On October 9, Microsoft released Decision-1, a scoring model for structured decisions. It writes no text — it picks the best option from a given set and returns probability scores. Built on Alibaba's Qwen3.5-9B, input costs $0.042 per million tokens, output is free. The decision-model track that startups heated up last month now has Microsoft, OpenAI, and Cloudflare all in.
StepFun shipped its flagship Step 5 Preview — 600B total parameters, 27B active, 1M-token context, aimed at coding, finance, and long-horizon agent tasks. The API is live on OpenRouter; open weights expected October 15.
Nvidia-backed Reflection AI released its first open-weight model Beam on October 5: 501B total parameters with only 23B active per token, 1M-token context. The company claims reasoning performance close to Zhipu's GLM-5.2 at one-quarter to one-third the inference compute, with weights coming later in October under Apache 2.0.
OpenAI Codex lead Tibo publicly committed: for the next 28 days, each day the team will either ship one clearly useful improvement for most Codex and Work users, or issue a full credit reset. The first time anyone turned "doing homework" into a public bet.
On September 22, Alibaba unveiled its full-stack AI strategy at the Apsara Conference in Hangzhou: Qwen 4 is training, future versions planned at 5–10 trillion parameters; its in-house Zhenwu V900 AI chip is 3x the previous generation, mass production in Q1 2027; Alibaba Cloud targets over 20GW of global data center capacity by 2032.