OpenAI's "o" Assistant, Claude Sonnet 5.5 and MiniMax M3.1: Three Model Bets Before DevDay

There is no quiet day in artificial intelligence. Three days before OpenAI's developer conference, a substantial leak began circulating: the company may be about to unveil a permanently running assistant. The name is a single letter: "o." That leak was only one of three stories the video treats as the week's real news; the same week brought a price-and-efficiency claim from Anthropic and a preview from Chinese lab MiniMax.

The product behind a single letter

According to TestingCatalog, "o" appeared on the upgrade screen of ChatGPT's $100 Pro subscription as a listed benefit: "o, your always-on assistant." The same team found "o" as a display name and a "-o" suffix in the application's configuration, and reads that suffix as a hint the assistant will handle email. Deriving a capability from two configuration fields is an inference, not a confirmation, and the outlet flags the distinction itself. The stronger clue is hidden in the branding. The @o account on X is currently suspended and may be held for the launch, though a one-character name can be suspended for countless reasons. wccftech.com reports OpenAI intends to position the product against Meta's Muse agent, and cnbc.com notes JPMorgan's analysts said Muse could become the most-used consumer AI application after ChatGPT, which explains the urgency on OpenAI's side.

The internal codename is the most muddled part. One account says "aeon" refers to OpenAI's existing custom-agent implementation for ChatGPT Workspace accounts, with a consumer-facing "o" built on top of it. runtimewire.com writes the opposite: that "o" would run on a GPT-6 Astra variant called aeon, tuned for long-running work. Three narratives, none from a primary source, and none of them settles the question.

When speed becomes a product feature

OpenAI's speed moves are better documented than the leak. testingcatalog.com reports a three-level speed selector spotted in the API playground: standard, fast and ultrafast. openai.com confirms the underlying shift, with developers.openai.com documenting that priority processing was renamed "Fast mode" on July 30, 2026 and delivers up to 2.5x faster responses. openai.com's Ultrafast preview, running on Cerebras infrastructure, reaches up to 750 output tokens per second. That speed carries a price tag in the subscription ladder. decrypt.co reports that ChatGPT's website code contains a $500 "Pro Max" tier that appears to buy faster processing rather than more usage, while testingcatalog.com says the tier is described as "Fastest Work and Codex" and likely runs on Cerebras. cerebras.ai confirms its side of the story, noting Ultrafast opened to a select group of customers and delivers the 750-token rate without a quality compromise. openai.com's partnership announcement supplies the scale: 750 megawatts of low-latency compute added to the platform.

Speed is no longer a feature but a new price tier. The jump from $100 to $500, as testingcatalog.com frames it, shows what users are now paying for: not weekly workload, but how much latency sensitivity they will absorb. The speaker's own inference points the same way, arguing that OpenAI has visibly stepped up infrastructure spending and that infrastructure could be a major DevDay theme.

Anthropic's price-to-performance flank

Anthropic's numbers are verifiable. In its September 22 Opus 5.5 announcement, anthropic.com says the model performs at the level of Fable 5.1, costs 40% less to run than Opus 5 and generates output more than 30% faster. anthropic.com's price table spells it out: $4 per million input tokens, $20 per million output tokens and $0.20 for cache reads. techcrunch.com notes the launch landed just two months after Opus 5 shipped on July 24.

The same anthropic.com post says Claude Sonnet 5.5 and Haiku 5.5 will follow "in the coming weeks," without a firm date. Leaks relayed by start-up-fortune claim partners reached a newer checkpoint and point to a late-September or October 1 launch window, while start-up-fortune stresses that Anthropic has committed to nothing beyond the September 22 statement. The pricing in the rumor is the solid part: anthropic.com's Sonnet 5 post locked $2 per million input and $10 per million output as permanent. The speaker's claim that the model beats GPT-6 on every axis rests on his own testing, not on Anthropic's announcement.

MiniMax M3.1 and the speed-efficiency trade

The third story carries the video's most striking timing: the speaker caught the leak in an early GitHub pull request, began assembling the video around it, and the M3.1 Flash preview was announced as the recording was nearly finished. orcarouter.ai reports that MiniMax left behind an architecture note dated September 22 and a 250 GB checkpoint on Hugging Face that only inference partners can open, with model watchers reporting that M3.1 is next for the coming week.

On output quality the speaker shows two tests. One produces pixel art, where the fast model adds the atmospheric detail more capable models often skip. The other asks for a line drawing from a short prompt, where the stronger model returns output so malformed it cannot be rendered while the fast model builds a readable, coherent scene. The speaker presents this less as a verdict than as a competitive signal, stopping well short of calling the model the best. A further fast model is floated: a free preview with a one-million-token context window, opening within two weeks in a coding client. That claim sets up the coding-agent demos that close the video, where agents turn a game into a working clone and the scale of the work moves well past small demonstrations.

Long-running agents: where the line sits

The closing project shows how far the stated promise actually goes: 100,000 lines of game code, 26,000 lines of test code, 1,000 automated tests, 775 files and roughly 25 hours of uninterrupted agent work. On top of that, hundreds of images and more than a hundred audio files were generated through code, along with an in-game guide. anthropic.com documents a comparable jump in agentic coding scale, reporting that one early tester completed a 680,000-line code migration in under a day. favtutor.com's compilation of these figures distills the idea: agentic coding is no longer a small demo but a mode of production that compresses weeks of engineering work into days. Against that, goldiebench.com points out the sector has begun building its own measurement, running one-shot builds and genuinely playtesting them rather than trusting screenshots. The industry is now auditing its own claims, which means fewer of them will survive unexamined.

The week's larger picture is infrastructure locking together with agent architecture. OpenAI is widening speed tiers, Anthropic is promising that same speed more cheaply, and MiniMax is packaging speed at a lower price still. All three point at the same thing: what is being promised to users is no longer how smart the model is, but how fast the work finishes. DevDay on September 29 will be the first time those promises can be measured.