- Book: System Design Pocket Guide: Interviews
- The series: Pocket Guides for Developers — 8 short books on agents, RAG, prompts, observability, events, databases and system design — the whole series
- My project: Hermes IDE | GitHub — an IDE for developers who ship with Claude Code and other AI coding tools
- Me: xgabriel.com | GitHub
You pick one entry in the Copilot model picker. Behind it, one model writes a draft and other models judge or critique it before you see a single answer.
That entry is HydraFusion, and on October 14 GitHub gave it a new model pool: Claude Opus 5.5, GPT-6 Astra, GPT-6.1 Sol and GPT-6 Luna. It also added a second tier, HydraFusion Max. The billing rule sits in GitHub's docs: you're billed for each model it uses, at that model's standard rate.
What changed on October 14
HydraFusion is not a model. GitHub's docs call it a runtime model orchestrator. For each prompt it chooses an execution pattern and a reasoning effort, picks the models, and returns one response. Copilot's Auto mode picks one model per request. HydraFusion picks a workflow and coordinates several models within a single turn.
It has three patterns, per the docs. Single sends the task to one model. Cascade has an efficient model draft and a judge accept the draft or escalate. Critique has one model draft, several models critique it in parallel, and the drafter revise once.
The October 14 changelog changes four things:
- The pool: Claude Opus 5.5, GPT-6 Astra, GPT-6.1 Sol and GPT-6 Luna.
- Effort per stage: medium for first drafts, high for review. Before, each model ran at its default.
- Cascade: it can now add a judging step that inspects the first attempt and asks for a repair only when it's needed.
- Critique: the critics lean more on tools that inspect your actual workspace, and they run in parallel.
HydraFusion Max uses the same architecture and workflows with higher reasoning effort, aimed at more demanding tasks. The docs add that Max also offers extra high reasoning.
Both are a research preview in VS Code, the GitHub Copilot app and Copilot CLI, on Copilot Pro, Pro+, Business and Enterprise with experimental or preview features enabled. GitHub's docs say plainly that "HydraFusion isn't intended for production workloads."
One prompt, three possible paths: Single, Cascade or Critique. You only see the final answer.
What one HydraFusion turn costs you
There is no separate charge for HydraFusion, and the discount for Auto model selection doesn't apply. Each model it calls bills at its own standard rate. GitHub's docs also note that, because several models can run per task, one task can use more AI credits than a single model would.
So the same prompt can cost very different amounts. A Single run costs roughly what one model costs; a Critique run pays for the drafter, every critic and the revision. The main conversation stays on one model when possible to keep cached tokens, which limits the overhead without removing it.
Two more lines in the docs matter for budgeting. You can't choose which models HydraFusion uses, and it only draws on models your plan includes and your organization's policy allows. If none qualify, it doesn't show up in the picker at all.
The context window is the other trap. The docs say HydraFusion doesn't have a context window of its own; the one it shows is based on the smallest limit among its models. If your conversation is already past that, compact it before you switch.
GitHub's numbers, and what they were measured against
When GitHub announced HydraFusion on September 4, it published offline results against Claude Opus 5:
| Benchmark | Quality vs Opus 5 | Estimated cost vs Opus 5 |
|---|---|---|
| TerminalBench 2.1 | +4.9 points | 67% lower |
| DeepSWE | -1.5 points | 36% lower |
| CheckpointBench | -0.1 points | 65% lower |
Read those with their limits. They come from controlled offline runs tied to specific benchmark revisions, workflow configurations, a model pool and pricing assumptions. CheckpointBench is an internal GitHub benchmark, and the cost figures are estimates, not billing data. Only TerminalBench shows a quality gain. The other two trade a little quality for a lower bill.
They also describe the September setup. The October 14 pool and per-stage effort are different, and the changelog publishes no new numbers. Omdia analyst Torsten Volk told TechTarget that "the economics of using LLMs for every coding task often do not work."
No HydraFusion fee, but every model it calls runs its own meter.
The edit a discarded draft leaves behind
One limitation in the docs deserves more attention than the benchmarks. When a draft is thrown away, changes that the draft already made in your workspace, such as file edits, aren't undone automatically. The final answer can be the escalated one while the rejected draft's edits are still on disk.
So give it a branch it can make a mess on. In Copilot CLI the docs enable it with --experimental and then /model. Do that inside a git worktree:
git worktree add ../app-hydra -b try/hydrafusion
cd ../app-hydra
copilot update
copilot --experimental
# in the session: /model -> HydraFusion (Research Preview)
For a one-shot prompt, the docs give copilot --experimental --model hydrafusion -p "YOUR-PROMPT". In VS Code 1.140 or later, turn on chat.copilot.hydraFusion.enabled and pick it from the Copilot Chat model dropdown.
Before you commit anything from that branch, run git status and git diff and read every file, including the ones the final answer never mentions. The docs ask for the same: review changes before committing.
Running it next to your other agents
I build Hermes IDE, so weigh this section with that in mind. GitHub Copilot CLI is one of the agents Hermes opens in a real terminal, next to Claude Code, Codex and the others, along with any command-line agent you add. ⌘N starts each task on its own hermes/<name> branch in its own worktree, which is the isolation the discarded-draft caveat calls for.
To get HydraFusion there, start a GitHub Copilot CLI task, enter /experimental on and restart Copilot CLI as GitHub's docs describe, then pick HydraFusion from /model. Give the same ticket to Claude Code in a second worktree, and you can compare the router's diff against a single model's. I haven't run HydraFusion myself.
Hermes 2.3.4's agent picker: Claude Code and Codex installed, GitHub Copilot CLI and the other supported agents listed, plus Custom agent.
Hermes is free, source-available under BSL 1.1, and runs on macOS, Windows and Linux. It isn't affiliated with GitHub, Anthropic, OpenAI or Nous Research.
Should you switch it on?
If you already pay for Copilot Pro or higher and have preview features on, trying it costs a setting and a branch, plus whatever credits each run burns. Take a task you've already finished by hand, run it on HydraFusion and on a single model you trust, and compare the diffs and the credits. That tells you more than GitHub's table can.
Would you let a router decide which models touch your code, if it bills you for every one it calls? Tell me in the comments.
If this was useful
HydraFusion is a system-design trade-off in miniature: when to escalate, when to pay for a second opinion, and what a cheap first attempt costs you when it fails. The System Design Pocket Guide: Interviews walks through 15 real designs where those same calls get made step by step. It's one of 8 short books in Pocket Guides for Developers, covering agents, RAG, prompts, observability, events, databases and system design. If you want to race Copilot CLI against Claude Code on one ticket, Hermes IDE gives each its own worktree.
Top comments (0)