GitHub shipped Project HydraFusion on September 4 — a layer that picks the model and the execution pattern for each coding task at runtime. Cost dropped across all three benchmarks, but quality beat Claude Opus 5 in only one of them.
Key takeaways
- Three execution patterns — Single, Cascade and Critique — selected before the task runs
- TerminalBench 2.1: 67% lower cost, 4.9 points higher quality than Claude Opus 5
- DeepSWE: 36% lower cost, 1.5 points lower quality
- CheckpointBench: 65% lower cost, 0.1 points lower quality
- Research preview on every Copilot plan, launched in the CLI with the /experimental flag
Routing instead of one model
HydraFusion does not send every request to the same model. Before starting it picks one of three execution patterns.
The three patterns HydraFusion switches between at runtime:
GitHub describes safeguards around it: execution timeouts, review in a tool-less context, route validation before execution, and a rule that incomplete changes never reach the code.
Cost falls everywhere, quality does not
The strongest result is TerminalBench 2.1 — a 4.9-point lead over Claude Opus 5 at 67% lower cost. In the other two measurements the balance flips.
| Benchmark | Cost vs Opus 5 | Quality vs Opus 5 |
|---|---|---|
| TerminalBench 2.1 | −67% | +4.9 pts |
| DeepSWE | −36% | −1.5 pts |
| CheckpointBench | −65% | −0.1 pts |
The “frontier-level quality” claim therefore rests on one measurement out of three.
So far, the reasoning and task solving capability is at or better than Opus.
A Principal Software Engineer at Microsoft, quoted on the GitHub blog.
Availability and limits
The feature ships as a research preview?Research preview: An early version of a feature released publicly for testing, with no stability guarantee and no commitment that it ships to production. on every GitHub Copilot plan and runs in Copilot CLI behind the /experimental flag. Billing follows each model's standard token rates, and the scope covers only single-turn tasks for now.
Why it matters
Model routing moves the decision of what to run from the user to the tool vendor. For teams paying per token, a double-digit saving is real and immediate. The price is predictability: the same prompt can take a different route and return a different result, and two of three benchmarks landed below the baseline model. That argues for measuring quality on your own repository before taking someone else's numbers as a guarantee.
What's next?
- HydraFusion handles single-turn tasks only — extending it to longer agent sessions has no announced date
- The /experimental flag means Copilot's default path stays unchanged until the research preview ends
- GitHub published a separate piece on AI coding costs on September 2 — spend optimisation is becoming a core Copilot theme
