OpenAI DevDay 2026: Ultrafast Inference, Decisions API and Codex Security

Simon Willison attended OpenAI DevDay 2026 at Fort Mason in San Francisco and live-blogged the keynote. He notes that OpenAI gave him a ticket and a seat in the creator area. The notes cover a long list of launches. We've skipped the consumer features and kept the ones that change how people build agents and ship code.

Speed is becoming a product tier

The headline for builders is Ultrafast. The live blog reports roughly 8x faster inference, up to 300 tokens per second, at about 6x standard pricing. It is said to be available for the Astra and Sol models across the API, ChatGPT and Codex. Availability was described as today for Astra and "soon" for Sol.

That trade-off is worth working through before you adopt it. Most agent loops are bottlenecked on sequential model calls, not on a single long generation. Paying more per token only makes sense where a human waits on each step, such as interactive coding, voice, or a trading-ops console. Batch and background work should stay on the standard tier.

Decisions API: a primitive for routing and guardrails

The Decisions API was described as returning responses "in a fraction of a second" by choosing among predefined options. We haven't seen the spec, so treat the details as unconfirmed. The shape suggests classification, routing, and allow/deny checks, which are the cheap, frequent calls in an LLM-in-the-loop system.

If you gate automated actions behind a model check, a constrained-output endpoint is a better fit than free-form text you then have to parse. The enum is validated by construction. Keep deterministic checks outside the model, though. A fast answer is not a safe answer.

Identity and the security pipeline

• Sign in with ChatGPT lets users authenticate into third-party apps with their ChatGPT account. Sam Altman is quoted as saying "we should have done it a long time ago". For small products, that removes a signup flow, but it ties your auth to one vendor.
• Codex Security Cloud was announced as generally available, with vulnerability scanning, deduplication and automated patching. A Codex Security CLI was announced as open source. Automated patches still need human review and your own tests.
• Agents API was updated with computer-use capabilities, which extends agents to UIs that have no API.

What we'd do with this

1. Benchmark Ultrafast against your slowest interactive path before committing to a pricing tier.
2. Prototype one routing or approval check on the Decisions API and compare it with your current prompt-and-parse version.
3. Run the Codex Security CLI on a non-critical repo and measure its false-positive rate yourself.
4. Keep authentication portable if you adopt Sign in with ChatGPT, for example behind a standard OIDC-style abstraction.

One caveat on sourcing: this is a live blog written during the keynote, and some numbers may be shorthand. Confirm prices, limits and availability on OpenAI's documentation before you rely on them.