Huge OpenAI DevDay Leak! “o” AI Agent, Sonnet 5.5 Beats GPT-6, MiniMax M3.1 Out & More! AI News

The headline before DevDay: OpenAI’s “o” agent appears in the wild

The working headline still matches the core story: “Huge OpenAI DevDay Leak! ‘o’ AI Agent, Sonnet 5.5 Beats GPT-6, MiniMax M3.1 Out & More! AI News.” The center of gravity is not a confirmed launch, but a cluster of late-September signals around OpenAI’s September 29 DevDay, starting with an apparently leaked ChatGPT Pro upgrade screen that included the phrase “o, your always-on assistant” [2].

The most concrete evidence is narrow but meaningful. A screenshot posted on September 25 showed a $100 ChatGPT Pro “Standard” plan with “o” listed among plan perks, alongside stronger models, more Work and Codex usage, maximum memory, storage and early access [2]. A separate configuration fragment attributed to TestingCatalog reportedly included “o” as a display name and “-o” as an email suffix, which has fueled the idea that the assistant may have its own task identity or mailbox-like routing [2]. That does not prove email control, but it does explain why observers are reading “o” as more than another model picker.

HelloBro’s source brief goes further, framing “O” as an always-on agent intended for long-running jobs that could persist for hours, days or even weeks, with possible support for 63 languages, paid tiers and email-style task management [1]. The careful reading is that these are still leak-layer details rather than OpenAI announcements. OpenAI has not confirmed the product name, availability, pricing, permissions or launch date in the sources reviewed here [2].

Why “always-on” changes the product category

If “o” is real, the important shift is not the letter. It is the operating model. ChatGPT today is mostly a session-based assistant: users ask, the model answers, tools may run, and the task usually ends. The leaked framing around “o” suggests a persistent assistant that could keep working in the background, resume state, manage jobs and perhaps operate as a named agent rather than a passive conversation window [1].

That puts OpenAI’s rumored product in the same strategic lane as other always-on assistant efforts: systems that log into services, monitor messages, run code, follow instructions over time and report back only when needed. The article that examined the leak explicitly warns that “email_suffix” alone does not prove email handling; it could be an alias, address suffix or internal routing label [2]. Still, the existence of such a field is exactly the kind of product plumbing one would expect if OpenAI were preparing account-linked agent identities.

The DevDay timing matters because developers do not only need a polished consumer assistant. They need APIs, permission models, background execution, audit trails, sandboxing, storage, billing and rate-limit policies. HelloBro’s brief also notes spotted “standard,” “fast” and “ultra fast” modes in the Responses API playground, suggesting that OpenAI may be positioning speed itself as a premium developer feature [1].

The 63-language and paid-tier claims need caution

The leak cycle has already produced stronger claims than the evidence can carry. One report says an aggregator treated Tuesday’s DevDay reveal as a given and repeated claims that “o” would support 63 languages and run with a Cerebras-powered “Fast Mode,” but the same analysis says those figures were not traceable to a primary source [2]. HelloBro’s version records 63-language support and multiple paid tiers as internal clues, but still places them inside a leak context rather than a verified product card [1].

That distinction is important for builders. If “o” appears at DevDay, the first questions will be practical: Can it read and send messages? Can it run unattended? Can it use Codex, Work, browsing and files? Can it be interrupted? Does it expose logs? Can enterprises restrict tools? And how is liability handled when an agent works for days instead of seconds? Until OpenAI publishes documentation, the safest conclusion is that “o” looks plausible, not proven.

Anthropic pressure: the “Sonnet 5.5 beats GPT-6” claim is partly shorthand

The headline’s Anthropic angle is also a mix of signal and shorthand. HelloBro reports that Anthropic has confirmed Sonnet 5.5 is due within weeks and that partner testing of a newer checkpoint has produced early descriptions of a fast, efficient and more natural model [1]. The same brief says a Sonnet 5.5 at or near current Sonnet pricing could become a major pressure point for OpenAI in coding and general-purpose work [1].

The strongest public numbers circulating in the 72-hour window, however, are about Claude Opus 5.5, not Sonnet 5.5. The ThursdAI episode published September 25 says Opus 5.5 beat GPT-6 Astra on Terminal-Bench 4.0, with 66.4% versus 57.9%, and describes Opus 5.5 as faster and cheaper than the prior Opus tier on typical workloads [5]. The same source says Sonnet and Haiku 5.5 are expected in the next few weeks, which supports the idea of Anthropic’s 5.5 generation arriving in waves rather than all at once [5].

