GPT-6 Astra Ultrafast in Codex: how to switch it on, what it costs, what it built

How to use it

• In the Codex app: the video shows a new speed picker at the bottom of the prompt box with two lightning icons for Ultrafast, above the earlier standard and fast options. He says the $500 Pro plan is listed on the business side of the pricing page rather than the main plans page, and that it also includes DOTS, "fastest work" and Codex.
• In the API (OpenAI's documentation, which is the source to trust here): set model to gpt-6-astra and service_tier to ultrafast. OpenAI recommends a persistent WebSocket connection for agentic work with many tool calls, since network overhead can erase the latency gains. See the Ultrafast mode guide.
• Price: OpenAI's pricing page lists Ultrafast for gpt-6-astra at $60 input / $300 output per million tokens for prompts up to 272K (standard is $10 / $50), so use it where speed justifies the cost. The changelog says it is available to API customers at low rate limits, with global processing and US data residency only. Sources: pricing and changelog.

The request shape, from OpenAI's guide (Python SDK; the guide recommends the WebSocket form for agents):

from openai import OpenAI

client = OpenAI()
response = client.responses.create(
    model="gpt-6-astra",
    input="Explain why the sky is blue in one sentence.",
    service_tier="ultrafast",
)

The plan price, plan contents and Codex picker are from the video; the API price and settings above are from OpenAI's docs.

What he timed

He says Ultrafast runs roughly 8 to 10 times faster, topping out around 300 tokens per second, and ran Astra at high effort, so the displayed token rate is lower because of reasoning. Each result below includes the agent testing its own work:

• Interactive black hole observatory (WebGL): first version in about a minute, about 2 minutes with tests.
• Floor plan to a furnished 3D walkthrough: about 2 minutes to a first version.
• Fluid art instrument, 8 min 14 s; miniature city with traffic simulation, 6 min 52 s; marble machine game, 7 min; particle portrait, 5 min 11 s; website redesign for his own site, about 5.5 min; neural network classifier lab, 3 min; a scroll-driven site needing a generated asset, 7 min 12 s (about 3.5 min of development).
• He notes some runs were slower because several ran at once on his machine.

His view, and ours

His argument is that the speed turns "one-shot and hope" into build, test and fix inside a few minutes, and lets you run many attempts or an optimisation loop in parallel. That is a claim from ten of his own demos, not a benchmark. The video ends with a plug for his paid community, which isn't covered here.