New data point: one early tester burned through $1,000 of AI tokens in about an hour and a half. That’s the price of running OpenAI’s new “Ultrafast” mode flat out. It comes from Matthew Berman, the creator who attended OpenAI Dev Day in person and covered every announcement in one quick recap video.
I went in expecting a model bump and a few API updates. There was a lot more than that. OpenAI showed an always-on personal assistant, a much faster premium mode, a cheap model that nearly matches their flagship, and a big push to move coding into the cloud. Here’s what matters and what you can do with it.
What OpenAI actually announced
Dots, an assistant that works while you don’t. Dots is OpenAI’s answer to the new wave of personal AI assistants like Grokbot and Muse. The expert credits OpenClaw and Peter Steinberger for setting the direction for this whole category. The big shift is that Dots is proactive. You don’t open ChatGPT and type a prompt. It runs 24/7 in its own cloud environment, connects to your Gmail, Docs and Calendar, and tries to handle tasks before you ask.
You can reach it through ChatGPT, Slack or Teams. Under the hood it controls your ChatGPT and Codex threads, which the author isn’t a fan of. He’d rather it were a standalone app. There’s also a billing catch. Chatting with Dots is free, but once Dots starts running ChatGPT or Codex threads for you, that counts toward your usage.
Ultrafast mode. This runs GPT-6 Astra on Cerebras chips instead of Nvidia GPUs. It’s about 8x faster and costs about 6x more. In the side-by-side demo, the standard model was still thinking while the Ultrafast version had already built and launched its rocket.
The new Pro 500 plan. It’s $500 a month for 25x the usage of Plus. Here’s the part that stings: the $200 plan dropped from 20x to 10x. So if you want roughly what you had before, you now pay more.
GPT-6.1 Sol. The creator called this possibly the most impressive release of the day. The numbers:
- GPT-6 Astra: $10 per million input tokens, $50 per million output
- GPT-6.1 Sol: $2 per million input, $10 per million output
- Sol cached input: 10 cents per million tokens
On the developer-sentiment benchmark the author trusts most, Sol matches or beats Astra. It holds up on a document benchmark too. On OS World (computer use), Astra still wins at the very top thinking tiers, but Sol gets close for a fraction of the price. The expert sees a pattern here. Anthropic did the same thing with Sonnet 5.5 sitting close to Opus. Both labs seem to have figured out how to shrink their giant models into cheaper ones that are almost as good.
The rest of the list:
- Codex Security Cloud: keeps scanning your code for vulnerabilities and flags issues as they show up
- Codex in the cloud: coding runs remotely, so your laptop doesn’t have to stay open
- Refreshed Codex CLI: now with voice control, plus a new code review experience
- Decisions API (preview): a light, fast model built for quick yes/no style calls, based on OpenAI’s smallest model
- Plugins relaunch: apps inside ChatGPT, take two, after the first attempt flopped
- Sign in with ChatGPT + bring your tokens: log into third-party apps with your ChatGPT account and use your existing subscription there instead of paying twice
- ChatGPT Space: a Notion-style shared workspace for teams and agents, covering slides, spreadsheets and websites
The insight underneath it all
The author makes a sharp point about pricing. People see a $500 plan and assume AI is getting more expensive. It isn’t. Intelligence that cost $10 to $30 per million tokens a year ago now costs pennies. What costs more is the very top: maximum speed and the best possible quality.
His second point surprised me. With Ultrafast turned on, the model isn’t the slow part anymore. Your computer is. Inference finishes almost instantly, then everything waits on tool calls and terminal commands. That’s exactly why moving Codex to the cloud matters.
3 practical ways to use this
🚀 Move routine API work to Sol. If you run content pipelines, summaries or agent workflows on a flagship model, test Sol first. At a fifth of Astra’s price, with 10-cent cached input, repetitive jobs that reuse the same system prompt get very cheap.
🤖 Hand Dots your busywork. Connect your calendar and inbox and let it prep meeting notes, chase follow-ups or draft replies overnight. The creator thinks simple, useful assistants like this will win over people who are still skeptical about AI, and I agree.
💰 Save Ultrafast for when speed pays. Live demos, fast prototyping and tight deadlines are the right use cases. Leaving it on for a whole workday is how you end up with a four-figure bill before lunch.
Tips and pitfalls
Tips:
- Benchmark Sol on your own tasks before switching. Public benchmarks are a starting point, not proof.
- Try “Sign in with ChatGPT” in tools you already use. You might be able to cancel a separate AI add-on subscription.
- Move long-running coding tasks to cloud Codex so they keep going after you close your laptop.
Pitfalls:
- Hidden usage from Dots. The conversations are free. The threads it runs for you aren’t.
- Ultrafast spend. Set budget alerts before you turn it on. The author’s $1,000 in 90 minutes is a real warning.
- Plan math. Check whether the $200 tier still covers you now that it’s down to 10x. Heavy users might get pushed to Pro 500 without noticing.
- Computer-use tasks. For the hardest agentic work, Astra still has the edge, so don’t assume Sol handles everything.
Worth the watch
This recap covers a whole Dev Day in about ten minutes, with demos and benchmark charts I can’t fit here. Watch the full video to see the Ultrafast rocket race and the Sol benchmarks for yourself before you change your stack or your plan.