Issue 2: Steer while it works

This week

Frontier model for reasoning, coding, computer use, research, and docs. First OpenAI model at Critical cybersecurity capability. Rolling out to a limited set of orgs today; Plus, Pro, Business, Enterprise, API, and AWS in the coming days. Daybreak less-restrictive defender access for Astra follows in the coming weeks, not on launch day.

Astra is OpenAI's first Critical cybersecurity capability model

Anyone running agents needs to know the capability bar and the gated defender access, not just the marketing leap. Also: https://openai.com/index/safety-overview-gpt-6-astra/

OpenAI added async tool calling and mid-turn steering for long Astra jobs

Harness features matter as much as the weights for people who work with agents.

Google shipped Gemini 3.8 Flash and 3.8 Flash Cyber

Third Flash in six weeks; cheap long-horizon coding/agents plus Fairwind cyber for defenders.

GPT-6 Astra is not interesting because OpenAI said AGI. It is interesting because the Responses API now assumes your agent job is long, interruptible, and tool-heavy.

Async tool calling means the model can keep working while your app runs a function or custom tool and returns later. Mid-turn steering means you can push new instructions over WebSockets while a response is still in flight. Reasoning effort can change mid-conversation without throwing away the cached prefix. Tool calling for Astra requires the Responses API. Chat Completions tool paths are a dead end for this model.

A destination is a harness that can take a correction without restarting the run. Not “prompt in, final answer out.” Not five agents that cannot accept a mid-flight note. One agent with a steerable loop beats a swarm that only listens at the start.

Constraints are the walls. Do not worship the model card. Critical-cyber capability and defender gating matter for how carefully you sandbox tool access. They are not the thesis. Stay concrete for operators. If the only interface is a chat box and a while-loop, you bought a chat model and bolted a harness on top.

Done is a check you can perform without taste. Your harness can start an Astra Responses job. It can run tools asynchronously while the model continues. It can inject a mid-turn steer without killing the response. It fails closed if you are still on Chat Completions tools. If any of those are missing, you are not ready for Astra. You are ready for a demo.

Treat the harness as the product. Steer while it works.