[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

[pas-mllr]

Sometimes you have to manually switch from plan to agent, set it to autopilot and tell it that you have switched it to agent…

[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.

[jukasper]

Thank you @hallatore for your feedback. The models are not subsidized, depending on user tasks we are picking the right execution pattern, we are talking about those in our blog post here, but let me know if this isn't enough information: https://github.blog/ai-and-ml/github-copilot/project-hydrafusion-frontier-quality-via-multi-model-orchestration/

1. yes, we are working on that
2. This is a real dimension and we think about it a lot. As of today, HydraFusion is already deliberately willing to spend more when it raises the probability of getting it right. that's exactly what the review-and-repair passes are for. It's not trying to find the cheapest path through your request. So this should already be built-in, but please let us know or share whenever you thought. HydraFusion didn't do the best job and another model would have come up with a way better output.

[hallatore]

Overall very impressed with the cost savings I'm seeing. And that I finally can "ask whatever" and not deal with picking the right model for the right task.

I wish it showed the output, but rather grayed out the stuff it reverted. Seeing what the model is thinking is important, even when it's wrong.

I'm used to vscode chat where I can see everything. So CLI itself is a bit of a learning curve. Not seeing command output or thinking.

[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

[gcapnias]

Last week I though it was HydraFusion that overspend my AICs, this week I has a similar experience with just the Copilot CLI.

It was the same pattern, an interactive session that controlled subagents implementing tasks. The truth is that is easier in the default agent to monitor the usage, so I had to file an issue "Subagents executing long tool-call sequences in a single turn fail prompt caching and compound token consumption" on this. When cashing is the only "tool" against overspending AICs, subagents should respect it.

[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.

[sirredbeard]

I came here to echo this sentiment and you captured it exactly. What is the model doing? Which model? I don't even have the visibility that you do.

[sirredbeard]

I made something that works for me for now. I love HydraFusion.

[hzarate-clgx]

I'd like to invoke hydrafusion via cli however I get an error currently even when using the experimental flag
Using Copilot CLI 1.0.83

$ copilot --experimental --model hydrafusion -p "hello world"
Error: Model "hydrafusion" from --model flag is not available.

--model auto does work however

[NillsF]

I found a workaround (copilot did).

Windows:

PS C:\Users\nilfran> $env:COPILOT_CLI_ENABLED_FEATURE_FLAGS = "HYDRAFUSION_ROLLOUT"
PS C:\Users\nilfran> copilot --model hydrafusion

Mac:

export COPILOT_CLI_ENABLED_FEATURE_FLAGS=HYDRAFUSION_ROLLOUT
copilot --model hydrafusion

[hzarate-clgx]

Thank you @NillsF! This is what I ended up with on mac with your guidance

COPILOT_CLI_ENABLED_FEATURE_FLAGS=HYDRAFUSION_ROLLOUT \
  copilot --model hydrafusion --context default \
  -p "hello world"

[NZCypher819]

Hi, I've been using Hydrafusion a lot since it came out, comparing it to other ways of working with squads etc. I got to say I like HydraFusion but there is really only one issue Ive had, it often asks me questions about something, but the explanation text is hidden, and I have to get it to explain the context of the question in chat first. Im using GHCP CLI for this. Apart from that, the results are pretty good, its fast become my favourite "model"

[njdnet101]

Nice. Becoming my default for non frontier (Fable/Astra) work.

To what extent do I have to still care about context window management? Is there any concrete guidance in this area? Compacting, clearing, etc

[matthewteeter]

Thank you for building this - it really saves a lot of cost without a noticeable degradation in output! Recently completed the start of a project and mai-code-1.1-flash was able to be used about 30% of the turns for very little cost - wonderful! Will it come to the rich client soon (non-CLI)?

[hallatore]

I feel cancelling with HydraFusion is not instant enough. It seems to work a lot before finally stopping.
Cancelling should hard stop whatever is going on.

[aditya81070]

The Hydrafusion do not show what it is taking approval on when asking for confirmations. For example, in the below screenshot, it is asking me to approve review plan it created but I can't se the review plan so I have to select other and then ask it specifically to show me the plan.
I do understand that thinking tokens might not be available right away but it should at least show some output.

[sunmeat]

HydraFusion is an interesting step toward agentic orchestration

The most important distinction here is that HydraFusion is not simply another model endpoint. It introduces an orchestration layer that can select between different execution strategies, including single-model execution, cascading, and independent critique followed by revision. That makes the feature substantially different from simply choosing a stronger model manually.

