OpenAI DevDay 2026: Dots, Ultrafast, Sol, and the Decisions API

1.Dots: OpenAI's Always-On Personal Assistant0:17

The biggest announcement, in Matthew Berman's telling, is Dots — OpenAI's entry into the personal-assistant race against Grokbot, Muse and Instinct, a whole generation of assistants he credits to OpenClaw (Peter Steinberger) for setting the tone. Dots is "always on" — working 24/7, proactive, trying to figure out what you need before you ask — rather than a prompt-and-response chat.

The mechanics are notable: Dots lives inside the ChatGPT app and controls your ChatGPT and Codex threads, with its own cloud environment and connections to Gmail, Docs, Calendar, Slack and Teams. Berman is ambivalent about this — "I kind of wish Dots was just its own app, its own thing" — but understands the reasoning: to get computer control, coding, and its own environment, they built it on top of Codex and ChatGPT. One caveat worth knowing: talking to Dots doesn't count against your ChatGPT usage, but the moment it drives your ChatGPT or Codex threads, that does.

2.Ultrafast: 8× Faster on Cerebras, 6× the Price3:06

The announcement Berman says he's personally most excited about is Ultrafast — GPT-6 Astra running on Cerebras chips instead of Nvidia GPUs, at 8× the speed. The trade-off is blunt: you pay 6× more per token, and those tokens "burn quickly." He had early access and burned through $1,000 in an hour and a half.

The demo makes the case: standard Astra and Ultrafast are both asked to build and launch a rocket; while standard is still thinking, Ultrafast's rocket is already launching. The deeper observation comes a moment later — the inference is now so fast that the bottleneck is your local computer, not the model. Tool calls and terminal commands are now the slow part.

3.Pro 500 and the Usage-Limit Shuffle3:41

Alongside Ultrafast comes a new Pro 500 plan at $500/month. Berman pre-empts the obvious objection — "everyone says AI is getting cheaper, but people are paying more" — with the distinction that matters: the same intelligence that costs pennies per million tokens today cost $10–30 a year ago. Intelligence is getting cheaper; the absolute frontier (and speed) is what you pay a premium for.

But he's critical of the limit shuffle: the old $200 Plus plan's 20× multiplier drops to 10×, while the new $500 tier gets 25×. "You're barely getting more than what you used to get at the $200 mark, but now you're paying a lot more."

4.GPT-6.1 Sol: Frontier Quality at a Fraction of the Price5:03

The release he calls "possibly the most impressive" is GPT-6.1 Sol — a model that's "basically as good as Astra, and much cheaper and faster." The pattern mirrors Anthropic's Opus 5.5 → Sonnet 5.5 move: both labs seem to have "solved" post-training, taking a massive frontier model (Fable, Astra) and distilling it into a smaller model that's nearly as capable.

"That is so cheap. It is so, so cheap. And the model's really good." The takeaway he draws is structural: when the flagship's post-training distills cleanly into a cheaper sibling, we all win — frontier quality spreads down the price curve.

5.The Benchmarks6:08

On the benchmark Berman considers the best reflection of how developers actually feel about these models, 6.1 Sol sits at the same level as — if not slightly higher than — Astra, at a fraction of the price. The same story repeats on a PDF benchmark ("basically performs like Astra") and on OS World (computer use), where Sol performs "very well, especially at the highest thinking tiers" — though Astra still edges it out at the very top end.

6.Codex Goes Cloud-Native, Plus a Security Cloud7:21

Two Codex announcements. First, Codex Security Cloud — AI that continuously scans your code for issues and vulnerabilities and proactively flags them. Second, and the one Berman is "very excited about": Codex going to the cloud natively. "Codex belongs in the cloud. You should not have to have your laptop open at all times."

His reasoning ties back to Ultrafast: with inference now lightning-fast, the biggest bottleneck is the local machine — so moving Codex into a cloud environment removes exactly the thing that's now slowing everything down.

7.The Dev Surface: Voice CLI, Code Review, and the Decisions API8:10

A refreshed Codex CLI adds voice control, and Codex gets a new code-review experience. But the announcement with the most signal for this site is the Decisions API — "watch out Jev," Berman says. It's a model built to make quick decisions, "based on Luna's intelligence," OpenAI's lightest, smallest model.

8.Plugins, Sign-in with ChatGPT, and ChatGPT Space8:48

Three platform moves close it out. Plugins relaunch: two DevDays ago OpenAI declared apps dead and scared a lot of startups, and "that entire experience did not work very well" — now apps run natively inside ChatGPT. Sign in with ChatGPT + bring your tokens: log into third-party apps with your OpenAI account and carry your subscription tokens with you, instead of paying the third party — "generous of them, doesn't seem like something Anthropic would do."

Finally, ChatGPT Space — "watch out Notion." A native workspace for you, your teammates, and agents to collaborate in: PowerPoint, spreadsheets, websites, all in one place.