A developer’s October 9 personal blog post describes building a newsletter index through Codex voice conversations while cooking. Before deployment, he reviewed the code and returned to typing to change the imports and presentation. This is a user’s account rather than an OpenAI announcement; the author discloses a past paid OpenAI preview appearance.
The same developer’s ttok 1.0 release on GitHub on October 9 changes its default tokenizer to o200k_base. His blog explanation that day says the inference that GPT-6 shares GPT-5’s tokenizer rests on experiments, without official OpenAI confirmation. A tool’s default setting should not be read as confirmation of a vendor’s model specification.
Prime Intellect published its account of rewriting Prime Agent in Rust on October 9. The company’s diagram shows 2,209 agents: 1,981 for Rust implementation and 228 for performance improvements; the performance measurements used a scripted model rather than a live one. The company also acknowledges bugs beyond the differential tests’ coverage, so passing those tests does not establish complete behavioral equivalence.
Nous Research corrected Step 5 Preview’s final Hermes Index score to 32.83 on October 9. The originally announced 33.89 is GPT-6 Luna’s score, so the correction removes the reported tie. This checks a dated announcement reproduced on the company’s homepage, rather than independently recalculating the benchmark.
Correction on the company’s homepage
GPT Codex is made by OpenAI; this collection includes a user’s experience with an OpenAI product and comparisons involving an OpenAI model.