OpenAI Dev Day Disaster Report: Usage Cut in Half, Demo Fails, and a New $500 Tier

OpenAI's Dev Day was supposed to be a showcase of cutting-edge innovation — instead, it turned into a controversy-laden launch event. From a quiet halving of plan usage to an awkward on-stage demo freeze, to a lineup of products that felt suspiciously familiar, this event generated far more criticism than applause. Based on a deep-dive breakdown from a YouTube creator, here's a look at the key moments from this event — both the genuinely impressive and the eyebrow-raising.

The Night Before: A "Warning Shot" on Usage

The day before the event, OpenAI posted a tweet as a kind of preemptive disclaimer, announcing changes to the $200/month subscription tier. The previously popular 20x usage plan was being cut to 10x, with the official justification being "twice the model efficiency."

The problem is that this "twice the efficiency" claim has no verifiable benchmark data to back it up — users simply have to take OpenAI's word for it. The deeper controversy is this: users had already paid for a plan at a specific price and usage level, and the provider unilaterally changed what that plan included. Many see this as one of the more blatant "rule changes" of the AI era, striking at the most sensitive trust issue in subscription products. You bought a specific price-for-usage combination, and changing one side of that equation mid-subscription will naturally trigger a strong reaction.

The New $500 Tier and "Ultra-Fast Mode"

To appease heavy users, OpenAI introduced a new $500/month plan offering 25x usage. Do the math: compared to the previous month's plan, you're paying 150% more for roughly 25% more tokens — a deal that left many developers shaking their heads.

The new tier also comes with a feature called "ultra-fast mode." It's 8x faster than the standard fast mode but costs 6x more. The creator tested it firsthand and found the speed genuinely addictive — but equally capable of burning through credits at a frightening pace.

In one test, the model was tasked with a V2 migration, completing 722 line insertions and 16 deletions in about 1 minute 41 seconds. The cost: weekly usage dropped from 54% to 52% — a single operation consuming 2% of the weekly quota. At that burn rate, the $500 plan's weekly allowance could be exhausted within 40 to 100 minutes of actual agent runtime. Some users have reported hitting their weekly cap in as little as 40 minutes.

The "unlimited usage" framing is also misleading — it's actually "unlimited usage for one month," and when an agent needs to delegate tasks to other agents, additional charges still apply. These fine-print details put a significant asterisk on the word "unlimited."

A Stumbling Live Demo and a Naming Debacle

The always-on agent showcased at the event (codenamed Dottie) was supposed to be the star of the show: a persistent agent capable of conversation, maintaining organization, and operating a computer. But during the live demo, when asked to "summarize last night's user testing," it repeatedly responded with "still looking" and never produced results — an excruciatingly awkward moment.

Wi-Fi was later blamed for the failure, but video footage clearly showed the agent receiving and responding to messages — it just couldn't complete the task. That made the Wi-Fi excuse hard to swallow.

Things got more dramatic with a domain name incident. According to the creator, SpaceX swooped in and registered the relevant domain, causing it to redirect to competitor Grokbot instead of OpenAI — a move reportedly costing SpaceX around $20 million. To top it off, OpenAI CFO Sarah Friar accidentally referred to the product by a competitor's name during her introduction, which inadvertently highlighted just how similar the two products are.

Always-on agents are one of the hottest battlegrounds in AI product development right now. The core design philosophy is to have an AI agent running continuously in the background — aware of the user's calendar, files, and communications — ready to respond immediately rather than starting from scratch each conversation. What sets this apart from traditional chatbots is "persistent context": the agent remembers past events and proactively tracks task progress. Grokbot (an xAI product), Cursor (AI agent within a code editor), and others have explored similar capabilities in their respective verticals. OpenAI's Dottie targets more general-purpose use cases and supports computer use (i.e., directly controlling desktop applications), which raises the technical bar significantly compared to pure conversational agents. This also partly explains why the demo was prone to instability — multi-step, cross-application task chains demand extreme system reliability, and any single step timing out can cause the entire demo to freeze.

