OpenAI Launches Ultrafast Speed Tier and New Pro 500 Subscription
OpenAI Launches Ultrafast and Pro 500
OpenAI has launched the Ultrafast speed tier and simultaneously introduced the new Pro 500 subscription. The company claims this tier generates tokens on Codex at a rate approximately 8 times higher than standard, about 300 tokens per second, and up to 6 times higher on the API.
The Ultrafast for gpt-6 astra is now available for API, as well as in ChatGPT Work and Codex under Pro 500 and enterprise plans; the Ultrafast for gpt-6.1 sol will follow later. The developer documentation states that this tier is fully open in the API for astra, but the default rate limits are lower: 500,000 tokens per minute for tiers 1-3, 1,000,000 for tier 4, and 5,000,000 for tier 5, with a recommendation to use WebSocket to maintain latency benefits during multi-tool calls.
According to the official comparison on the Codex side, Ultrafast is about 8 times faster than the standard astra tier and about 4 times faster than the existing astra Fast tier. The API documentation states that this tier is more expensive, supports only U.S. data residency and global processing, and does not support processing endpoints in the EU and other non-U.S. regions.
The Pro 500 subscription costs $500 per month and is positioned as the highest usage tier, approximately 25 times that of Plus, unlocking Ultrafast in Work and Codex. The company has also reopened the Pro 200 subscription, promising to retain astra and the daily usable gpt-6.1 sol, and will no longer restore the 5-hour limit; reports indicate that starting October 30, the usage multiplier for the $200 tier in Work and Codex will drop from 20 times that of Plus to 10 times, and the weekly message limit for gpt-6 Pro in Chat will decrease from 200 to 100 messages, with existing user limits retained until October 29, after which they will receive a usage credit expiring at the end of the year.
In August, the Ultrafast preview achieved approximately 14 times the output of gpt-5.6 sol using cerebras inference, with a maximum of about 750 tokens per second, but was limited to a small preview. This time, the speed tier is being made available as a formal sale tier for astra, tied to higher-priced subscriptions.
In terms of market mechanics, buyers are those coding and proxy teams constrained by quotas, as well as API developers needing real-time response experiences; sellers are those separating latency from model capabilities into a standalone SKU. Funding is shifting from standard inference to accelerated tiers and the $500 tier; beneficiaries are heavy Codex users and enterprises able to convert iteration speed into delivery density, while the burden falls on the original usage expectations of the $200 tier and individual developers relying on mid-tier subscriptions to handle peak loads. This is an event-driven stratification: the same model is repriced based on 'can wait' and 'cannot wait.'
The standard price for the API side of astra remains $10 per million tokens for input and $50 for output; the existing Fast tier was previously about 2.5 times the standard price. Ultrafast raises the speed premium another tier, making it more likely for resident proxies and Dots-type background tasks to remain at standard or sol, pushing interactive coding and real-time loops onto accelerated billing.
Source: Public Information
ABAB AI Insight
OpenAI's sale of speed is not new. The August Ultrafast preview connected 5.6 sol to cerebras, demonstrating flagship capabilities approaching real-time with about 750 tokens per second; subsequently, astra introduced a Fast tier, trading 2.5 times the price for 2.5 times the speed. DevDay formalizes this line into a subscription product: standard, Fast, Ultrafast tiers, plus a $500 usage cap. Historically, this mirrors how cloud computing breaks down the same chip into on-demand, reserved, and accelerated instances—no model upgrade, just a change in billing granularity.
The capital path is a two-way cut. Upwards: Pro 500 uses a 25 times Plus quota and exclusive acceleration access, siphoning heavy coding from the $200 tier. Downwards: reopening the $200 tier but cutting usage and eliminating the 5-hour window effectively forces users to burn through their weekly quota faster, hitting the purchase threshold sooner. The API opens astra Ultrafast to all developers, but with low default tpm and regional restrictions limiting scale, high usage must seek quota from customer teams. Resource allocation focuses on the fastest segment of the inference cluster, specifically feeding Codex typing rhythm and WebSocket proxies, rather than Dots-type overnight background tasks.
Comparable examples include NVIDIA selling different service levels for the same GPU, and Cursor and Anthropic turning speed and context into paid switches. The industry phase shifts from 'scoring flagship points' to 'grabbing the seconds while the fingers are still on the keyboard.' Whoever first makes coding feel close to local completion will dominate the default paths in IDEs and workspaces; closely scoring sol and sonnet continue to compete on unit price, while latency becomes another product line.
The structural judgment indicates a transfer of pricing power. The mechanism separates 'intelligence' from 'waiting time for results' for sale: when safety assessments constrain larger models, what can be immediately monetized is the accelerated copies of already passed flagship models. Quotas dropping from 20 times to 10 times, then introducing a new 25 times tier, creates upgrade justifications through product tiers rather than proving the $500 tier with new capabilities alone. The result is a division of labor in the supply chain between inference speed wholesalers and subscription wholesalers, with model labs operating on both sides.
ABAB News · Cognitive Laws
Same intelligence, repriced by waiting time
Where quotas decrease, new tiers are born
Real-time sold to fingers, overnight left for background