FrontierGemini 4 Carbon, which Google calls its strongest model, goes to cyber defenders first, and everyone else is promised access as soon as possible, with no dateOn October 28 Google introduced Gemini 4 Carbon and called it its strongest model, with leading cybersecurity capabilities. This week it goes to a subset of cyber defenders, then to more trusted partners through Fairwind. The wider rollout is a plan without a date.FrontierOpenAI releases GPT-6.1 Terra: a fast computer-use model at $1 per million input tokensOn October 26 GPT-6.1 Terra arrived in OpenAI's API at $1 per million input tokens and $5 for output. According to Vercel, though, OpenAI still recommends the larger GPT-6 models for the hardest code. OpenAI's announcement would not open, so there are no benchmarks here.FrontierSiri AI speaks five more languages, but in the EU it is missing on iPhone, iPad and Apple WatchOn October 26 Apple released Siri AI in French, Japanese, Korean, Portuguese and Spanish, and promises four more by the end of the year. Bulgarian is not mentioned, and in the EU Siri AI is unavailable in iOS, iPadOS and watchOS.FrontierMiniMax released M3.1-Flash: a multimodal model with about 428 billion parameters, 23 billion active and a context of up to 1 million tokensThe model page on Hugging Face is dated 25 October. According to MiniMax, 23 of the roughly 428 billion parameters work on each token, and in its own table the model leads in some rows and trails Opus 5 and DeepSeek in others.FrontierOpenAI's models arrive in Google Cloud: Astra, Sol and Luna, for now in a closed preview for select companies, billed through the cloud accountOn October 21, Google Cloud added GPT-6 Astra, GPT-6.1 Sol and GPT-6 Luna to its model garden. Access is a closed preview for select enterprise customers; payment runs through Google Cloud Marketplace, out of existing cloud commitments. OpenAI plans general availability in the coming weeks.FrontierDeepSeek opened its search to everyone: $0.75 per 1,000 calls, free through Claude Code and Codex until October 31On October 17 DeepSeek released V4.1-Flash-1017, a minor update to its model with shorter outputs and weights under MIT. The big item in the announcement is the Search & Fetch API: $0.75 per 1,000 calls per endpoint, free through DeepSeek Harness, Claude Code and Codex until billing starts on October 31 at 16:00 UTC.FrontierReflection released Beam: an open-weight model with 501 billion parameters under Apache 2.0, but even the NVFP4 version weighs about 280 GBOn October 16, Reflection released Beam: a mixture-of-experts model with 501 billion parameters, 23 billion active per token, a 1 million token context and an Apache 2.0 license. Per Hugging Face, the NVFP4 version is about 280 GB and BF16 is about 1 TB. The quality numbers are Reflection's.FrontierMeta opened Muse Spark Open at 816 billion parameters: free, but with a “Built with Muse Spark Open” badge and a contract with Meta if you sell it as a service with revenue above 20 million dollarsOn October 15 Meta released the weights of Muse Spark Open: 816 billion parameters in total, 42 billion active, a 1 million token context. The license is Meta's own: use is free, but every product carries a badge, and anyone selling model access above 20 million dollars in combined revenue signs a contract with Meta.
80 stories