OpenAI's fastest GPT-6 Astra costs 6x, but burns subscription allowances at 8x
Ultrafast charges six times Standard API token rates and uses included ChatGPT Work and Codex allowances at eight times the Standard rate. Its speed claims still need independent testing.
OpenAI's new Ultrafast mode for GPT-6 Astra charges $300 per million output tokens, six times the Standard rate. The multiple is the same on every line of the price list, so prompt caching cuts the bill but leaves the 6x gap intact.
OpenAI added the tier to its Responses API on Sept 29, according to the OpenAI API changelog. Developers turn it on by setting service_tier to ultrafast on gpt-6-astra, OpenAI's frontier model. The OpenAI Ultrafast guide calls it "the fastest service tier in the OpenAI API" and says it is open to all API users at low rate limits.
For prompts of 272,000 input tokens or fewer, the OpenAI pricing page lists Ultrafast at $60 per million input tokens, $6 cached, $75 for cache writes and $300 for output. Standard is $10, $1, $12.50 and $50. Above 272,000 input tokens, Ultrafast rises to $120, $12, $150 and $450, against Standard's $20, $2, $25 and $75. The full ladder is simple: Batch costs half of Standard, Fast costs double and Ultrafast costs six times. VentureBeat described Ultrafast as three times the Fast price, which is the same fact seen from one rung up. OpenAI's pricing page lists no Ultrafast price for the cheaper GPT-6.1 Sol.
For identical token usage, the token bill scales by six too. Take a hypothetical coding agent running 100 steps, each reading 100,000 cached tokens and 20,000 fresh ones and producing 2,000 billable output tokens. Using OpenAI's published rates, that is $40 on Standard, $80 on Fast and $240 on Ultrafast. This assumes the cache is already populated and excludes cache-write charges, tools and any additional reasoning tokens. It is a worked example, not a test run. An earlier AGI Upside Signal examined how retries and caching change the cost of a finished task.
What the extra $200 buys rests on vendor claims. Nvidia says Ultrafast runs on its Blackwell GPUs and delivers "up to 8x faster token generation" than Astra in Standard mode. Nvidia sells the chips. At DevDay, OpenAI claimed up to 8x faster generation in Codex, up to 6x through the API and up to 300 tokens per second, VentureBeat reported. OpenAI's current guide gives no speed multiple at all. Those generation-speed claims do not establish that a whole agent task finishes six or eight times faster. The sources reviewed for this article do not include an independent benchmark of Astra Ultrafast.
Inside ChatGPT Work and Codex, the math changes. The OpenAI Codex pricing page says Astra Ultrafast draws down included subscription usage at 8x the Standard rate, but purchased credits at 6x. Fast mode follows the same pattern, at 2.5x of included usage and 2x of credits. A plan's allowance therefore runs out faster than the API price multiple suggests. In ChatGPT Work and Codex, the mode is limited to the "Pro $500" plan, the top of OpenAI's $100, $200 and $500 monthly Pro options, and to eligible Enterprise and Edu plans.
Two limits narrow who can use it at scale. Ultrafast's default ceiling is 500,000 tokens per minute for usage Tiers 1 to 3, 1 million at Tier 4 and 5 million at Tier 5, according to OpenAI's guide. The OpenAI model page for GPT-6 Astra allows 40 million tokens per minute on Standard at Tier 5, eight times as much. Ultrafast also supports only US data residency and global processing. The guide says it does not work with EU or other non-US regional endpoints, which leaves it off the table for teams whose data has to stay in Europe.