1300 166 138 Tell us what you need
News / OpenAI

OpenAI’s fastest GPT 6 costs six times more. Is it worth it?

By Monika Tantau • Published 1 month ago
A business owner works on a laptop while checking a tablet, weighing whether a faster and pricier OpenAI model is worth paying for

Key takeaways

• OpenAI opened Ultrafast mode for GPT 6 Astra to all API customers on 29 September
• It generates output up to six times faster in the API and up to eight times faster in Codex
• Every token costs six times the Standard rate: about $421 per million output tokens, against $70
• Ultrafast runs only on US or global processing, not on OpenAI's Australian data residency

Most Read

OpenAI’s most capable model now has a fast lane, and it is expensive. Ultrafast mode for GPT 6 Astra generates output up to six times faster through the API and up to eight times faster in Codex, where it reaches about 300 tokens per second. Every token costs six times the Standard price, according to OpenAI’s pricing page.

OpenAI added the mode to its API on 29 September, during its DevDay event. Two days later NVIDIA, whose Blackwell chips run it, published its own account of how the speed was achieved.

What Ultrafast actually changes

Ultrafast does not make Astra smarter. It is the same model, the most capable in the GPT 6 family, served on a premium tier that cuts the wait between generated tokens. OpenAI’s own documentation tells developers to use it when speed justifies the higher cost.

The gain matters most where a model works in loops. Coding agents edit, test and fix in many short rounds, and every pause adds up. NVIDIA says faster generation shortens those edit, test and debug cycles, cuts the time between tool calls and makes interactive apps feel more responsive.

Six times the price, line by line

Through the API, Standard Astra costs about $14 per million input tokens and $70 per million output tokens. Ultrafast costs about $84 and $421. Cached input and cache writes carry the same six times multiplier, and prompts above 272,000 tokens cost more again. OpenAI bills in US dollars, so these figures use the Reserve Bank rate of 22 September and will move with the exchange rate.

Default rate limits are modest, starting at 500,000 tokens per minute and reaching 5 million at the top usage tier. In ChatGPT the speed is not sold on its own. Among personal plans it comes only with ChatGPT Pro 500, which costs $769.99 a month in Australia, and it is also offered to Enterprise customers.

A man holding a notebook and pen looks up in thought, deciding whether the extra cost of GPT 6 Astra Ultrafast is justified

Who should actually pay for it

For most workloads, nobody. A support bot, a document summary or a nightly report gains little from finishing a few seconds sooner. OpenAI’s own model guide points everyday coding, research and computer use to GPT 6.1 Sol, which offers near Astra results at about a fifth of the price, and keeps Astra for the hardest reasoning work.

The case for Ultrafast is narrow but real: live coding sessions where a developer waits on every turn, agents that chain dozens of tool calls, and customer facing tools where response time shapes the experience. Even then, it is worth measuring the minutes saved against a bill that is six times larger.

Why Australian businesses should care

There is a catch for organisations with data rules. OpenAI’s documentation says Ultrafast supports only US data residency and global processing. OpenAI does offer an Australian data residency endpoint for its API, but Ultrafast cannot run on it, so requests sent through the fast tier fall outside that local arrangement.

For businesses with obligations under the Privacy Act 1988, or client contracts that require data to stay onshore, that rules Ultrafast out for anything sensitive. Our guide to AI data residency in Australia explains what onshore processing does and does not protect.

Speed becomes the new premium

AI vendors used to compete mainly on intelligence and price. Ultrafast adds a third option: paying more to wait less. OpenAI’s inference lead told NVIDIA that its models have become good at programming Blackwell chips and the coming Rubin generation, so faster tiers are likely to keep arriving.

For now, the sensible approach is to treat Ultrafast like express shipping: worth it for the jobs where time genuinely costs money, and an expensive habit everywhere else.

Scroll to Top