Skip to content
NewsProduct launch

GitHub adds a Max tier to Copilot's HydraFusion multi-model orchestrator

· by Pondero Newsdesk

The short version

GitHub's October 14 changelog added HydraFusion Max for higher-effort tasks and refreshed the orchestrator's model pool to Claude Opus 5.5, GPT-6 Astra, GPT-6.1 Sol, and GPT-6 Luna, with both options still in research preview.

GitHub adds a Max tier to Copilot's HydraFusion multi-model orchestrator

Copilot's HydraFusion just split into two tiers, and the real change is how it decides a draft is good enough before paying for a stronger model. A new judging step can send a cheap first pass back for a targeted repair instead of escalating straight to a pricier model, and Critique can now run several critic models in parallel when a task's error cost justifies the extra spend. GitHub detailed both changes, plus a new HydraFusion Max tier for harder work, in an October 14 changelog entry, per GitHub's changelog.

What changed

HydraFusion's model pool now draws from Claude Opus 5.5, GPT-6 Astra, GPT-6.1 Sol, and GPT-6 Luna, with the mix used for a given task varying by which orchestration workflow handles it, GitHub said. Reasoning effort is no longer fixed at each model's default: the system now assigns medium effort to initial drafts and high effort to review steps, a split GitHub says keeps drafting fast while putting deeper reasoning only where checking happens. Two of HydraFusion's three execution modes picked up structural changes. Cascade, which previously sent a task to an efficient model and escalated to a stronger one only when needed, can now run a judging step that inspects the first model's work and asks for a repair, rather than escalating wholesale. Critique, which runs a second model to check the first one's output, moves away from leaning on text summaries and instead lets its critic models inspect the actual workspace directly, and now runs multiple critic models in parallel instead of one.

The new HydraFusion Max option sits on the same architecture and reuses the same Cascade and Critique mechanics, but applies higher reasoning effort by default, aimed at tasks GitHub categorizes as needing the deepest reasoning and most thorough analysis rather than a balance of speed and cost. GitHub's own comparison table describes the original HydraFusion as "moderate token savings" for tasks that need a balance of speed, cost, and performance, versus Max's "frontier capability" framing for more demanding work, while saying Max still factors in token efficiency through its orchestration rather than abandoning cost control outright. Both options are available now across VS Code, the Copilot app, and Copilot CLI, on Pro, Pro+, Business, and Enterprise plans, gated behind each surface's experimental or preview feature flag and subject to an organization's model policy. Neither HydraFusion nor HydraFusion Max has exited research preview; GitHub's changelog states plainly that both remain there after this update.

The upgrade is a workflow change, not a bigger model

The practical shift for a Copilot admin is in how review happens, not just which models are available. Cascade's new judging step means a cheap first pass gets checked and sent back for a targeted repair before the system decides to pay for a stronger model, instead of jumping straight to escalation. Critique's move to workspace inspection, plus running several critic models at once rather than one, raises the compute spent on review itself. GitHub frames Max as still tracking token efficiency through orchestration even at its higher effort setting, but a team choosing Max over base HydraFusion is deliberately picking depth of review over the cost profile the original mode was built to protect. Because billing accrues from the tokens each model in the selected workflow consumes, at that model's standard rate, the question for an admin evaluating Max is not whether it costs more in the abstract. It is whether a given task's error cost is high enough that paying for parallel critics and high-effort review steps beats a single escalation to one strong model.

Context and reactions

GitHub shipped the original HydraFusion as a research preview on September 4, confined to Copilot CLI, with Single, Cascade, and Critique as its three execution patterns and an initial model pool built around Claude Opus 5 as the comparison point. GitHub's own offline benchmarking at that time claimed up to 67% lower estimated cost than Claude Opus 5 alone on TerminalBench 2.1, with task quality 4.9 percentage points higher, and a 65% cost reduction on CheckpointBench. Those figures came from GitHub's controlled evaluations rather than production data, a caveat that still applies to this update: the October 14 changelog makes no new benchmark claim for the refreshed model pool, the retuned reasoning-effort split, or HydraFusion Max, so there is no public number yet showing whether the changes move the needle on GitHub's own prior tests. The October 14 release also folds in VS Code and the Copilot app as surfaces, following HydraFusion's earlier expansion beyond CLI-only availability on October 1. The underlying model roster has also moved since September: Claude Opus 5.5 is the version GitHub has been steering Copilot admins toward since a deprecation cutover completed October 2, after an earlier notice had named the now-superseded Claude Opus 5.

What to watch next

Watch whether GitHub publishes updated TerminalBench or CheckpointBench numbers for the new model pool and the Max tier, since the September figures were tied to a configuration that has since changed. Also watch whether HydraFusion's preview-feature gate lifts to default-on the way other Copilot model-picker additions have, since both HydraFusion and HydraFusion Max remain behind that flag after two rounds of expansion.

Sources