OpenAI Launches GPT-6.1 Sol's Ultrafast, Speed Up to 8 Times Standard Sol, Intelligence Level Close to Astra
OpenAI's developer channel announced that the Ultrafast tier of GPT-6.1 Sol is now available for API, Codex, and ChatGPT Work, with speeds reaching up to 8 times that of the standard Sol tier, and intelligence levels approaching Astra.
The API pricing is $12 per million input tokens and $60 per million output tokens, which the company states is 1.2 times the cost of Astra. The standard tier remains at $2 for input, $0.10 for cached input, and $10 for output, approximately one-fifth of Astra's standard price. The Ultrafast tier is aimed at troubleshooting, in-app agent navigation, and real-time experiences billed by the second.
Access to Codex and ChatGPT Work is limited to Pro 500, eligible pay-as-you-go enterprise plans, and credit-based educational plans. Enterprise administrators must enable permissions first. This tier covers all supported regions and supports data residency in the US and EU. The same batch also added EU data residency for GPT-6.1 Sol Fast and GPT-6 Luna Fast.
This is the fourth day of changes experienced over the past 28 days. OpenAI employees stated that the shift to control has been changed to instant response, with the model reacting faster to mid-course adjustments, allowing users to change direction in real-time and reducing the model's idle time on old instructions. Two changes were implemented together.
GPT-6.1 Sol was launched on Developer Day, September 29, first entering the Plus, Pro, Business, Enterprise, and Edu work versions and Codex, with regular chat not available at that time. The API model name is gpt-6.1-sol. Previously, Astra's Ultrafast tier had been available in Codex at a maximum of 8 times token output, approximately 6 times on the API side, with subscription limits consuming at 8 times and external credits billed at 6 times. The company clarified that the multiplier measures token output speed, not the overall task completion time.
The buyers are agents and programming users who need to change direction in real-time, while the sellers are OpenAI's high-speed inference tiers. The event is driven by the speed tier being made available for Sol, not a new model release. Funding shifts from $2 input in the standard tier to $12 input and $60 output. Beneficiaries include Pro 500 and enterprises that cannot wait for standard speeds, while other APIs that only sell at standard prices but are compared in real-time scenarios are under pressure. Tool calls and testing times are not included in this 8 times.
Source: Public Information
ABAB AI Insight
OpenAI's speed tiers did not start with Sol. In August, it previewed GPT-5.6 Sol running at 750 tokens per second using Cerebras hardware. On Developer Day, September 29, Ultrafast was formalized, with Astra in Codex at a maximum of 8 times, approximately 300 tokens per second, and about 6 times on the API side, with prices being 6 times that of the standard tier. GPT-6.1 Sol was then marked as coming soon. It only became available for API, Codex, and work versions on October 8.
The pricing structure features two price lists for the same model. Standard Sol input is $2 and output is $10, about one-fifth of Astra's price. The Ultrafast tier has an input of $12 and an output of $60, noted as 1.2 times Astra. Only Pro 500, pay-as-you-go enterprise, and credit-based educational plans can access the subscription side, with enterprise administrators defaulting to closed. The motivation is to keep Astra-like capabilities in the low-price tier, sell waiting time separately as a premium, and prevent regular Plus users from exchanging quotas for speed.
Similar offerings exist among cloud vendors' on-demand and reserved instances, as well as inference chip manufacturers selling throughput as a product. Cerebras and Groq sell tokens per second, while OpenAI sells accelerated service tiers with the same weight. The industry position is stratified: intelligence is nearing flagship status, and competition has shifted to token output speed and who can change instructions mid-course.
Essentially, the pricing power has shifted. Previously, charges were based on model tier; now, they are based on waiting time premiums. The mechanism occurs as agents wait for the next output in cycles; 8 times token output can change real-time feel, but the overall task still suffers from tool and testing delays. After the shift to instant response, idle time is reduced, and users are more willing to pay the 6 times bill for this tier.
ABAB News · Cognitive Laws
Intelligence is sold by tier, waiting is sold by multiples.
Token output speeds up, but tasks do not necessarily finish earlier.
Being able to change direction mid-course is worth paying for speed.