A Lineup Full of Déjà Vu

Most of the headline announcements at this Dev Day had a strong "we've seen this before" flavor:

• Always-on agent: Widely compared to Grokbot, with a UI reminiscent of Cursor.
• Decision API: A general-purpose classifier API with image support. The catch? It was announced at the event but wouldn't be available for roughly another week — what the creator mockingly called "an announcement about an announcement." Cloudflare and others already offer similar classifier products.
• Spaces: A Notion or Google Docs-style product with integrated to-do lists and notes, described as "yet another note-taking app."

The creator's core critique: a company that sells itself on "AI is the future and can create entirely new experiences" spent most of its launch event showcasing product categories that already exist. This "reinventing existing products" strategy doesn't do a compelling job of demonstrating AI's innovative potential.

Two Genuinely Noteworthy Highlights

Despite the criticism, the creator did acknowledge two real bright spots from the event.

Sign in with ChatGPT: Users can log into third-party apps using their ChatGPT account and share a portion of their token quota with those apps. This means no more manually copying and pasting API keys when using applications, and it may support spending caps — so even if an app is compromised, it can't drain your entire account. The creator views this as a genuinely forward-thinking design, predicting that within the next year or two, we'll see a wave of "bring your own token" apps: applications that provide value on top of your own tokens while charging a service fee.

Open model access in Codex: GLM 5.3 Flash and Kimi K3 will be natively integrated into Codex, with their usage counting toward a user's OpenAI commit quota. OpenAI calls this "embracing openness." The creator is cautiously skeptical, however: a closed-source frontier AI company loudly proclaiming "open is the way" may ultimately be motivated by a desire to funnel open model usage into its own closed ecosystem. The true "openness" here deserves scrutiny.

Final Thoughts

Taken as a whole, this Dev Day exposed several contradictions in OpenAI's product strategy and user communication: adjusting plan terms in ways that erode trust, padding the launch event with products that feel derivative, while simultaneously showing genuine forward-thinking vision in identity authentication and open model integration. When AI can reach this level of intelligence and write code this fast, what likely follows is a tsunami of low-quality code (slop) — and that may be a future we need to start thinking about now.

Background Notes

"Sign in with ChatGPT" is essentially an OAuth-like identity authentication and authorization protocol extension. OAuth (Open Authorization) is the dominant third-party login standard on the web today — it's what powers "Sign in with Google" or "Sign in with GitHub." OpenAI has added an AI-specific layer on top: when users authorize, they can choose to share a portion of their account's token quota with a third-party app, with the ability to set a spending cap. The value of this design is that it makes "AI capability" — not just identity information — a shareable, authorizable asset. For developers, there's no need to build and maintain API key management and billing systems. For users, token usage rights are precisely carved up so that even if a third-party app is attacked or abused, the damage is capped at a preset limit rather than exposing the entire account. If widely adopted, this model could give rise to a new class of lightweight AI applications — apps that provide vertical-specific interaction logic while relying entirely on the user's own OpenAI account for compute and model capability, letting developers focus on product value rather than infrastructure.

Codex here refers to OpenAI's cloud-based AI coding agent platform — distinct from the earlier code-completion model of the same name. This generation of Codex can autonomously execute multi-step programming tasks in a sandbox environment: reading code repositories, running tests, committing changes, and so on — functioning much like an asynchronously operating "virtual programmer." Integrating GLM 5.3 Flash (Zhipu AI's lightweight model) and Kimi K3 (Moonshot AI's reasoning model) into the Codex ecosystem means that calls made to these third-party open or semi-open models will be counted toward the user's OpenAI platform quota. The benefit for OpenAI is obvious: it converts third-party model usage into traffic and billing flowing through its own platform, using an "openness" posture to reinforce its position as the central hub. For users, unified quota management is genuinely more convenient in the short term — but whether this creates deep dependency on a single platform in the long run is worth watching.