[Research Preview] HydraFusion is live in GitHub Copilot CLI: Frontier quality via multi-model orchestration

[ebndev]

We are excited to introduce Project HydraFusion, a research preview that delivers frontier intelligence through runtime orchestration. It creates a full execution plan, choosing from models across multiple providers to draft, critique and revise, or cascade to more powerful models to complete your task.

How to try it

HydraFusion is available to users on all GitHub Copilot plans through /experimental in GitHub Copilot CLI. Usage is based on the tokens consumed by the models HydraFusion uses, priced at each model’s standard rate.
To try HydraFusion in Copilot CLI:

1. Run /update to install the latest version
2. Run /experimental on
3. Run /model, then select HydraFusion (Research Preview)

How HydraFusion works

HydraFusion treats workflow selection as an optimization problem. It uses capability signals for reasoning, code generation, debugging, and tool use to select the most efficient execution pattern to meet the quality bar.
For each request, HydraFusion currently chooses one of three execution patterns:

• Single. One selected model solves the task directly.
• Cascade. An efficient model drafts a solution and a quality gate decides whether to accept it or escalate to a stronger model.
• Critique. One model drafts a result, an independent read-only critic from a different model family reviews it (following the same review pattern as Rubber Duck), and the drafting model revises once.

For further information, please check out our blog post on Project HydraFusion: https://gh.io/hydrafusion

Why we built HydraFusion

HydraFusion fills a key role in our overall strategy to deliver automated semantic routing between local, cloud, and compound models. For developers, that complexity stays behind the scenes: you select HydraFusion like any other model, and it chooses a workflow that balances performance, cost, and latency for each task.
In offline evaluations across three agentic coding benchmarks, HydraFusion consistently demonstrated frontier-level quality with substantial estimated cost savings. On TerminalBench 2.1, it improved verified task quality by 4.9 percentage points at 67% lower estimated cost compared with Claude Opus 5.

What we need from you

Try our research preview today. This preview is designed to learn which tasks benefit from compound workflows and how orchestration affects latency and cost in practice. When sharing feedback, please include copilot debug logs or your session ID through /session id

📍Frequently asked questions

Q: Where can I use it?

Today, only in GitHub Copilot CLI. It shipped in the latest release as an experimental research preview. The GitHub Copilot app and VS Code are both targeting September as a fast follow.

Q: What is the major difference between Auto and HydraFusion?

Both Auto and HydraFusion do the choosing for you. Auto selects one model per request. HydraFusion runs multiple models in a single turn, has them review and critique one another, and fuses the result. Auto picks the optimal model; HydraFusion picks the optimal workflow, possibly using multiple models.

• Progression: Auto V1 (Jan 2026, capacity/SKU-aware per-request selection) → Auto V2 aka. HyDRA (May 2026, intent-scored routing) → HydraFusion (Aug 2026, per-turn orchestration and cache-aware workflows).

Auto and HydraFusion are both intelligent model routing, and we expect them to come together into a single experience over time. We’re still evaluating what that looks like.

Q: Can I select which specific models HydraFusion draws from or exclude certain models?

Not today. HydraFusion picks from a curated mix, lighter models for straightforward work and stronger reasoning models for harder problems. The lineup shifts as new models ship and our evals show what performs best, so we don't publish a fixed roster. The point is that you get the right model without choosing one. We know some teams need more control here and we're looking into it.

Q: How does billing work for HydraFusion?

There is no separate HydraFusion charge. You pay for the constituent models it runs, so the cost of a turn is the sum of its phases.

Q: Why can't I see HydraFusion's output while it's working?

HydraFusion isn't a single model. It can run several models that draft, review, and revise before producing an answer. Those intermediate passes are working drafts that may be revised or discarded, so we show you the final reviewed result rather than text that could change moments later. We're working on showing more while it runs. Tell us what you would like to see.

Happy Coding!

[aditya81070]

Hey team, thanks for great feature, excited to try this. Just noticed one thing: when in the plan mode and trying to switch to HydraFusion model, it switches back to previously selected model. Can we add a warning or message which specify that HydraFusion will only work in normal mode.

[jukasper]

Hi @aditya81070 , thank you so much for this feedback. This seems to be a bug, we are currently investigating. So thanks for bringing this to our attention.

