Copilot Fusion: The Compute Line

title: "Copilot Fusion: The Compute Line" slug: "2026-10-29-copilot-fusion-half-the-cost-compute-line" meta_title: "Copilot Fusion: Costing the Compute Line" meta_description: "Copilot Fusion claims half the API cost per task. The other line item is compute minutes for every agent session, and per-minute billing keeps it small." primary_keyword: copilot fusion author: Krova Cloud Engineering date: 2026-10-29 schema: BlogPosting+FAQPage category: AI Code Sandboxing secondary_keywords: AI coding cost per task, agent compute billing, per-minute microVM pricing, sandbox idle cost, task economics excerpt: "Fusion cuts model cost per task by about half. The forgotten half is compute: every agent session runs somewhere that bills by some unit."

Copilot Fusion: The Compute Line

GitHub's copilot fusion reached general availability at Universe on 28 October, coordinating multiple models per task at about half the average API cost, according to Remio's event rundown. Every cost story this week stops at tokens. The other line item on every agent task is the computer the session runs on, and its unit of billing decides whether the saving survives.

The headline numbers are real. Remio reports Fusion Max at 60 on the Artificial Analysis Coding Agent Index, within two points of Claude Code on Opus 5.5 Max, with roughly 40% less time per task. The Moneycontrol coverage adds that GitHub logged 34.3 billion commits in 2026 as the backdrop, and RuntimeWire frames the orchestration shift.

What copilot fusion changes about task cost

Task cost has three parts: model tokens, compute minutes while the session works, and compute minutes while it waits. Fusion attacks the first part. The second and third depend entirely on where sessions run and how that place meters time.

Halving tokens shifts the mix instead of shrinking the bill proportionally. In our model of a twenty-minute session on vendor-published sandbox rates, the compute share of total task cost roughly doubles when the token line halves. An afterthought becomes a line item worth auditing.

The billing units that decide the compute line

Wall-clock billing charges every second a sandbox exists, including the minutes an agent spends waiting on a model response. Per-second or per-minute meters with power-off charge only for live compute, and the gap between those two policies is pure waste at fleet scale.

Vendor rates published in October 2026 put a busy two-vCPU, four-gibibyte hour near $0.17 on one popular sandbox and near $0.24 on another. We compiled those figures in per-minute microVM costs. Our own published comparison puts Cube pricing 92% below those rates at the same size, and we re-checked that page on 29 October.

Idle discipline matters more than headline rate. A session left running between phases bills like a working one under wall-clock meters. Per-minute hosting economics shows the monthly delta for teams that forget to power off.

GPU phases follow the same logic at steeper rates, and the split pattern in GPU sandboxes for coding agents keeps accelerator minutes attached only to inference windows.

Costing a Fusion-style task end to end

Take a task that dispatches three sessions of twenty minutes each with five idle minutes between phases. On wall-clock billing you pay for thirty-five minutes per session; on per-minute billing with power-off you pay for twenty. At fleet volume that third is the difference between a saving and a rounding error.

Add snapshot restore instead of cold rebuilds and the compute line shrinks again. Boot and install minutes are compute minutes too, and cloned golden images remove most of them.

Verdict

We recommend costing every agent task as tokens plus live compute minutes plus idle minutes, then choosing a sandbox whose meter charges only the middle term. Fusion makes the token line cheaper; per-minute billing with power-off and snapshot restore is what keeps the compute line from eating the difference.

Limitations

• Our task-cost model uses vendor-published October 2026 rates; however, negotiated and spot pricing move both lines.
• The compute-share doubling is arithmetic on stated assumptions, but not a measured invoice from a Fusion workload.
• Benchmark scores describe model quality per task, and we cannot verify GitHub's internal cost accounting behind the half-cost claim.
• Snapshot restore savings depend on image size and registry proximity, which vary per team.

FAQ

Why does halving token cost raise compute's share?

Because the compute line stays constant while the token line falls, so the same minutes represent a larger slice of a smaller total. Auditing meters then matters more, not less.

Which billing unit wastes the least for agent sessions?

Per-minute or per-second metering with power-off between phases, since waiting time stops billing. Wall-clock meters charge for existence rather than work.

Do snapshots really cut compute cost?

Yes, by removing boot and install minutes from every session. A cloned golden image starts where a rebuild would have finished, and those minutes bill like any others.