OpenAI is rolling out a multimodal preview of GPT-6 to Plus subscribers this week, pairing a redesigned realtime API with voice and vision routing that targets sub-200ms round-trip latency on common prompts.
What changed in the API
The preview ships with a new session model: developers open a persistent websocket, stream audio or frames, and receive structured tool calls alongside natural-language replies. Early testers report smoother barge-in during voice mode and fewer dropped partial transcripts.
Pricing stays on the existing Plus tier for consumer access; API usage is metered separately with a modest uplift on realtime endpoints compared to GPT-5. OpenAI says the uplift reflects dedicated inference capacity reserved for low-latency paths.
Who should care
For builders, the biggest shift is unified multimodal context — a single thread can mix uploaded screenshots, live camera frames, and spoken follow-ups without manually swapping modes. Custom GPTs inherit the same routing automatically.
Unified multimodal context means one thread can mix screenshots, live camera frames, and spoken follow-ups — no mode switching.
If you rely on voice for brainstorming or field notes, revisit prompts that assumed higher latency: shorter clauses, explicit “wait for my pause” instructions, and tool schemas that tolerate partial input all behave better under the new stack.
