KDCube
← AI Industry News

OpenAI Ships Presence, Kimi K3 Opens Its Weights July 27

OpenAI launched Presence on July 22, a limited-GA, FDE-led platform for governed voice and chat agents that already resolves ~75% of its own support calls. Meanwhile Moonshot's 2.8T Kimi K3 opens its weights on July 27 at $3/$15 per million tokens — frontier-adjacent quality that signals the end of rock-bottom Chinese ...

Highlights

Enterprise agent platforms harden

  • OpenAI launched Presence on July 22, a limited-GA enterprise platform for realtime voice and chat agents that run under company-defined policies, permissions, and evaluation standards — deployments are led by OpenAI Forward Deployed Engineers, not self-service (OpenAI · VentureBeat).
  • Presence is already answering OpenAI's own English-language phone support line, where it resolves ~75% of inbound calls without a human — a rare hard adoption number from a launch-day product (VentureBeat).
  • Moonshot's Kimi K3 — a 2.8T-parameter mixture-of-experts model live via API since July 16 — sets its full open-weight drop for July 27, one day before the stateless MCP cutover (The Decoder · Axios).
  • K3 lands 4th on the Artificial Analysis Intelligence Index (57), trailing Fable 5 (60) and GPT-5.6 Sol (59) but matching Opus 4.8 — priced at $3/$15 per million tokens, a sharp jump from K2.6's ~$0.95 input (The Decoder).

Key Signals

  1. OpenAI Presence puts governance in the runtime, not the app

    July 22, 2026

    Presence pairs model reasoning with policies, guardrails, simulations, evaluations, approved actions, and human escalation — the same governed-deployment surface AgentCore and Gemini Enterprise have been racing to own (OpenAI). The catch for builders: access is gated behind Forward Deployed Engineers and select integrators, so this is an enterprise engagement, not a self-serve SDK (VentureBeat · Help Net Security).

  2. Kimi K3 signals the end of rock-bottom Chinese inference

    open weights July 27

    K3's $3 input / $15 output pricing is comparable to Claude Sonnet 5 and roughly triples its predecessor's input cost — a deliberate move up-market as frontier-adjacent open weights arrive (The Decoder). For teams that self-host, the largest open-weight model shipped to date is about to be downloadable, but the cheap-token era that made Chinese models a budget default is closing (Axios).

  3. Two platform deadlines collide at month-end

    July 27–28

    Kimi K3's weight drop (July 27) and the stateless MCP 2026-07-28 spec land within 24 hours of each other. Teams planning a self-hosted K3 evaluation and an MCP transport migration in the same sprint should sequence them, not stack them (Axios).

Why It Matters / What To Watch

  1. Governed agent deployment is consolidating into managed platforms.

    • Compare Presence's policy/guardrail/eval loop against AgentCore Policy (Cedar + natural-language rules) and Evaluations before committing to a stack (OpenAI).
    • If you need self-service, note Presence is FDE-led today — budget for an engagement, or stay on the Agents SDK (VentureBeat).

  2. Re-baseline your inference cost model this week.

    • K3's pricing shift means "open weight" no longer implies "cheapest" — recompute per-task cost against GPT-5.6 Sol (~$1.04/task) and K3 (~$0.94/task) before switching (The Decoder).
    • Watch the July 27 open-weight release if you self-host: frontier-adjacent quality at 2.8T params changes the buy-vs-host math for high-volume workloads (Axios).

Quick Links