So the clean editorial version is this: Anthropic’s 5.5 family is pressuring OpenAI now, but “Sonnet 5.5 beats GPT-6” should be treated as a leak-and-test-community claim until Anthropic publishes Sonnet 5.5 model cards, pricing and benchmark tables. The confirmed comparison in the freshest public source is Opus 5.5 versus GPT-6 Astra on a named coding benchmark [5].

MiniMax M3.1: out, leaked, or still behind the curtain?

MiniMax is the messiest part of the story because the sources disagree in a useful way. HelloBro states that what first appeared in an early GitHub commit has become an official “MiniMax M3.1 Flash Preview,” positioned as a text model for everyday software development with speed, reliability and support for bug fixes through full feature building [1]. That framing fits the broader theme: fast models are closing the quality gap, especially for coding workflows [1].

OrcaRouter’s September 26 analysis is more conservative. It says no public MiniMax M3.1 exists as of September 27: no model card, no blog post, no public weights, no pricing page and no API model ID [3]. It does, however, document a private checkpoint named MiniMax-M3.1-preview-private, reportedly around 250 GB with 62 files and 48 safetensors files, plus a second private drop and a leaked architecture note reproduced in a public partner repository [3].

That means the safest synthesis is not “nothing happened” and not “fully launched.” The current public record supports “M3.1 is in the preview/leak/partner-integration stage, with some claims of Flash Preview availability, but MiniMax has not published the normal artifacts of a public release.” Developers should watch for the basics: a public Hugging Face model card, a MiniMax blog post, an API model ID, pricing and independent benchmarks [3].

Space Bunny and the stealth-model pattern

The MiniMax thread also overlaps with Space Bunny Alpha, an anonymous model listing that drew attention for long context, speed and a MiniMax-compatible tokenizer signature. Space Bunny Alpha’s own benchmark page reports OpenRouter snapshots around September 24 with 87 tokens per second median writing speed, 1.07 seconds median response latency and 94.98% three-day inference availability [4]. It also says token-count probes showed a MiniMax-family resemblance but did not establish the exact checkpoint [4].

That matters because stealth releases have become an industry habit. Labs and routing platforms increasingly test models under anonymous names, then reveal the vendor later. In this case, the evidence supports “possible MiniMax family,” not “definitive M3.1.” Combined with OrcaRouter’s caution, Space Bunny looks like another clue in the M3.1 watch cycle rather than proof of a public launch [3][4].

The bigger picture: agents, latency and price-performance

Taken together, the OpenAI, Anthropic and MiniMax threads point to the same competitive axis. The frontier is no longer only “which model is smartest in chat.” It is becoming: which system can keep working, which one is fast enough to feel interactive, which one is cheap enough for agent loops, and which one can survive long-running tasks without losing control.

OpenAI’s rumored “o” would push the race toward persistent agency [1]. Anthropic’s Opus 5.5 numbers, and the expected Sonnet 5.5 follow-up, push on coding quality and cost [5]. MiniMax’s M3.1 trail, whether leak or preview, pushes on efficient long-context and developer-oriented deployment [3]. DevDay may therefore be less about a single surprise and more about whether OpenAI can define the next interface for AI work: not a chatbot you prompt, but an agent you assign.

Sources from the last 72 hours

1. [1]Huge OpenAI DevDay Leak! “o” AI Agent, Sonnet 5.5 Beats GPT-6, MiniMax M3.1 Out & More! AI NewsSep 27, 2026, 6:17 AM UTC
2. [2]OpenAI leak points to "o", an always-on assistant inside ChatGPT ProSep 26, 2026, 12:00 AM UTC
3. [3]MiniMax M3.1 Leak: What the Preview Documents RevealSep 26, 2026, 12:00 AM UTC
4. [4]space-bunny-alpha Benchmarks: HLE, GPQA, MMLU-Pro & AI BENCHYSep 25, 2026, 12:00 AM UTC
5. [5]ThursdAI - Sep 24 - Opus 5.5 beats Fable for 40% less, OpenAI halves GPT-6 prices & moreSep 25, 2026, 5:48 AM UTC

AI-generated article based on recent web research, then preserved as a dated editorial snapshot.