OpenAI DevDay 2026 live blog

What it claims

This is a first-hand log of OpenAI's DevDay launches. Its value over the summary is in the operational details it records:

• GPT-6.1 Sol is pitched as "near-Astra level intelligence at a fifth of the price" of GPT-6 Astra.
• Ultrafast is up to 300 tokens/second, which OpenAI calls 8x faster. It costs 6x the standard price. Today it is available only for Astra 6; Sol 6.1 is "soon".
• Pro 500 is a new plan that includes Ultrafast and 25x Plus usage. The $200/month plan is back on sale for new subscribers.
• Decisions API is a preview. It gets sub-second responses by having the Luna model pick from "a predefined set of options".
• dots run on Astra. They are for ChatGPT Pro and Enterprise only. At launch, creating one told the author to switch to a desktop.
• Codex Security results from OpenAI's internal sprint: 36% of findings were duplicates, and auto-generated patches had a 1% rollback rate.

How it works

Codex Security. The open-source CLI and the Cloud product follow the pipeline below. The Cloud product uses Daybreak, a security-specialised model, and is distributed through the Codex plugins marketplace.

1. The tool builds a knowledge base of architecture details that are not visible in the source code.
2. It reads SECURITY.md files, where product owners define their own threat model and secure-coding guidelines.
3. The scanner loops until it finds no more issues.
4. Deduplication is a separate CLI workflow.
5. codex-security patch generates fixes and opens PRs.
6. A verify-fix command acts as an adversarial agent against each fix. The speakers credit it for the 1% rollback rate.
7. Auto-patching was presented as a "dial you turn up" as trust in result quality grows.

Cloud adds scheduled scans and automatic de-duplication.

flowchart LR
  K[Knowledge base + SECURITY.md] --> S[Scan]
  S -->|issues found| S
  S --> D[Deduplicate]
  D --> P[patch]
  P --> V{verify-fix}
  V -->|passes| PR[Open PR]
  V -->|fails| P

The diagram's fail-and-retry branch is an inference. The page describes verify-fix only as adversarial and does not say what happens when a fix fails it.

ChatGPT Sites:

• Each Site gets a SQLite database.
• Scheduled tasks for Sites launched at DevDay.
• Sites support Sign in with ChatGPT and default to private.
• "Plugins in Sites" lets a Site show data personalised from the visiting user's plugins.
• Codex can watch a source, such as a shared sheet, and keep the Site updated.
• Editors can publish updates themselves.
• OpenAI says it hosts 8M sites.

GPT-Live delegation. In the demo, a voice agent kept listening while GPT-6 Astra wrote code in the background to draw images.

Other platform pieces:

• Sign in with ChatGPT lets users bring their own paid tokens to a third-party app.
• A (possibly new) Agents API with Computer Use.
• Plugin extensions, which are full apps inside ChatGPT and Codex.
• An OpenAI Marketplace.
• A session recommended WebMCP over Computer Use as more efficient.

Evidence

• Internal security sprint: about a quarter of product engineers took part, and they fixed 53 critical findings on day one. 36% of discoveries were duplicates and the patch rollback rate was 1%. These are OpenAI's own figures with no external baseline.
• Computer Use: OpenAI says model-optimised harness work gave a 2x latency improvement that has shipped to all users.
• Ultrafast was credited with merging the ChatGPT and Codex desktop apps in 28 days. This is an anecdote.

How to use it

• The page links "Getting started with GPT-Live" and the openai-live-console starter app on GitHub.
• The Codex Security CLI is open source. Codex Security Cloud is in the Codex plugins marketplace and should be generally available now.
• Budgeting: Ultrafast multiplies cost by 6x. If the "fifth of the price" claim holds, Sol is the cheaper option once it gets Ultrafast.

Limits

• Live demos failed repeatedly:
• A Dot stalled silently on stage.
• Voice mode failed and the presenter had to type instead.
• GPT-Live interrupted the human speakers once.
• The author hit Codex Cloud problems that morning and switched to another tool.
• Decisions API is preview-only.
• Ultrafast for Sol is not yet available.
• dots do not inherit what you have already shared with ChatGPT; you grant access explicitly.
• No benchmarks or per-token prices are given for Sol beyond the relative claims above.
• The author attended on a free ticket from OpenAI.