OpenAI Codex head Tibo announced that the default speed of GPT-6 Astra and GPT-6.1 Sol has been optimized by about 50%. This applies to subscription users and products and partners logged in with ChatGPT accounts, including OpenCode, Pi, Amp, and Devin. Users do not need to change settings and should feel the difference within two hours.

He later added that throughput has increased from 30 TPS to 50 TPS, stating that the most optimized tokenizer has been used. The number of tokens required for these two models to complete tasks is also lower. The post did not provide benchmark tests, price changes, or capacity limits.

This is the first day of his 28-day plan. The day before, he mentioned that over the next 28 days, he would either deliver a clear improvement to most Codex and office users daily or reset the entire project. Speed was the first item, with no reset.

The speed increase refers to the default speed, not new models. Astra and Sol are included in this range. The path goes through subscriptions and ChatGPT logins, so partners do not need to release separate versions. The increase from 30 TPS to 50 TPS is about a 67% increase in throughput, but he maintains that it is about 50% faster for users. Both figures come from his post, with no third-party retesting.

The tokenizer was specifically mentioned. Faster generation combined with fewer tokens will shorten the wait time for a coding task. He did not mention whether reasoning quality, tool invocation, or long context has also sped up.

This is a release of reasoning capacity, not a price reduction. Buyers are those using Codex by subscription and coding tools that embed ChatGPT login into their products, while the seller is OpenAI, providing reasoning for Astra and Sol. The event is driven by the default speed switch, not user migration. Similarly, subscriptions can achieve higher throughput within two hours. Beneficiaries are partners who have integrated login and retain users by reducing wait times, while those still experiencing 30 TPS without this default optimization are under pressure.

Source: Public Information