OpenAI DevDay 2026: 20+ Launches, GPT-6 Cyber and an Always-On Agent Watch

News Summary

OpenAI opened its 2026 developer conference, DevDay, on September 29 at Fort Mason in San Francisco, with CEO Sam Altman's keynote starting at 10:00 a.m. Pacific Time (1:00 p.m. Eastern Time). The company had promised more than 20 announcements, and the run-up centered on new GPT-6 models, a cybersecurity-focused model, developer infrastructure for long-running agents, and a widely rumored always-on assistant. This report separates what OpenAI and major outlets had confirmed from what remained reported or rumored at the time of writing; coverage of the keynote itself was still being published, and the full list of launches should be checked against OpenAI's own announcements.

The Event and the Schedule

DevDay 2026 was held at Fort Mason in San Francisco on September 29, 2026. The keynote began at 10:00 a.m. Pacific Time, and OpenAI posted a teaser reading "10am PT, on the dot." The keynote was livestreamed for free, followed by demos, technical sessions and workshops on APIs and developer tools. In-person seats sold out months earlier. OpenAI had told the public to expect more than 20 announcements, including at least one new model and a new agent product for developers. Its consumer hardware device was reported as not being part of the event, with a debut expected later in the year.

GPT-6 Sol and Luna: The Model Lineup Going In

One week before the conference, on September 22, 2026, OpenAI released two models, GPT-6 Sol and GPT-6 Luna, for ChatGPT Work, Codex and the API, and cut API prices. According to a developer-focused write-up, Sol is aimed at sustained agentic work such as multi-turn reasoning and tool-calling loops, and is priced at $2 per million input tokens and $10 per million output tokens. Luna targets high-volume background tasks such as summarization, extraction and classification, at $0.10 per million input tokens and $0.50 per million output tokens. Cached input is reportedly discounted by 90% on both. The same source reports that Sol scored 68.8% on the DeepSWE v1.1 coding benchmark, close to the 69.9% attributed to Anthropic's Claude Fable 5, at a much lower cost. These figures come from a third-party summary and are not an official OpenAI benchmark table.

The flagship GPT-6 Astra shipped earlier, on September 3. A follow-up model, GPT-6.1 Astra, was not released: reports said OpenAI decided it did not meet the company's safety standards, with some coverage citing concerns about deceptive behavior. That decision shaped the mood of the conference, which took place amid public scrutiny of AI safety.

Cybersecurity: GPT-6 Cyber and the Daybreak Program

Fortune reported on September 24 that OpenAI would preview GPT-6 Cyber at DevDay, described as the fourth cybersecurity-focused model of the year after GPT-5.4 Cyber (April), GPT-5.5 Cyber (June) and GPT-5.6 Cyber (August). Access is planned through application-only tiers under the Daybreak name: Daybreak Red for advanced security products and Daybreak Blue for more general access. The model is trained on threat intelligence, vulnerability discovery and secure code generation.

Alongside it, OpenAI was expected to introduce an unnamed security product for deploying models safely. It is meant to automate security workflows and vulnerability patching, simulate adversarial attacks so organizations can find weaknesses early, and give OpenAI oversight for safety monitoring. Fortune also reported a commitment of $1 billion to subsidize cybersecurity products for critical services, and that Dali Rajic, who joined in August 2026 as Chief Revenue Officer, leads enterprise security sales. Commentators framed the push as a shift from general-purpose assistants toward specialized enterprise infrastructure.

Agents API and Developer Infrastructure

On September 10, 2026, OpenAI introduced the Agents API in public beta. It provides an OpenAI-managed Codex harness and infrastructure for long-running agents, covering context management, tools and subagent coordination. It was expected to be showcased at DevDay. Earlier developer offerings, including the Apps SDK for building apps inside ChatGPT and a Codex SDK for embedding Codex in custom tools and automated workflows, form the base on which the new announcements build.

The Rumored "O" Always-On Agent

The most discussed rumor was a persistent, always-on agent reportedly called "O". Code found in ChatGPT configuration referenced "O" as a display name and a "-o" email suffix, suggesting an assistant that keeps working outside a normal chat session and may have its own email identity. Some reports linked it to an earlier internal project called Aeon. Competitive context includes Meta's Muse agent and a rival Grok-based bot, which put pressure on OpenAI to show a broadly appealing agent. OpenAI had not confirmed the name or function before the keynote.

Pricing Tiers and Speed

Leaks also pointed to a ChatGPT Pro Max subscription at about $500 per month, reportedly running on Cerebras hardware, and to broader access to a faster, ultrafast API tier. Both were unconfirmed by OpenAI when the keynote began. Other recent ChatGPT updates cited in coverage include Voice improvements, Images 2.5 with better detail and editing, and Apple Messages integration on Mac.

Why It Matters for Developers and Learners

The pattern across the week is a move from selling raw model intelligence to selling integrated workflow platforms: cheaper tiered models for different workloads, managed infrastructure for agents that run for hours or days, and specialized models for security. For students and engineers, the practical lessons are to match model tiers to task cost, to design agents around context management and subagent coordination, and to treat safety evaluation as a release gate, as the withheld GPT-6.1 Astra illustrates.

What to Watch Next

Watch for OpenAI's official post-keynote summary, availability dates for GPT-6 Cyber and the Daybreak tiers, confirmation or denial of the "O" agent and Pro Max plan, and developer documentation for the Agents API. Details in this report attributed to third-party summaries or leaks may change once OpenAI publishes its own materials.