BuildNot yet confirmed elsewhere1 publisher2 min readPublished
GitHub's HydraFusion preview plans each Copilot prompt as a workflow of model calls
GitHub says its HydraFusion preview costs an estimated 67% less than Claude Opus 5 on TerminalBench 2.1 by planning model calls per prompt. Whether that saving reaches your bill depends on how often a cheap draft clears the quality gate.
The Engineer · Build desk
Three figures. GitHub estimates HydraFusion costs 67% less than Claude Opus 5 on TerminalBench 2.1. Auto model selection gives a 10% discount that HydraFusion does not get. So on a prompt sent to the same model, HydraFusion costs roughly 11% more than Auto.
| Figure | Value | Note | Claim |
|---|---|---|---|
| Est. cost vs Claude Opus 5 | 67 % lower | GitHub estimate on TerminalBench 2.1 | 16 |
| Auto model selection discount | 10% | Does not apply to HydraFusion | 6 |
| Same-model prompt vs Auto | 11 % more | Full rate vs Auto's 90% | 15 |
What happened
- Each prompt gets one of three plans: a direct solve by one model, a cheap draft behind a quality gate that may escalate, or a draft reviewed by a read-only critic from another model family before one revision.
- GitHub published HydraFusion on 4 September in Copilot CLI and extended it to VS Code and the GitHub Copilot app on 30 September.
- GitHub bills every model HydraFusion uses at that model's standard rate, with no separate HydraFusion charge and no Auto discount.
- GitHub's guidance is to start with first-turn, single-prompt coding tasks, and it names multi-turn performance as the next focus.
- The documentation labels HydraFusion a research preview with no SLA and says it is not intended for production workloads.
Why it matters
- cost On a prompt that Auto and HydraFusion's direct plan send to the same model, HydraFusion costs about 11% more, because the 10% Auto discount does not apply.
- contradiction GitHub's 67% is measured against Claude Opus 5, yet its documentation positions HydraFusion against Auto and says additional model passes "may be worth the extra time and AI credits".
- constraint Teams on JetBrains, Visual Studio or the cloud agent cannot test it, and Business and Enterprise users still need an administrator to enable preview features after the 22 October default-policy change.
- decision Teams that hand-built draft-then-critique or cheap-then-escalate code now have to decide whether to keep it or move to a runtime that offers per-leg cost accounting, bounded execution and isolated review.
The picker survives. HydraFusion is an entry in it, and the documentation says the entry does not appear when no model qualifies under your plan and policies [11]. Auto model selection, which makes one choice per request, stays alongside it [6]. So HydraFusion adds a third option.
The choice among plans is made by a scorer. Per prompt it rates the task on reasoning, code generation, debugging and tool use, then picks, in GitHub's words, "the least complex workflow expected to meet its needs, using additional model calls only when they are likely to improve the result" [10].
Each plan has its own cost shape. A direct solve is one call. A draft-then-critic plan is at least three billed legs: draft, review and one revision [14]. A draft-then-escalate plan pays for the draft every time and for the strong model only when the gate fails. In a simplified model, it beats calling the strong model directly only when the draft costs less than the strong call multiplied by the share of prompts that never escalate [13].
Now the 67%. The dev.to write-up says the figure deserves scrutiny and that the architecture is the more durable story [17]. As a ratio, 67% lower means HydraFusion costs about a third of what Opus 5 does on that benchmark, or Opus 5 costs roughly three times as much [19]. GitHub's figure is an estimate, and it is paired with "higher verified quality" [16]. The write-up's opening sections do not describe how the estimate was built or which plans produced it.
For the number to transfer, four things would have to hold on your workload. Your tasks would resemble the benchmark's. Your cheap model would clear the gate about as often. The price gap between your cheap and strong models would match the one in the benchmark run. And the estimate would match what your plan bills. We would test it directly: a fixed set of your own prompts, run once under Auto and once under HydraFusion, with credits divided by accepted changes.
Two design choices look right to us. The critic is read-only and comes from a different model family [1], so it can inspect the draft but cannot change it. The runtime also keeps the main conversation on one model where it can, to keep the benefit of cached tokens, while assisting models such as a reviewer "receive only the context they need" [7].
What to watch
- Whether GitHub publishes the TerminalBench 2.1 method: task count, which plans ran, and how cost was estimated.
- Whether multi-turn support, which GitHub names as the next focus, lands and gets its own cost numbers.
