Operators can now set a hard cap on per-session spend (sessions pause at budgetreached rather than continuing to charge); a designated advisor model can be called mid-turn for strategic guidance without spawning a full second agent; inferencegeo pins inference to US or global for data-residency compliance.
Why it matters: Budget caps make agentic deployments commercially predictable; geo control unblocks EU and data-residency compliance requirements that have been blocking enterprise AI adoption.
A new ROI section in the Copilot Impact Dashboard pairs licensing cost-per-developer against pull request output, split by agent-first vs passive users and expressed as percent of payroll; Code Review's Lite and Balanced effort levels (replacing Low/Medium) are now generally available with org-wide defaults.
Why it matters: First-party ROI data from GitHub gives teams concrete justification for AI tooling spend — a meaningful shift from relying on third-party benchmarks or internal estimates alone.
Users can now upload and query files mid-voice-conversation; GPT-Live Voice also runs inside Projects, inheriting project instructions and referencing recent project chats.
Why it matters: Voice users can now analyze a document or data file mid-call — opens up document-in-context voice workflows that weren't previously possible in ChatGPT.