OpenAI adds Codex cloud environments, a Decisions API and an Ultrafast tier

At DevDay, OpenAI gave Codex reusable cloud environments and repository security scans, and added computer use to its Agents API, a Luna based Decisions API in preview, and Ultrafast, a speed tier billed at six times standard rates.
What it means for founders
- Speed is now a line item. Ultrafast turns latency into something you buy per token at a steep premium, so reserve it for user facing paths where waiting loses customers.
- Routing may get cheaper. If Decisions API pricing tracks Luna's $0.10 input and $0.50 output per million, teams using a large model to classify or route could move that step to a smaller, faster call.
- The platform keeps absorbing tools. Hosted browsing, security scanning and code review inside Codex overlap with what many startups sell, and EU customers should note Ultrafast has no EU residency.
- Watch for the Decisions API general release and its price sheet in the coming days, which will show whether it undercuts building your own classifier.
The story
OpenAI used DevDay on September 29 to push Codex off the developer's own machine and to add three pieces to its API: computer use for hosted agents, a narrow Decisions API, and a premium speed tier called Ultrafast. The changes sit beside the GPT-6.1 Sol, Dots and ChatGPT platform news covered separately, and are detailed in OpenAI's event recap.
Codex keeps its setup between tasks
Until now each Codex cloud task began in a fresh sandbox. OpenAI says teams can now save a reusable environment, with settings and permissions an admin has approved, so jobs boot faster and every engineer works from the same base. Work can be steered from a laptop, phone or browser. Every paid tier from Plus upward gets it, including Edu and Healthcare.
The command line tool gains voice control, an /agents view for juggling parallel jobs, and better worktree and session handling. The ChatGPT desktop app picks up a review screen where Codex summarizes diffs and flags likely problems before a human comments on a GitHub or GitLab change, and it can run a first review unattended in the cloud.
Codex Security Cloud scans GitHub repositories on a schedule or as commits land, triages results, drops duplicates and drafts patches without a laptop open. It bundles the Daybreak Blue defensive models without a separate application, and is limited to Pro, Business, Enterprise and Edu.
Three API additions
The Decisions API points GPT-6 Luna, the cheapest current model, at questions with a fixed menu of answers. Feed it text or images and it returns one of those answers, useful for sorting inputs, sending requests to the right handler or choosing what an agent does next. It opened as a limited preview, with wider release due within days.
The Agents API, in public beta, can now drive a browser that OpenAI hosts, with the developer's app approving sites and handling sign-ins. It also inherits multi-agent delegation, tool search and context compaction from Codex. Through Amazon's Bedrock Managed Agents, the same agents can run entirely inside AWS.
For Ultrafast, OpenAI claims up to six times faster output in the API and up to eight times in Codex, where it quotes 300 tokens per second. API calls on this tier cost six times the normal rate; for GPT-6 Astra, The Decoder puts that at $60 input and $300 output per million tokens. It is enabled with a service tier flag, supports only US data residency, and in ChatGPT and Codex needs Enterprise or a new $500 a month Pro 500 plan.
What we don't know yet
OpenAI has not published Decisions API pricing or latency, usage caps for reusable environments, or Ultrafast rate limits. A Sol version of Ultrafast is only promised soon.
Sources
Primary sources
Reporting
More in Products & Launches
- Meta open sources firmware and an SDK so anyone can build Muse gadgets

For founders: Hardware startups get a free agent with Meta's distribution behind it to build on.
TechCrunch · 46d ago - Suno launches Speech, generating voiceovers and background music together

For founders: If you produce podcasts, app onboarding audio, meditation content or ads, Speech is a cheap way to prototype a narrated track with a score in one pass.
The Verge · 47d ago - OpenAI launches Dots, always-on agents for ChatGPT Pro and business plans

For founders: Budget for usage, not seats. The chat is included, but anything a dot builds in Codex draws on the same limits your team already spends, so a busy dot can…
TechCrunch · 50d ago - OpenAI turns ChatGPT into an app platform, with no revenue share yet

For founders: Distribution is real, monetization is not. A spot in the sidebar of a product with a billion weekly users is a serious channel, but you still need your own…
TechCrunch · 49d ago - Google will turn Gemini Gems into skills starting November 17

For founders: Audit anything built on Gems now. If your team or customers rely on a Gem that leans on Canvas, Deep Research or GitHub files, it may lose that behavior in…
TechCrunch · 51d ago