- Whether HydraFusion leaves preview or reaches JetBrains, Visual Studio or the cloud agent, which the documentation currently excludes.
Clarity's read
What the record supports and how the coverage leans. The claims behind it follow.
Reality
- Evidence35
- Adoption
- Insufficient
- Hype gap+20
- Incentives55
- Confidence55
Claim ledger
Ranked by verification strength, evidence, and original report placement.
- [1]
Instead of choosing a model, the runtime builds an execution plan per prompt: one model solves it directly, or a cheap model drafts and a quality gate decides whether to escalate, or one model drafts and a read-only critic from a different model family reviews it before a single revision.
- [2]
GitHub published HydraFusion on 4 September 2026, launching it in Copilot CLI; it reached VS Code and the GitHub Copilot app on 30 September.
- [3]
HydraFusion is a research preview; the documentation says "there's no service level agreement (SLA), and HydraFusion isn't intended for production workloads."
- [4]
The documentation lists only Copilot CLI, VS Code 1.140 or later (or Insiders) and the GitHub Copilot app; Copilot cloud agent, JetBrains IDEs, Visual Studio and Copilot Chat on github.com are not included.
- [5]
Per the documentation, "you're billed for each model it uses, at that model's standard rate. There's no separate charge for HydraFusion, and the discount for auto model selection doesn't apply."
- [6]
Auto model selection makes the model decision for you, one model per request, and gives a 10% discount.
- [7]
The documentation says HydraFusion keeps the main conversation on the same model whenever possible to keep the benefit of cached tokens, and that assisting models such as a reviewer "receive only the context they need".
- [8]
GitHub's scope guidance says first-turn, single-prompt coding tasks are the best place to start, and that multi-turn performance is explicitly the next focus.
- [9]
The documentation says to use Auto for everyday work and HydraFusion for tasks such as fixing a complex bug or making a change across several files, "where additional model passes may be worth the extra time and AI credits".
- [10]
Per prompt, HydraFusion scores the task on capability signals (reasoning, code generation, debugging, tool use) and picks "the least complex workflow expected to meet its needs, using additional model calls only when they are likely to improve the result".
- [11]
The documentation says HydraFusion only uses models that your plan includes and your policies allow, and that it does not appear in the picker if none qualify.
- [12]
For Copilot Business and Enterprise, an administrator must enable preview features; the new default enablement policy taking effect on 22 October does not change that because "preview features remain opt-in".
ReportedSupportedSource: GitHub changelog of 30 September and documentation, per the dev.to write-upView cited source - [13]
In a simplified model where an escalation costs one full strong-model call S on top of the draft, a draft-then-escalate plan costs draft + p*S, where p is the share of prompts that escalate. It beats calling the strong model directly only if draft + p*S < S, that is, draft < (1 - p)*S.
- [14]
A draft-then-critic plan runs at least three billed legs: the draft, the critic's review and the single revision.
- [15]
For a prompt that Auto and HydraFusion's direct plan send to the same model, HydraFusion pays the full rate while Auto pays 90% of it, so the HydraFusion bill is about 1 / 0.90 = 1.11 times higher, roughly 11%.
- [16]
GitHub's headline number is 67% lower estimated cost than Claude Opus 5 with higher verified quality on TerminalBench 2.1.
ReportedContestedSource: dev.to write-up reporting GitHub's headline claim2 sources— create a free account to open themView cited source - [17]
The write-up says GitHub's headline number deserves scrutiny, and that the more durable story is architectural.
ReportedContestedSource: dev.to write-up author2 sources— create a free account to open themView cited source - [18]
Teams have been hand-rolling draft-then-critique and cheap-then-escalate patterns in their own agent code for eighteen months; HydraFusion moves those patterns into the Copilot runtime with cost accounting per leg, bounded execution and isolated review.
- [19]
A 67% cost reduction leaves 1 - 0.67 = 0.33 of Opus 5's cost, so HydraFusion costs about one third of Opus 5 on that benchmark, and Opus 5 costs about 1 / 0.33 = 3.0 times as much.
Sources
1 independent publisher whose own reporting we read for this story.
- dev.toProject HydraFusion: GitHub Copilot Stops Picking a Model and Starts Building a Workflow
1 article · October 12, 2026
Topics and entities
Follow any of these and your For You feed starts watching them — no settings page required.
Entities
- GitHubFollow
- GitHub CopilotFollow
- Project HydraFusionFollow
- Claude Opus 5Follow
- TerminalBench 2.1Follow
- Copilot CLIFollow