PRE-KEYNOTE BRIEF · no announcements confirmed yet Research lock: Sep 29, 6:50 AM PT

OPENAI DEVDAY · TUESDAY, SEPTEMBER 29

DevDay starts at noon CT.

The real question is not “which model?” It is whether OpenAI turns persistent agents, ultrafast inference, and security-gated capability into a coherent platform.

LIVE FEED

DevDay, as it happens.

The keynote starts at 12:00 PM CT. Confirmed announcements land here as they happen — this feed refreshes itself, no reload needed.

Waiting for the keynote…
    01 / SIGNAL

    Start with the line between fact and rumor.

    As of the research lock, the keynote had not happened. These are the confirmed pre-event facts versus the claims today must settle.

    CONFIRMED CONTEXT
    • Agents API is in public beta.Managed Codex harness, durable sessions, context compaction and hosted or self-hosted sandboxes. No fee beyond tokens, tools and container time.
      Source ↗
    • GPT-6 Sol and Luna launched Sep 22.API pricing: Sol $2 input / $10 output; Luna $0.10 / $0.50 per 1M tokens—reported at about half GPT-5.6 pricing.
      Source ↗
    • GPT-Live 1 is generally available.Full-duplex voice, reported at $0.05/min plus model and tool usage.
      Source ↗
    • Assistants API was retired Aug 26.Migration path: Responses API or Agents API.
      Source ↗
    UNCONFIRMED
    • “o,” an always-on assistant.Product strings briefly surfaced in ChatGPT. Evidence supports something in development—not a DevDay launch or an API.
      BleepingComputer ↗
    • GPT-6 Cyber and “Daybreak.”Fortune reporting is credible, but revised timing reportedly removed a confirmed DevDay unveiling; “coming weeks” is the safer frame.
      Timing caveat ↗
    • $500/month Pro Max.A leak, not an announced plan.
      Source ↗
    • UltraFast consumer/API expansion.The underlying speed exists; a consumer rollout rests on lower-credibility reporting.
      Rumor source ↗
    Do not overread the hardware chatter. Court-filing reporting says OpenAI’s Jony Ive-designed device will not ship before the end of February 2027; a DevDay reveal remains possible but is weak speculation. Source ↗
    02 / RUN OF SHOW

    What is actually on the menu.

    Only the opening keynote is confirmed for livestream. The rest of this agenda comes from multiple outlets citing OpenAI’s event page; breakout names were not available.

    8:00 AM PT
    Breakfast / attendee program opens
    IN PERSON
    10:00 AM PT
    Opening keynote · Sam Altman
    LIVESTREAM
    11:15–3:30
    Breakouts, technical programming, API/tool demos and workshops
    RECORDED LATER
    11:30 AM PT
    Lunch
    IN PERSON
    4:00 PM PT
    Closing session
    IN PERSON
    4:45–7:00
    Reception
    IN PERSON

    Venue: Fort Mason Center for Arts & Culture, San Francisco. Applications are closed; no day-of tickets. Invited attendees reportedly pay $650. Agenda source ↗

    03 / LIVE WATCHLIST

    Six questions that decide whether this matters.

    Ignore adjectives. Capture availability, interfaces, constraints, and price.

    Does “o” exist?

    The strongest rumor is an always-on assistant that works beyond a chat window. The developer hinge: consumer feature, workspace agent, or API platform?

    Listen for: permission scopes, approval gates, email access, memory boundaries, API availability.

    Agents API: beta to what?

    The platform is already public beta. Today could supply the pieces that turn it into an enterprise control plane.

    Listen for: GA date, pricing, SLA, regional residency, zero-data-retention, sandbox policy.

    Is speed a tier?

    ~750 tok/s inference is technically real. What is unconfirmed is packaging it as a consumer toggle or paid API processing class.

    Listen for: model limits, price multiplier, latency guarantees, quality tradeoffs.

    How gated is cyber?

    Astra reached OpenAI’s “Critical” cyber capability level. A GPT-6 Cyber preview could arrive with identity, attestation and hardware-key controls.

