OpenAI DevDay 2026: what’s confirmed and what’s still rumor

OpenAI DevDay will take place on 29 September 2026. Two days before the conference, rumors have already surfaced in discussions about a new “o” agent, a Pro Max plan costing $500, and upcoming model releases. As of 27 September, some of this can be checked against company pages, while the rest remains hearsay based on leaks and personal tests.

The distinction matters: an internal name in an interface, a tester’s claim, and a public model with documentation represent different levels of evidence. Below, confirmed releases are separated from assumptions.

What is officially known about DevDay

OpenAI has scheduled DevDay for 29 September in San Francisco. The opening keynote is scheduled for 10:00 Pacific Time and will be streamed. The public page lists technical sessions on APIs and tools, demos, and hands-on sessions, but does not name any specific models or a product called “o” (OpenAI DevDay 2026 schedule).

So the event date is confirmed, but the lineup of announcements is not yet. Speculation about an agent, email features, and weeks-long operation does not become official specifications simply because similar strings were spotted in an interface or configuration.

Rumors about the “o” agent and the $500 plan

In retellings of the leak, the “o” product is linked to the internal name AON and described as an always-on assistant. It is said to have a cloud environment, code execution, email capabilities, and several interacting agents. OpenAI has not yet described any of these capabilities in public documentation. Until an announcement, these are theories about what may be behind the labels that were spotted.

The same caution is needed with the $500 Pro Max plan. OpenAI’s help center describes Pro plans at $100 and $200; as of 10 September, new purchases and upgrades to Pro $200 have been temporarily suspended. There is no separate $500 Pro Max plan in that help article (current ChatGPT Pro pricing help article). The leaked price should not be included in a budget until an official page with the terms is available.

Fast mode already exists, but it is not a new DevDay announcement

OpenAI does in fact offer a faster processing mode for API requests. On 30 July 2026, the company renamed Priority processing to Fast mode and said it can process gpt-5.6-sol requests up to 2,5 times faster than standard processing. The mode is billed separately under API pricing; the list of compatible models and the price depend on the documentation (Fast mode guide).

The official announcement of speeds of up to 750 tokens per second concerned the launch of GPT-5.6 Sol on Cerebras infrastructure. This is a separate model-delivery scenario, not confirmation that every API request will run at that speed (OpenAI announcement about GPT-5.6 Sol on Cerebras). So Fast mode is an already available capability, while a new three-mode switch and its possible connection to Pro Max remain speculation.

Opus 5.5 has already been released; Sonnet 5.5 is still awaited

Anthropic officially introduced Claude Opus 5.5 on 22 September. The company named Sonnet 5.5 and Haiku 5.5 as the next models to be released in the coming weeks. As of 27 September, no public results for Sonnet 5.5 are available yet, so claims that it outperforms GPT-6 cannot be verified as a public comparison.

Opus 5.5 already has published test results. For Terminal-Bench 4.0, Anthropic reports 66,4% for Opus 5.5 and 57,9% for GPT-6 Astra, but the first result was obtained at the xhigh effort level and the second at high; the documentation also includes a margin of error. This is the result of one benchmark for Opus, not Sonnet, and not proof that one model is better at every task. Anthropic itself warns that differences in scores do not always predict differences in real-world performance. The company also reports that Opus 5.5 generates responses more than 30% faster than Opus 5, while typical workloads cost approximately 40% less according to its internal estimates (Claude Opus 5.5 announcement and comparison methodology).

A comparison involving a single Minecraft clone or one SVG image may suggest what to test next. It does not replace a benchmark: the result depends on the task, model version, parameters, available tools, and who evaluated the finished work.

MiniMax and LongWriter: do not turn claimed properties into specifications

Reports about MiniMax M3.1 Flash mention an early reference in GitHub and access to a preview version through MiniMax Code. Integration requires a stable model identifier and provider documentation. Without them, availability, speed, limits, and whether the model will remain available after the trial period cannot be reliably verified. Image-generation examples show individual successful outputs, but they do not compare models for overall reliability.

The same approach should be taken with LongWriter 2.5: claims of a million-token context, two weeks of free access, and zero data retention require a link to the model card and current service terms. Until those claims are confirmed, do not treat such information as either a technical specification or a guarantee about data processing.

A large agent demo project is experience, not an independent test

In one demonstrated run, Opus 5.5 is credited with creating a Minecraft clone in about 25 hours: 775 files, around 100 thousand lines of game code, 26 thousand lines of tests, 1000 automated tests, 783 images, and 184 sounds. The scale is impressive, and it shows how far a long-running task involving code and asset generation can go.

But counters alone are not enough to assess quality. You need the repository and commits, a list of errors and fixes, test-run results, tokens consumed and cost, as well as the amount of manual work required after the agent. Until then, the figures describe a specific demonstration, not how many weeks a developer will save on a working project.

How to verify the next announcement

1. Find the provider’s release page or API documentation, and check the date, exact name, and model identifier.
2. Distinguish an available product from trial access, an internal name, and a prediction about future features.
3. Compare models using the same tasks and tools. Record versions, settings, repeated runs, and evaluation criteria.
4. For agent tasks, consider not only autonomous operating time but also cost, the number of human corrections, test success rates, and the quality of the final project.
5. Check data-retention terms separately from the marketing phrase “zero retention”: they depend on the provider and the specific access method.

DevDay may confirm some rumors, change them, or introduce entirely different products. Until the company publishes specifications and terms, it is more useful to prepare test tasks than to plan work around leaked names.