The Day in AI

• OpenAI and Google said on 21 October that GPT-6.1 Astra, Sol and Luna will run in Google Cloud's Model Garden, starting with a private preview.
• DeepSeek open-sourced DeepInfer, the inference framework behind its production services, under Apache 2.0 with support for NVIDIA SM100 and Huawei Ascend 950.
• Google Cloud, Red Hat and IBM released llm-d 1.0, a stable open-source Kubernetes inference stack, and Google reports time to first token cut by more than 85%.
• Mistral AI launched Voxtral 3, a speech-to-text model, a text-to-speech model and a voice agent SDK, with transcription from $0.0035 per minute.
• Anthropic launched Claude Duet, a research preview agent for professionals and small business owners built on Claude Opus 5.5.
• Recursive released Flint, an open-weight 40B mixture-of-experts coding model that the company says runs at over 125 tokens per second on a MacBook Pro.

The models will run entirely on Google's infrastructure, inside Google Cloud's access controls and data governance. Billing goes through Google Cloud Marketplace, so companies can spend down Google Cloud commitments they have already signed. Model Garden lists more than 200 models. OpenAI says Codex workflows are part of the offer and general availability is planned in the coming weeks, and it states that Microsoft remains its primary cloud partner.

GPT models are also due to appear in Antigravity, Google's coding tool, and in the Gemini Agent. Google has not said which models those will be.

Two open-source inference stacks arrived on 20 October, the same day Google released TorchTPU. DeepSeek published DeepInfer, the framework that serves its own products, with launch support for DeepSeek-V4.1-Flash on NVIDIA SM100 and Huawei Ascend 950 chips. It bundles DeepSeek's FlashMLA, DeepGEMM and DeepEP kernels with expert parallelism, draft-token decoding and deterministic sampling. It started from vLLM and has been heavily reworked since. This release covers offline inference only, and DeepSeek kept back its KV cache system, prefill and decode disaggregation, and load balancing across replicas. Technology Org counted over 12,000 GitHub stars within a day.

Those missing pieces are what llm-d 1.0 provides. Google Cloud, Red Hat and IBM released it at PyTorch Conference as a Kubernetes layer that turns many vLLM or SGLang servers into one pool. It routes requests to servers that already hold the matching cached prompt, moves KV cache to cheaper memory tiers, and splits prompt processing from token generation onto separate machines. Google reports more than 85% lower median time to first token and 2.4 times the throughput on an Ironwood TPU coding workload. It runs on NVIDIA, AMD and Intel GPUs and on Google TPUs.

Mistral AI shipped Voxtral 3 on 20 October as a full voice agent stack. Voxtral Small Transcribe 3 handles more than 50 languages with speaker labels and end-of-turn detection, and Voxtral Mini TTS 3 covers over 35 languages. The LiveKit-based SDK runs listening, reasoning, tool calls and speech in one streaming pipeline, so an agent can act on outside systems during a live call.

• Anthropic launched Claude Duet, which sends a morning brief, schedules meetings and builds dashboards, and asks for approval before payments or deletions.
• Anthropic released Claude API version 2026-10-21 with new major SDK versions that send it by default and move from the MIT licence to Apache 2.0.
• Anthropic added enforcement status fields to API user profiles, so resellers can see whether one of their end customers is restricted or banned.
• OpenAI expanded ChatGPT Ads to 89 countries and says more than 100,000 advertisers have run campaigns since launch.
• Recursive built Flint in under four months with an automated research system,.
• Aleksander Madry, former head of OpenAI's Preparedness team, has reportedly rejoined OpenAI to work on safety and security with chief scientist Jakub Pachocki.
• Anthropic reportedly pitched a draft AI safety bill that would limit federal power to block model releases in the United States.
• US federal government reportedly banned DeepSeek across federal civilian agencies over security concerns.
• OpenAI reportedly acquired Orchestra Bio to expand its biotechnology work.
• AMD reportedly acquired World Labs, bringing Fei-Fei Li into its leadership.
• OpenAI says thousands of GPT-Rosalind agents, with PerturbAI and Scripps, analysed data on millions of mouse brain cells in 32 minutes.