What changed

GitHub expanded Copilot Fusion to more clients and models, reported Fusion Max on public Terminal-Bench 4.1 leaderboards, and previewed matched-model harness comparisons across Copilot CLI, Claude Code and Codex CLI with run-to-run intervals and cost.

Why it matters

Model selection and harness behavior are now presented as separate evaluation dimensions, giving teams a more useful comparison shape while still relying on vendor-run configurations and benchmark task coverage.

What to do

Reproduce one matched model and reasoning level on a private repository task set, pin every harness version and compare accepted outcomes, review effort, wall time and total cost.

Caveat

Fusion scores and harness results are published by GitHub. Public leaderboards improve inspectability but do not make the configurations independent, and Copilot-only Fusion runs lack matched harness pairs.