The architecture also raises an important question around observability. The current decision to hide intermediate model outputs makes sense from a UX perspective because intermediate drafts can be discarded, but it becomes problematic for tasks where the intermediate reasoning represents useful research or evidence that the user explicitly needs to inspect. This has already surfaced in the discussion.

A useful direction would be to expose structured execution telemetry rather than raw chain-of-thought, for example:

• Which execution strategy was selected: single, cascade, or critique
• Which model families participated
• Whether escalation occurred
• Number of orchestration stages
• Aggregate token usage and estimated cost
• Whether a critic rejected or modified the initial result
• A concise summary of what changed between draft and final output

That would provide transparency without exposing private chain-of-thought or internal reasoning.

I also think the distinction between model selection and workflow selection is the most interesting part of HydraFusion. Auto chooses a model for the request, whereas HydraFusion chooses how the request should be executed, potentially involving several models. That is a much more general abstraction and could become particularly valuable for coding agents where planning, implementation, testing, and review have different capability requirements.

One additional feature I would strongly consider is an explicit execution trace that can be exported alongside the session ID. That would make debugging unexpected behavior much easier, especially when the final result is incorrect but the individual stages appeared reasonable.

The research-preview approach makes sense here. Real-world usage should reveal whether the additional orchestration quality justifies the latency and token costs across different classes of coding tasks.

[ciacco85]

I can't see hydrafusion, may be because i'm using copilot under ****.ghe.com? If it isn't available on ghe, do you have any ETA?

[JanIvarM]

Does HydraFusion’s internal routing strictly honor our enterprise-level disabled-model policies for every workflow phase, including critic, judge, fallback and escalation models?

Is the model usage covered by existing data processing agreements?

[hallatore]

Any plans on updating HydraFusion to support latest models like Opus 5.5 and GPT-6.1 Sol?

[Astrophizz]

I'm curious about this because I have seen it pick GPT-5.6 Terra, which AFAIK loses outright to GPT-6 Luna and GPT-6.1 Sol in cost/quality.

[hallatore]

This is the main reason I've stopped using HydraFusion for now. If it could route to those two models I would probably jump straight back.

[philjones88]

Does it respect the enabled list of enterprise copilot enabled models?
For context we struggle with auto because it does not and it uses get old outdated models over cheaper and better new models.

[spoltier-SCOR]

Same here, it would be useful to have documentation generally on how the curated model mix is managed

• does automatically update to newer versions within the same family + size class ?
• how often is it revised ?
• what restrictions / nudges can be set by the user ?

[fowl2]

Is it possible to use HydraFusion with BYOK? We have limited GH credits but use Azure/OpenAI + AWS/Anthropic behind a corporate billing/auth proxy.

[sirredbeard]

It is possible with Lerna, and experimental plugin, to route to Microsoft Foundry on Azure, where you can deploy OpenAI and Anthropic models. This is a temporary hacky solution while the team builds out HydraFusion. See the caveats, for example, the models must match 1:1 with the models HydraFusion expects from GitHub.

[lvde0]

I hope they will add official BYOK support, but I wouldn't hold my breath. This is probably one of the new Copilot "selling points" that will help trapping users in the expensive Copilot model pricing.

[Csaba8472]

I've been testing the Copilot CLI over the past few days and used HydraFusion exclusively. Looking at my dashboard, the bulk of my spend somehow went to GPT-5.4 instead of the newer 6 or 6.1 Sol or Luna models. I thought the point of HydraFusion was to route intelligently to modern, cost-effective models, but 5.4 took up most of the bill and doesn't even have the Auto: prefix attached to it. Does anyone know why HydraFusion leans so heavily on 5.4?

[lvde0]

It's cheaper, I guess? Still the biggest gripe with these "Auto Routing" features is the complete lack of transparency and control. They just mention a "curated mix" of models but that could literally mean anything. At least you're currently just billed for the actual model usage. Might change in the future, I fear.

[Csaba8472]

GPT-5.4 is both worse and ten times more expensive than luna

[JKomoroski]

Feedback: can't tell what HydraFusion is doing during long runs (Copilot CLI)

Copilot CLI v1.0.92

During longer agentic tasks, the progress display only shows

◉ HydraFusion · Single · working · 32m 34s
◐ Main pass · primary worker
◐ Running tool operation
Receiving model output · 13.1 KiB
Esc to interrupt.

I can't tell which tool is running or whether the task is progressing. The Ctrl+T thinking toggle has no visible effect with HydraFusion.

Requests:

• Show the tool name and a short summary of the call in progress (command, file, or query)
• Make Ctrl+T work with HydraFusion, or state that it doesn't apply.
• Option to stream pass-level output or reasoning summaries
• Improve task visibility