OpenAI Product Head: Astra Long Tail Consumption Can Be Reduced by Three to Four Times

Thibault Sottiaux, head of OpenAI's core products, stated that improvements have been made for heavy long-tail scenarios using ChatGPT accounts: the model quality remains unchanged, and the subscription usage extraction can be reduced to about one-third to one-fourth of the original, meaning a maximum reduction of about 75%. He presented this as pure profit without disclosing the calculation formula.

The optimization targets are limited to ChatGPT subscription quotas, not the API pricing list. The official rate card remains at 250 points for every million input tokens and 1250 points for output in Astra, with Fast mode at 2.5 times; API input costs about $10 and output about $50 per million tokens. He also suggested that Astra's low inference tier is already stronger than the high tier of GPT-5.6 Sol, and users accustomed to running Sol at high tiers should lower it to low or medium when using Astra. Astra is included in the regular usage pool, and paid plans can allocate 100% of their quota to Astra.

Within two days of launch, reports of excessive consumption have piled up in the code repository. ChatGPT Plus users reported hitting limits in about 20 minutes; some feedback indicated that a single Astra prompt exhausted a 5-hour window. In GitHub discussions, Pro 20x users compared Astra and Sol on the same day and machine: the weekly quota interface dropped from about 45% remaining to about 5% remaining, with Astra consuming at least about 20 percentage points; local telemetry showed Astra input cache at about 96.33% and Sol at about 98.45%, with the weekly window calculated at 10080 minutes. The official estimate for Plus tier indicates that Astra local messages range from about 5 to 45 in a 5-hour window, while Sol ranges from about 10 to 100 messages, with the interval itself already stating "one heavy task consumes one-fifth of the window" in the documentation.

This is a continuous firefighting effort within the same product line. When Astra was launched in batches, reset cards were issued daily; after early release, full resets were given to Plus, Pro, and Business users. At the end of June, the official acknowledgment of automatic review, sub-agents, and backend suggestions indicated that quotas were consumed quickly, leading to full resets. This change occurred within about 72 hours after full release, targeting "heavy long-tail users logging in with ChatGPT accounts," without stating that all plans and tasks would be equally reduced.

In terms of market mechanisms, what has been reduced is the speed of deduction on the subscription progress bar, while the model weight and API pricing have been preserved. The event was driven by the progress bar dropping faster than the community expected after the release. Computing costs still occur per token, and the ledger can be rewritten on the subscription side. Benefiting are the paid users logging in with ChatGPT who run long-tail Agents; API pay-per-use customers do not see this discount. The pressure is on the need to simultaneously fulfill "quality unchanged" and "less deduction" in inference scheduling on the same set of GPUs. The undisclosed mechanism means users cannot verify if they fall into that long tail.

Source: Public Information