    Listen for: access criteria, auditability, deployment model, Daybreak Red versus public API.

    What breaks next?

    A secondary source reports Chat Completions shuts down Nov 30. This is not officially verified—treat on-stage confirmation as a migration alert.

    Listen for: deprecation dates, Responses parity, tool-calling behavior, model aliases.

    Where does margin move?

    Recent Sol/Luna cuts make high-frequency agents cheaper. Another price move—or a premium Pro Max tier—would sharpen segmentation.

    Listen for: cache rates, batch discounts, long-context premiums, fast-tier billing.
    04 / IF–THEN

    The implications are architectural, not cosmetic.

    The decisive layer is who owns long-running state, permissions, runtime isolation and cost control.

    For builders

    • Responses migration becomes the first budget line if a Chat Completions sunset is confirmed.
    • Sol/Luna plus durable agent sessions could make always-on products economically viable.
    • UltraFast can change voice and agent UX—but only after task-level evals show what speed trades away.

    For enterprise

    • Residency, ZDR and human-approval controls matter more than demo quality.
    • Key-expiry governance is a compliance positive; persistent agents expand the approval surface.
    • Cyber models may require applications and attestations, making access itself a procurement variable.

    For the industry

    • OpenAI and Anthropic are competing on cost per completed task, not only benchmark peaks.
    • The control plane—sandboxes, identity, oversight—looks like the next enterprise moat.
    • Capability gating benefits open-weight alternatives even as it lowers frontier-model risk.
    05 / REACTION PROTOCOL

    Prepare to decide, not merely to watch.

    A compact operating plan for turning announcements into action without chasing stage demos.

    Before noon CT 20–30 MIN

    1. List every production dependency on Assistants API, Chat Completions, model aliases and tool calling.
    2. Record current cost, latency and quality baselines for one representative workflow.
    3. Open a clean note with three columns: claim / official source / action threshold.

    During keynote CAPTURE EXACTS

    1. Write model and API identifiers verbatim.
    2. Separate “available today,” beta, preview, waitlist and future roadmap.
    3. Capture price units, regional limits, migration dates and links—not marketing adjectives.

    First 30 minutes after TRIAGE

    1. Verify against OpenAI’s own docs or announcement pages.
    2. Classify each item: ignore / evaluate / migrate / commercialize.
    3. Do not rewrite a roadmap from the keynote alone; require documentation and an API response.

    Same day SAFE REACTION

    1. For a real API launch, build one narrow canary behind a feature flag.
    2. Re-run the same quality, latency and cost eval—include failure recovery.
    3. For client-facing claims, publish only confirmed availability and explicit constraints.

    Capture sheet

    • Exact product / model name
    • Availability date and region
    • API, SDK and endpoint
    • Input / output / tool pricing
    • Latency or throughput promise
    • Data retention and residency
    • Permission and approval model
    • Migration deadline
    • Official documentation URL
    • One workflow worth testing
    06 / WHY TODAY

    The week arrived preloaded.

    The keynote lands amid a live price war and a race to define the agent safety stack.

    SEP 22

    OpenAI launches GPT-6 Sol/Luna and cuts API prices; Anthropic ships Opus 5.5 roughly two hours later.

    Source ↗
    SEP 24–25

    Fortune reporting raises expectations for GPT-6 Cyber, a security product and “a dozen or more” launches.

    Source ↗
    SEP 26

    “o” strings surface on a ChatGPT upgrade page and in client code.

    Source ↗
    SEP 28

    Anthropic and NVIDIA announce an agent safety platform—one day before DevDay.

    Source ↗
    One more signal: OpenAI says its internal research organization now runs 3.1 agent-workdays per human workday. That is less a product launch than a reference deployment for agentic coding at scale. Source ↗
    SOURCE DESK

    Research conducted approximately 6:30–6:50 AM PT on September 29, before the keynote. No keynote announcements were confirmed at that time. Re-check official OpenAI documentation before acting on any rumor, availability claim or migration date.