[shawn-cummins-dsg]

Hi. Regarding the hidden outputs while its running. When i need it to do some research, show me an analysis and recommendation, then ask me for what to do about it, this becomes a problem because all I can see is the question - "What do you want to do about it?" - when I cancel the interaction and it dumps all the info into the cli I can finally see what it thought/wanted from me.

[jukasper]

ok, this is great feedback, so thanks for sharing. We will def. take this into consideration

[casebycase]

I've been patching something like this together using an orchestrator copilotcli, which has instructions on how to force use copilotcli for subagents and logging everything to disk… and the orchestrator knows how to and has authority to cross-validate and critique.

Removes this need for me to do that if i can just pass to /model Hydrafusion to get that behavior so looking forward to testing this.

[jukasper]

thank you for sharing! yes, please let us know how your testing is going, we are super keen on knowing what worked and didn't work for you!

[NillsF]

It looks like starting a session with the copilot --model hydrafusion flag doesn't work.

Model "HydraFusion" from --model flag is not available. Using "claude-sonnet-5" instead.

[NillsF]

I do have it available in the model picker.

[aditya81070]

This is experimental so you might have to do --experimental and then try it out.

[Liebeck]

TLDR: I tried out HydraFusion and it worked flawlessly. (Session ID:  562b9f26-9b74-4987-9de2-7590fa45e905 )

Background

For one of my projects, I tried to fix a bug yesterday using Auto mode (gpt-5.3-codex) in the GitHub Copilot app. Copilot produced a plan but was somehow unable to start implementing the changes, even though I explicitly approved the implementation several times, escalating to me typing "DO THE IMPLEMENTATION" in caps. That session seemed stuck, and I had to manually switch to an Opus model to get any code changes at all. I wasn't happy with the result, so I stopped the implementation there and asked it to summarize the issues found so far, plus the test data, into a markdown file describing new E2E tests. That file was the basis for my HydraFusion test.

Session steps

• Start: claude-sonnet-5 was picked for reading my memory bank files.
• Planning: Cascade ran while the markdown file was being read. I was asked three follow-up questions about the implementation, answered them, and let the CLI start.

{"type":"session.fusion_handoff","data":{"fusionId":"fusion-2d2a6f1d-b6b8-4e55-9f0e-e79903e7afc6","sourcePhaseId":"fusion-2d2a6f1d-b6b8-4e55-9f0e-e79903e7afc6:judge","targetPhaseId":"fusion-2d2a6f1d-b6b8-4e55-9f0e-e79903e7afc6:repair","targetModel":"gpt-5.6-sol","message":{"role":"user","content":"HydraFusion Cascade escalation.\nComplete the original task in the current workspace. The workspace may contain an incomplete prior attempt; inspect it independently, preserve correct work, and fix or finish what remains.\nDo not assume any prior completion claims are correct.\nOriginal task as JSON string:\n\"I want you to work on a bugfix for which I have already written down the problem. I have already added test credentials for the docker environment for which I need you to write tests. Can you read the file @memory-bank\\\\plans\\\\tablet-dashboard-public-access-requirements.en.md and tell me what you think?\""}},"id":"b2d9584f-f3f8-4f7e-bf48-bab7f942d706","timestamp":"2026-09-05T05:04:30.880Z","parentId":"e7f32f40-0748-469d-88a3-0ae0c36616c4"}

• Implementation: HydraFusion ran as Single, probably because I started from an already solid analysis. The CLI did not show which model was actually doing the implementation.
• Tests: executed with claude-haiku-4.5
  However, I feel that haiku was spawned as a subagent and not as a result of HydraFusion

{"type":"subagent.completed","data":{"toolCallId":"call_GWXMe3ms1aBYnzkpiFMr2Im9","agentName":"task","agentDisplayName":"rerun-dashboard-tests","model":"claude-haiku-4.5","firstDispatchedModel":"claude-haiku-4.5","totalToolCalls":4,"totalTokens":116345,"durationMs":397761},"agentId":"2d5b3798-c3b6-4177-8580-6d46df038537","id":"9369e9d9-cc67-4143-a84f-b4787deaa725","timestamp":"2026-09-05T06:02:25.515Z","parentId":"b6fe26bc-e2be-462a-8202-8a7af1593dc3"}

• Code review: done with gpt-5.6-sol

I had multiple rounds of running tests and code review until all new tests passed.

