 on the Artificial Analysis Coding Agent Index, "within two points of Claude Code with Opus 5.5 Max," at about half the average API cost per task.
That is GitHub's number about GitHub's product. It may well hold up. It is still the vendor's claim, and you can check it on your own tickets in a week.
What shipped on October 28
Copilot Fusion is not a model. In the model picker you choose Copilot Fusion plus an effort level, and Copilot routes and orchestrates models for each prompt, per the changelog. GitHub describes it as HydraFusion's multi-model orchestration combined with the fine-tuned model selection behind Copilot's auto mode.
The facts from the announcement:
- Where: all Copilot plans, in VS Code, the GitHub Copilot app, Copilot CLI and JetBrains IDEs. More clients are listed as coming soon, with no dates.
- Effort levels: Standard (versatile models for everyday coding), High (more capable models, deeper reasoning) and Max (the most capable models for demanding tasks).
- Patterns: Single (one model does the task), Cascade (an efficient model drafts, a quality check accepts it or escalates to a stronger model) and Critique (one model drafts, another reviews, the first revises).
- Migration: preview users who had auto, HydraFusion or HydraFusion Max enabled move automatically to the matching Fusion effort level.
GitHub launched it during its Universe conference week in San Francisco, The AI Economy reports.
Fusion's job in one picture: one prompt comes in, and the router decides which model, or chain of models, handles it.
The benchmark, and what's behind it
AlphaSignal's write-up lays out the numbers on the Artificial Analysis Coding Agent Index v2.0:
| Fusion Max | Claude Code, Opus 5.5 Max | Devin Fusion | |
|---|---|---|---|
| Composite index | 60 | 62 | 62 |
| FrontierCode | 51 | 56 | 54 |
| Cost per task | $5.59 | $11.42 | $8.00 |
The same piece puts Fusion Max at an average of 7.2M tokens per task and notes that these costs come from benchmark runs, not from Copilot subscription pricing. GitHub adds that Fusion Max used about 40% less time per task (changelog).
The index itself is an equal-weight composite of FrontierCode v1.1, Terminal-Bench 4.0 and SWE-Atlas-QnA-AA, per Artificial Analysis. At the time of writing, that public page did not list a Copilot Fusion entry, so the 60 above comes from GitHub and the outlets quoting it.
Read the table straight. Fusion Max is two points behind on the composite and five points behind on FrontierCode. Its case is cost: roughly half of Claude Code's per-task spend in that run. GitHub also cites a DeepSWE run in the GitHub Copilot harness, where Fusion High and Max scored up to 8.7 points higher than comparable frontier-model efforts and cut estimated cost by up to 77% (The AI Economy). That one is GitHub's harness and GitHub's measurement.
The bill: every model it picks charges you
Fusion has no fee of its own. You pay for each model it uses, at that model's standard rate (changelog). The preview docs said the same: "There's no separate charge for HydraFusion" (GitHub Docs). Since June 1, Copilot usage is metered in AI Credits, charged at each model's listed API rate (GitHub).
That makes the pattern matter. A Single run bills one model. A Critique run bills a drafter and a reviewer, then the drafter again. A Cascade that escalates bills the cheap draft and the expensive answer. I covered that per-model billing when HydraFusion added Max, and GA doesn't change it.
What GA adds is a discount: "For a limited time, users on paid Copilot plans will receive a 30% discount on model costs when using Copilot Fusion." The changelog gives no end date, so budget for the day it stops.
No router fee, but every model in the chain lands on the receipt. The 30% discount is for a limited time.
Your model policy is now Fusion's menu
For teams, the admin side is the practical change. Fusion only picks from models enabled in your enterprise and organization model policies. Enterprise admins can see which models Fusion chooses from, set each model's policy state and disable a model for Fusion without disabling it for everyone. They can also make Fusion, at any effort level, the default model through enterprise managed settings (changelog).
The changelog doesn't name the models in the pool. AlphaSignal lists Claude Opus 5.5, Claude Sonnet 5.5, GPT-6 Astra and GPT-6 Sol. If your company has only approved some of those, Fusion's choices shrink to match, and the benchmark run above may not describe what your developers get.
There's also a local angle. The AI Economy reports that when you're offline, Fusion falls back to the Single pattern with a model on your machine, and quotes GitHub's chief product officer Mario Rodriguez calling the cloud-plus-local approach "the best of both worlds" (source). Copilot CLI already finds local models on an Ollama instance through /model since version 1.0.94-0, and GitHub notes that choosing a local model does not turn on offline mode or disable telemetry (changelog).
Check the claim on your own repo
A composite index is an average over somebody else's tasks. Before you set Fusion as the team default, give it a week of your tickets:
- Pick 5 to 10 real, closed tickets with tests that prove the fix.
- Run each in Copilot CLI with Fusion High or Max, and in the agent you use now, on separate branches.
- Record what you care about: tests passing, diff size, review comments, wall time, and the AI Credits each run used.
- Repeat a couple of runs. A router can take a different path on the same prompt, so one run is an anecdote.
I build Hermes IDE, so weigh this part with that in mind. Hermes runs GitHub Copilot CLI and Claude Code each in its own terminal (README), and ⌘N starts every task on its own hermes/<name> branch in its own worktree. That is the setup GitHub's comparison implies: the same task, Fusion in Copilot CLI against Claude Code, without the two agents touching each other's files. Each agent signs in with its own CLI login, and Hermes never asks for tokens. It's source-available and runs on macOS, Windows and Linux. I haven't run Fusion myself.
The ⌘N New task sheet in Hermes 2.3.4. GitHub Copilot CLI sits in the same agent list as Claude Code and Codex.
What to watch
GitHub's Mario Rodriguez says more orchestration patterns are planned (The AI Economy), and the changelog says more clients are coming. Three questions are still open for working developers: when the 30% discount ends, whether Artificial Analysis lists Fusion on its public leaderboard, and whether your team can see which model handled which prompt when a change goes wrong.
Would you let a router pick your coding model, or do you want the model name on every diff? Tell me in the comments.
If this was useful
A router like Fusion turns "which model wrote this?" into a question you have to answer from logs. Observability for LLM Applications is about exactly that: tracing which model handled each step, running evals before you trust a cascade's quality check, and keeping cost per task visible when the path changes under you.
If you want to race Fusion against your current agent on the same ticket, Hermes IDE keeps each run on its own branch and worktree (GitHub).
Top comments (0)