OpenAI speeds up GPT-6 models by about 50% in ChatGPT

OpenAI made GPT-6 Astra and GPT-6.1 Sol about 50% faster in ChatGPT on October 6, and the speedup starts a 28-day pledge of daily improvements for Codex and Work users, according to a daily AI brief on Threads. The change that reaches your output is textGrain, OpenAI’s text watermark: if your team calls the API, nothing is marked until you opt in, but ChatGPT and Codex text in the EU gets the invisible watermark in the coming weeks.
For API customers the opt-in is a dashboard setting chosen per model, under Organization settings, Data controls, Text provenance. That path comes from Max Nardit’s breakdown of OpenAI’s help center. The EU part is about compliance. The EU AI Act’s transparency rules took effect on August 2 and require AI companies to mark generated content in a way other systems can identify, TechCrunch reported on the October 5 announcement.
The watermark is not a symbol. It nudges the model’s word choices in a pattern readers can’t see but a detector can. OpenAI’s technical report, co-written with researchers from the University of Pennsylvania and Yale, describes using a secret key to sort the next-word predictions. Add up hundreds of those nudges and the detector can flag the text using only the text and the key. Because the mark lives in the words, it survives copy and paste.
Rewording does weaken it. In one OpenAI test, swapping 10% of words for synonyms cut detection from about 92% to 66%, and OpenAI says short passages, math answers and translated text are harder to detect. Nardit’s reading of OpenAI’s figures adds more detail. At a 1% false positive rate, 200-token passages were caught about 80% of the time and 400-token passages about 95%. With 25% of words replaced, detection fell to 17%. Across the 24 official EU languages it ranged from 42.2% for Romanian to 69.0% for Spanish, and OpenAI then raised the watermark strength for languages below 60%.
On quality, OpenAI said it saw no meaningful change with the watermark on. Its own table for GPT-6 Astra shows GPQA Diamond at 94.44% with the watermark against 93.94% without it. But Terminal-Bench 4.0 dropped from 56.06% to 53.90% with the watermark on.
For now, only approved researchers and expert organizations get the detector. OpenAI ties that to the limits above and says those groups can help evaluate how reliable it is and what it should be used for. Nardit lists the first partners as John Thickstun at Cornell, Martin Vechev at ETH Zurich and INSAIT, and the Kempelen Institute of Intelligent Technologies.
OpenAI built a text watermark before and held it back, partly because users might move to rivals that didn’t watermark, The Wall Street Journal reported in 2024. Since then the rival has moved first. Two months ago Anthropic said it would watermark Claude text worldwide, and some Claude users pushed back.
OpenAI also warned against reading too much into a negative result: a missing watermark “does not prove human authorship.”
Up next
cc-limit-pacer drops Claude Code a model tier before lockout
Your Claude Code window is running hot, so a community plugin starts your next session on Sonnet instead of Opus and holds your batch runs. In a sample replay, that cut the hours spent locked out by about half.
Related stories
- GPT-6 Astra Pro splits one task among parallel agents
- ChatGPT answers can now be a game or a budget calculator
- Alphabet Q3, OpenAI API credits and Copilot Fusion
- GPT-6.1 Sol matches Astra on DeepSWE at a fifth of the cost
- Hugging Face sends language model calls via OpenRouter
- ChatGPT users in the EU report getting OpenAI’s dots
Comments
No comments yet. Be the first.
Join the conversation
Sign in to leave a comment. It appears after moderation.