Feedback and suggestions

1. The FAQ asks what we'd like to see while HydraFusion is working: for me the most useful single addition would be the name of the model chosen for the implementation phase. I could see model names for every other step, but not that one.
2. I would love to have a slash command to be able to see which models have been chosen by HydraFusion and which percentage of my session was run with which model. I have taken a look at the events.jsonl but gave up soon.

Why this is important: I ran an online event two weeks back and during a quick poll everyone said that they don't trust auto v1 at all and never use it. So actually have a breakdown of which model ran where could have a big impact in generating trust

However, I ran this prompt "Can you tell me which models have been chosen from HydraFusion, how often they were used, how many input tokens, cached input tokens, output tokens, AIC used and their percentage?"

For some reason, the total AIC counter seems to mismatch for the CLI counter and the generated response 🤔

Suggestion: Can you offer such a breakdown to the user?

3. I saw that the Copilot app and VS Code are targeting September as a fast follow. Coming back to the CLI after months of using the app was a noticeable step down for me 😅 I'd love to test out the next features in the Copilot app 😄

---

If you want additional information (took a bunch of screenshots) or have questions, feel free to reach out

[hallatore]

Is this mode subsidized, or is it just this good?

First impression: It picks models pretty well. The cost/quality ratio is good.
And I think it answers a great problem: It's really hard for me to know what model is "good enough" at any given time.

I've been using it for different tasks from small to big. The overall quality has been "not too bad" compared to running every session with Opus 5. But the credits burn has been 10% of what I'm used to.

1. I hope it can come to vscode insiders soon.
2. I wish I could set a cost/quality temperature. Sometimes I'll rather waste a bit of credits for a overall higher probability of success.

[gcapnias]

My first impression was that it got the job done, successfully planning and implementing a small application. The planning phase took about 500AIC, which I’d say was fine, while implementation required around 2300AIC—more than I expected, as I thought 1800AIC would suffice.

The application ran smoothly with no glitches or noticeable issues during first use. The app was a movie finder built using the TMDB API, impressively completed without having a TMDB key at any point during development.

These are end-to-end the sessions:

O Implement Movie Finder Specification
E:\Workspaces\playgrounds\copilot-playground\flow03 . 43b5b020-6839-43dd-8d96-f28a3371df28 . just now

Build Movie Search App
E:\Workspaces\playgrounds\copilot-playground\fLow03 . 5643a43a-88a7-44d1-8b4d-e83374fa458b . 1h ago

o Setup Matt Pocock Skills
E:\Workspaces\playgrounds\copilot-playground\flow03 . b6451271-e8b7-4fb8-8300-5a8d0f876597 . 4h ago

I’m not sure how to view a breakdown of the model usage or how the tokens were spent. After all, AICs used, it’s 40% of a Pro+ package!

If one wants to review the repository, you can find it here: movie-finder - branch: /feature/movie-finder

[faizankshaikh]

Hi - my overall experience for HydraFusion has been positive. The code worked as expected and the model delivered what was requested. Session ID 8544fe94-5765-401b-a559-3b30739a8e6a

One issue and one question

• The turn ended (it gave final summary, and it answered followup questions), but it still says 'Working'. I presume its a bug

• Can you give a suggestion on when and how often to run /compact when using HydraFusion? When I use orchestration-execution sessions, I rigorously compact executor and frugally compact orchestration. For HydraFusion, its not super clear if the same should be done.

[Wrcz]

I haven't tried HydraFusion yet, but I've read your article and it seems to have great potential. I have a question: can I work with HydraFusion and SpecKit simultaneously? If so, what will the synergy between them be like? I don't know if you've considered this. Thank you for your response.

[aditya81070]

I have started the session with a complex task which involves research. Now, as this is completely research task and there is no execution, I don't have any insights on what model is doing. I used to /ask to get some insight but it answers same thing even after significant interval and HydraFusion ignores the follow up messages.
At this point, I am not sure whether it's doing research or it's just stuck at some blocker. Is there anyway we can at least get some insights/output on what it's doing? If it's not possible to share complete intermediately output, can we share some sort of results or a clear signal that model is making a progress similar to TODOs list ?

I saw below text after a session start so not sure this is directing something.

Static context is using 82% of available input tokens. Automatic compaction will wait longer because system messages and tool definitions cannot be reclaimed. Run /context and reduce context by disabling unused MCP servers or tools, trimming custom instructions, or switching to a larger-context model.