▶ YouTube Summaries

• YT :: https://www.youtube.com/watch?v=p0fybvFyOlM
• Original title :: A total disaster

A sardonic recap of OpenAI's 2026 Dev Day: the $200 Pro plan's usage was halved the day before the event, a $500 tier with "ultrafast" mode was introduced, and most headline launches (Dots, the decision API, Spaces) are clones of existing products. Two announcements get real praise: "Sign in with ChatGPT" token sharing and open-weight models served inside Codex.

The usage rug pull

The day before Dev Day, OpenAI (via Thibault Sottiaux) announced the $200/month plan drops from 20x to 10x usage, justified by the claim that new models are "twice as efficient" - with no published evidence. Prime calls changing usage mid-subscription the biggest rug pull of the AI era so far.

For heavy users there is a new $500/month plan at 25x usage: 25% more tokens than the old $200 plan for 150% more money.

Ultrafast mode

The $500 plan includes "ultrafast" mode (built on the Cerebras partnership): roughly 8x faster and 6x more expensive than normal. Prime tested it himself: one feature implementation from a migration doc (no tests, no review agent) took 1m41s, produced 722 insertions, and burned 2% of his weekly limit. That puts a whole week's $500 allowance at roughly 40-100 minutes of agent runtime. He admits the speed is addictive.

Dots: an always-on agent

The "new thing" is Dots, an always-on, organized agent with computer use. Prime sees it as a direct copy of Muse / Grokbot, with a Cursor-like layout. Lowlights:

• The CFO accidentally called it "Muse" on stage.
• The live demo stalled ("still checking"), then got blamed on Wi-Fi although the agent visibly responded.
• SpaceX bought the dots.com domain (reportedly for $20M), which redirects to Grokbot.
• "Unlimited usage" for Pro turned out to be one month only, and any delegation to other agents still costs money.

Decision API and Spaces

• Decision API - OpenAI's take on fast general-purpose classifiers (like Jev, or Cloudflare's Cleft). Its unique feature, image input, isn't available for another week - "an announcement for the announcement".
• Spaces - a Notion-like workspace with to-dos and notes. Prime's point: an AI company releasing clones of existing products undercuts its own pitch that AI enables new experiences.

What was actually good

• Sign in with ChatGPT lets apps spend a share of the user's tokens - no API key copying, possibly with spending limits. Prime expects "bring your own token" to be everywhere within a year.
• Open models in Codex - GLM 5.3 Flash and Kimi K3 served natively in Codex, billed against OpenAI commit. A welcome contrast to Anthropic's stance on Chinese models, but Prime suspects the angle is a closed company becoming the gatekeeper for open models.