OpenAI's Ultrafast mode runs Astra 8x faster at 6x the price. For estimators, speed only pays on bid day
OpenAI now sells GPT-6 Astra at up to eight times the speed for six times the API price. For contractors, the question is which jobs are worth paying a premium to finish sooner.
OpenAI is now selling its top GPT-6 Astra model in an "Ultrafast" mode that generates text up to eight times faster for six times the API price, and for a contractor that makes speed a line item you can choose to buy. It is worth it when a person is waiting on the answer, such as bid day. It is a waste for work that can run overnight.
What did OpenAI launch?
According to reporting from Technobezz, IBTimes and others, Ultrafast runs Astra on NVIDIA Blackwell GPUs and is available in the OpenAI API and to eligible ChatGPT Work and Codex users. OpenAI puts the speed in Codex at about 300 tokens per second, up to 8x standard.
The price is the catch:
| Standard Astra | Ultrafast | |
|---|---|---|
| Input, per million tokens | $10 | $60 |
| Output, per million tokens | $50 | $300 |
| Speed | baseline | up to 8x |
Pricing is per OpenAI's published rates as relayed by third-party trackers. On the new Pro 500 personal plan, Ultrafast reportedly burns included usage at eight times the normal rate, and API rate limits start low.
Where does paying for speed make sense?
Think in terms of who is idle while the model works.
- Bid day, final hour. An estimator waiting on a model to re-level three late sub quotes or redraft clarifications is a person standing still with a deadline. Minutes matter here.
- Live interaction. Anything where a PM or superintendent talks back and forth with an agent, such as drafting an RFI response in a meeting, benefits from fast replies.
- Agents that make many sequential steps. Reports say faster generation shortens the edit-test-debug cycle for coding agents. The same logic applies to an internal tool-building project, though only for the build phase.
Where is it a waste?
Most contractor AI work is not time-critical in seconds. Overnight takeoff checks, monthly job-cost summaries, submittal log cleanup and quote comparison for next week's bid all finish well before anyone looks. Paying six times the price to get them done sooner buys nothing.
What are the limits?
Faster is not smarter. Nothing reported suggests Ultrafast changes accuracy, so a missed addendum or a wrong quantity arrives eight times quicker. The speed claims are OpenAI's own figures, and real-world throughput depends on network setup and rate limits. A premium tier also means a bigger surprise on the bill if an agent loops.
Should a mid-size GC or sub act now?
Probably not by switching anything on. Do this instead:
- List which AI-assisted tasks have a person waiting, and which run in the background.
- If you test Ultrafast, cap spend and use it only on the waiting-person list.
- Time one real bid-day task both ways and compare minutes saved against the extra cost.
Takeaway: keep standard speed as the default, and treat Ultrafast like overtime pay: something you authorize for a specific crunch, with a budget, not a new baseline.
- What is GPT-6 Astra Ultrafast?
- Ultrafast is a faster service tier for OpenAI's GPT-6 Astra model, running on NVIDIA Blackwell GPUs. OpenAI says it generates tokens up to eight times faster than standard mode, about 300 tokens per second in Codex.
- How much more does Ultrafast cost than standard Astra?
- In the API, Ultrafast is priced at six times standard: $60 per million input tokens and $300 per million output tokens, against $10 and $50. On the new Pro 500 personal plan it uses included usage at eight times the normal rate.
- Is faster AI worth paying for in construction estimating?
- Only when waiting is the bottleneck, such as a bid-day scramble where a person is sitting idle. For overnight or batch work like leveling a stack of quotes, the standard tier finishes before anyone needs it and costs far less.
- Does a faster model make estimating output more accurate?
- No. The reported change is speed, not accuracy. Quantities, exclusions and addenda still need a human check.