Scale · founder · 6 min read

OpenAI DevDay 2026: what founders should actually do about Dots, Sol and the $500 plan

OpenAI's September 29 DevDay shipped always-on agents, GPT-6.1 Sol and a $500 plan. Here's what matters if you build with AI tools and what to ignore.

Published October 1, 2026 ·

openai devday codex agents pricing vibe coding

OpenAI held DevDay on September 29 and announced more than twenty things. Most of them are aimed at enterprises and developers who run their own infrastructure. A handful matter if you’re a founder or PM building with AI tools. This is the filtered version.

The short version

Three announcements deserve your attention: GPT-6.1 Sol (a cheaper model), Codex Cloud (agents that keep working when your laptop is closed), and Dots (always-on agents). Two are mostly noise for you: the Ultrafast tier and the $500 Pro plan they’re bundled with. The rest is plumbing.

GPT-6.1 Sol: the one that touches your bill

Sol is OpenAI’s “near-flagship, much cheaper” model. API pricing is $2 per million input tokens and $10 per million output, with cached reads at $0.10. It has a context window of about 1.05 million tokens and can write up to 128,000 tokens in one response. OpenAI’s pitch is near-parity with its top model, GPT-6 Astra, at a fifth of the cost, focused on coding, computer use and professional workloads.

Treat that claim as a claim. The numbers OpenAI published are from its own tests. Our earlier piece on why Astra isn’t a coding upgrade still applies: model launches are marketing events first, and the useful comparison is the one you run on your own task.

What to do: if you pay for an AI coding tool that lets you pick a model, try Sol on one real task you’ve already done with another model. If the output is as good, you’ve found savings. If you’re on a flat subscription, nothing changes on your invoice, though usage limits may stretch further. We covered how that mechanism works in the model price war.

Codex Cloud: delegating work that outlasts your laptop

Codex tasks can now run in isolated cloud environments and carry on if your computer goes to sleep or goes offline. You start something on your desktop, close the lid, and pick it up on your phone later. OpenAI also shipped a refreshed command-line tool, GitHub code review as generally available (GitLab is in preview), and a vulnerability scanner called Codex Security Cloud for paid business tiers.

This matters because the bottleneck for non-technical builders is rarely typing speed. It’s babysitting. An agent that finishes a long task while you’re in meetings is worth more than one that’s twice as fast while you watch it.

The catch is the same one we flagged in agents that don’t stop to ask: the longer an agent runs unattended, the more it can get wrong before you see it. Keep backups, keep it in a repo with history, and review what it did.

Dots: always-on agents, with real caveats

Dots are persistent agents that run on their own cloud computers using GPT-6 Astra. They connect to more than 4,000 apps and you reach them through ChatGPT, voice calls, Slack or Teams. They’re available to Pro and Business Premium subscribers in eligible markets, and in beta for Enterprise, Edu and healthcare where an admin has to switch them on. Rollout is gradual, so you may not have access yet.

For a PM, the interesting use is recurring work: a weekly competitor scan, a status report assembled from your tools, inbox triage. For a founder, it’s a cheaper first hire for admin. The risk is the one our piece on what your agent is allowed to touch describes: an always-on agent with access to 4,000 apps is an always-on agent that can make mistakes in 4,000 apps. Start with read-only access and one narrow job.

Ultrafast and the $500 plan: skip, probably

Ultrafast generates output up to 8x faster in Codex (6x in the API) at six times the standard price. The speed is real, and it’s bundled into the new $500/month Pro 500 plan, alongside a reopened $200 Pro tier. Unless your team is losing hours of paid time waiting on the agent, this is a premium you don’t need. Fast and cheap are different purchases; Sol is the cheap one.

Two smaller items worth a note

Sign in with ChatGPT lets users spend their own ChatGPT plan allowance inside partner apps, with Devin, Notion and Vercel named. If you ship an app, this could eventually shift AI costs from you to your users. It’s early; watch it rather than build on it.

Private Intelligence is OpenAI’s zero-data-retention offering, with a confidential-computing option promised for fall. If you handle customer data, read our note on AI coding tool data retention before assuming any vendor’s default is safe.

What to do this week

1. If you use Codex, check which model it defaults to and test Sol on one real task.
2. If you’re on GPT-5.5 in Codex or ChatGPT, it retires October 14. Migrate before then.
3. If you try Dots, give it one narrow, read-only job first.
4. Ignore Ultrafast unless you can name the cost of waiting.

DevDay was big on volume and modest on change for non-engineers. The signal is the direction: agents that run without you, at prices that keep falling. Build your habits around backups and review, not around any single model name.