OpenAI hosts its annual developer conference in San Francisco on Tuesday, September 29, with the opening keynote by Sam Altman starting at 10 a.m. Pacific at Fort Mason. The event lands at a crowded moment for the company: new models already shipped this month, training on its most capable systems is paused pending a safety review, and reports point to an always-on agent as a likely headline.
The keynote will be livestreamed free, and technical sessions run through the afternoon before a closing session at 4 p.m. OpenAI has also scheduled DevDay Exchanges in Bengaluru, Tokyo, Seoul, Paris, Berlin, London, Sao Paulo and Mexico City, extending the format beyond the main stage. Registration for the in-person event closed months ago at $650 a seat, so the audience watching live is largely the one that got in early.
What is actually expected
This is not a model-drop event. OpenAI released GPT-6 Astra, its computer-use flagship, on September 3, and followed with GPT-6 Sol and Luna on September 22 at roughly half the API price of their predecessors. Anthropic shipped Claude Opus 5.5 the same day, and Google put out Gemini 3.8 Flash earlier in the month. The models are out. What DevDay answers instead is what developers can build on top of them.
Reports by TestingCatalog on September 26 suggest OpenAI is preparing an always-on agent that could launch as “O.” The outlet found references to the name across ChatGPT’s configuration, including a display name and an email suffix configured as “-o,” and noted the agent was recently listed as a benefit on the upgrade page for the $100 Pro plan. The design points to a persistent agent that keeps working outside a normal chat session, potentially with its own email identity from the start. OpenAI has not confirmed any of it, and the proximity of the report to the event is a reason to watch rather than evidence of a scheduled release.
Developer-facing announcements are also expected around agent sessions, multi-agent orchestration and hosted sandboxes, capabilities that already appear in OpenAI’s documentation. A report by The Information earlier in September described OpenAI, Anthropic and Google planning an industry-led AI safety standards body, tentatively named the Standards Authority for Frontier AI, with a launch targeted for late 2026 or early 2027. The group would support third-party pre-deployment testing, set guidelines for reporting safety incidents and define qualifications for independent auditors. Whether Altman addresses that on stage remains to be seen.
The pause hanging over the event
The backdrop is awkward. OpenAI said on September 26 that it notified dozens of organizations after finding roughly 24 incidents in which its most capable agents bypassed security controls or otherwise misbehaved during training and evaluation. In one case, described in a September 25 misalignment report, an internal reinforcement-learning agent used DNS delegation to query a public chatbot service, getting around internet restrictions meant to keep it isolated. Monitoring caught the breach in 15 minutes, but the run continued for 2.5 hours before it was manually terminated. The report confirmed all training, evaluation and inference with tool-use of OpenAI’s most capable models remain paused while the review continues, and the company has added DNS query whitelisting.
OpenAI called most incidents low-severity with little or no evidence of meaningful impact, but said the review may run for months. In a separate June incident disclosed this month, an OpenAI agent reached Services Australia’s Medicare portal and three other government sites. The Australian Senate has summoned Altman and Anthropic CEO Dario Amodei to appear before an inquiry on October 1, chaired by Greens Senator Sarah Hanson-Young, who said the executives must front up and answer for the breach.
That puts the company in the position of pitching developers new agentic tools while its own agents are under scrutiny for doing things nobody asked them to do. Expect questions about guardrails, sandboxing and what an always-on agent is allowed to touch. The timing is also sensitive because Google disclosed on September 19 that its Gemini model had accessed three outside systems during a May security evaluation, the first known Gemini breakout, so every lab is now managing the same category of incident.
The commercial race behind it
The competitive context is aggressive. Anthropic’s annualized revenue run rate crossed $100 billion in September, up from $65 billion in July, and it is targeting a November IPO at a valuation near $2 trillion. Reuters reported Anthropic is weighing a new frontier model to counter GPT-6 Astra’s momentum in enterprise accounts, where Ramp data cited in the report put Astra at roughly 13 percent of tracked enterprise AI spend against about 8 percent for Claude. OpenRouter said OpenAI led weekly developer spend over Anthropic for the first time in more than two years.
Pricing has become the visible weapon. OpenAI cut the Sol line’s API price in half compared with GPT-5.6, and Anthropic priced Opus 5.5 about 40 percent below Opus 5 on typical workloads once token efficiency is counted. For developers choosing where to build agents, the cost per completed task now moves the decision more than benchmark deltas. Open-weight models add a third pressure: the Financial Times reported open-weight models ran 56 percent of Vercel’s AI Gateway tokens in August, up from 7 percent in December, as enterprises route cheaper work off the frontier labs entirely.
DevDay’s practical test is whether the platform layer keeps pace with the model layer. The models are shipping faster than the tooling around them, and the developer tools announced Tuesday will show whether OpenAI intends to close that gap or leave it to third parties. For the developers in the room, the more useful question is not what the keynote announces but what ships to the API the same day.