Summary
Nate Herk puts Codex’s new Ultrafast mode through real-world tests to see whether it justifies the $500-per-month plan it requires. Ultrafast is available only with GPT-6 Astra, is marketed as eight times faster than standard and four times faster than the existing fast mode, and can reach about 300 tokens per second. It is also available through the API and is expected to come to GPT-6.1 Sol.
The first test asked Astra to cut a 60-second vertical showreel from a Google San Jose trip, using roughly 11 minutes of narration and more than 145 GB of raw footage. Both runs used extra-high reasoning. Ultrafast finished in about 19 minutes and consumed 17% of the weekly Pro usage, while standard took 50 minutes and used 4%. That is roughly 2.5 times faster, at more than four times the usage. The presenter judged the Ultrafast cut to be the better story.
A second test generated a YouTube resource guide from a video comparing OpenAI Dots and Meta’s Muse. Standard took 13 minutes and Ultrafast 1 minute 43 seconds, close to the advertised speedup, with only a small increase in usage. The takeaway is that Ultrafast suits short, time-pressed tasks, while long agentic jobs can burn through usage quickly.
📺 Source: Nate Herk | AI Automation · Published October 01, 2026
🏷️ Format: Benchmark Test
