OpenAI Accelerates GPT-6 Models by 50 Percent, Deploys Invisible Text Watermarking System

OpenAI has delivered a significant performance boost to its latest large language models, increasing token generation speed by 50 percent across ChatGPT and pushing forward with a compliance-driven text watermarking system. On October 4–5, the company rolled out performance optimizations to GPT-6 Astra (released September 4) and GPT-6.1 Sol (released September 29), raising the default inference speed from 30 tokens per second to 50 tokens per second across all official ChatGPT endpoints. The acceleration applies to the entire GPT-6 series following an ecosystem-wide switch to an optimized tokenizer that requires fewer tokens to complete the same tasks.