May 2026 Recap

Two production-facing shifts stand out. First, the LLM Gateway adds a one-parameter “reasoning” toggle and automatic JSON repair. The toggle promises unified access to provider-specific reasoning modes, but there’s no evidence it improves task accuracy or reliability. Expect higher latency and cost, plus uneven behavior across providers. The JSON repair is more consequential: it reduces brittle tool integrations but risks silently masking model errors unless you log originals and repairs.

Second, the streaming stack gets materially more deployable. Per-word speaker labels and “unknown” tags reduce misattribution and cross-talk issues—useful for call analytics, compliance, and agent coaching. The claimed error cuts are meaningful, but there’s no dataset detail (language, accents, domains), so validate on your traffic. Continuous partials offer mid-turn updates every ~3 seconds, improving long-form dictation and IVR capture, but will drive token, bandwidth, and UI churn costs. Build guardrails for rate, debouncing, and reconciliation of revisions.

Streaming PII redaction is the compliance headline: it removes sensitive data in real time and disables partial transcripts by default to prevent leakage. This is the rare vendor default that favors privacy over UX, but it also means slower agent feedback loops unless you explicitly re-enable partials and accept exposure risk. Region availability (US/EU) reduces data residency friction.

Playground and voice demos are noise for decision-making. The operational work is governance and observability: flagging redaction/partials at session-level, monitoring diarization accuracy by line of business, metering costs from continuous partials and reasoning modes, and keeping audit logs for JSON repairs and data handling.