Maya: Hi, I'm Maya. Welcome to the Codex Insider daily briefing. Today we're tracking GPT-6 Astra's rollout, a third-party billing route, prompting techniques, a reported browser workflow, and OpenAI's developer demonstration.

James: Hi, I'm James. Let's start with GPT-6 Astra, which OpenAI began shipping September 3 across ChatGPT Plus, Pro, Business, Enterprise, the API, and Amazon Bedrock.

Maya: YourCloudDude's release rundown describes native computer use, roughly 1.05 million tokens of context, and output of up to 128,000 tokens. The computer-use capability is aimed at multi-step browser and desktop workflows alongside coding tasks.

James: Lian Lim and Shane Spencer report OpenAI's first Critical cybersecurity classification, following internal tests of autonomous vulnerability discovery and exploitation. The rollout put defenders and selected enterprise partners first, with stricter isolation and monitoring around access.

Maya: Tibor Blaho and Pankaj Kumar report standard pricing of ten dollars per million input tokens and fifty for output, matching Fable 5.1. They also describe a faster mode at double the price. Pro reasoning has separate access limits.

James: Breakdowns from Vox and George Hanu put GPT-6 Pro outside the twenty-dollar Plus chat tier. They describe two hundred weekly Pro messages on the two-hundred-dollar plan, followed by a downgrade after that allowance. Standard Astra use is counted separately.

Maya: A separate billing story comes from Reddit user masterkain, who describes Codex Pooler. The third-party routing tool is designed to let developers use Cursor through an existing OpenAI Codex subscription while keeping Cursor's editor interface.

James: The pitch is consolidation, per a Reddit post in r/cursor. You keep Cursor's editor but point its model calls at Codex Pooler's servers, so you stop tracking a second pool of credits.

Maya: The post warns that reasoning tokens don't currently pass through to Cursor. Workflows that depend on that intermediate output may therefore behave differently after switching the connection, even when the developer keeps using the same editor.

James: The configuration directs requests through Codex Pooler's servers, and the post says setup instructions are available. That is the practical change to evaluate: a third-party route for model traffic, alongside the stated limitation on reasoning output.

Maya: Once you've got Astra, the next question is how much it decides on its own. Four prompting techniques circulated this week, all treating its autonomy as a dial, not a fixed setting.

James: Angel recommends prompting Astra to infer intent and act on routine tasks, while reserving pauses for destructive or irreversible operations. The approach also puts that instruction into AGENTS.md so polite requests are understood as directions to complete work.

Maya: On the other end, Alex Prompter's three-list framework forces the agent to sort every task before acting: what it does alone, what it drafts for review, and what stays with a human.

James: Alex Prompter's approach pauses before sending, spending, or deleting. God of Prompt instead emphasizes defining which decisions need human input, while Miles Deutscher suggests asking Astra to identify recurring workflows it could complete end to end.

Maya: Here's what that autonomy looks like in practice. The How I AI channel says Astra cleared coding tasks stuck for six months in a single pass, working through Codex.

James: The creator describes a CRM automation in Chrome. They returned to an existing Codex conversation and asked for an update; Astra added an email object containing the recipient, subject, and body without manual dragging or typing.

Maya: Their framing is telling. They call it a browser-use and computer-use problem, not just coding. Before Astra, navigating those nodes was unreliable and forced them back into manual editing.

James: In that creator's account, the change was reliable execution of the browser steps as well as understanding the task. It is one reported workflow showing how computer use can remove manual interaction from an existing automation project.

Maya: OpenAI made its own developer pitch in a September fifth video highlighted by Simon Willison. The clip focuses on the model's attention to detail and its ability to turn a prompt into more sophisticated visual output.

James: The video says Astra excels at building 3D models and shows gardens, shipyards, animals, cityscapes, and Dyson spheres. Those demonstrations give developers examples of the shapes and scenes OpenAI chose to emphasize in the launch.

Maya: At one minute fifty-nine, Willison spotted a pelican riding a bicycle with a red neckerchief. He has used that unusual prompt for years as an informal visual test, making the appearance recognizable to readers following his model comparisons.

James: Willison highlighted the familiar creature within the broader developer reel. Alongside the complex 3D scenes, it offers a small, specific example of translating a written request into a visual result that people can inspect.

Maya: So the through-line this week is one launch and everything adapting to it, from billing workarounds to prompting frameworks that decide how far you let it run.

James: The models keep moving; the guardrails and the billing math are yours to set. That's your Codex Insider daily briefing.