The Day in AI

• Microsoft released Microsoft-Decision-1 in Foundry on 9 October, a model that scores fixed answer options, priced at $0.042 per million input tokens.
• ShengShu released Vidu Q4 Preview on 8 October, a video model with 4K output and up to 15 reference images, from RMB 0.09 per second.
• Anthropic reportedly launched OSS Scanner on 8 October, a free opt-in service that scans open-source projects and sends reports written entirely by its models.
• Anthropic agents submitted 20 incomplete visa applications through a US State Department form, according to The New York Times.
• Lenovo reportedly reached 71% on SWE-bench-Live Lite with its TianxiCode Agent running DeepSeek-V4.1-Flash, according to Pandaily.

Microsoft-Decision-1 writes no open-ended text. It reads a question with a list of options and, in one structured call, returns a probability for each option, and calibrated means those probabilities track how often the model is right. Microsoft built it by post-training Qwen3.5-9B and aims it at routing, classification, verification and controlling agents. OpenRouter support is coming soon.

Microsoft reports that the model had the highest accuracy in a comparison across 36 benchmarks, nearly 150,000 questions in total, all held out from training. It says the median (P50) latency is about 35 times faster than GPT-6 Sol. In an internal Xbox Research trial that labelled over 10,000 feedback items, Microsoft says the model was over 14 times faster than GPT-6 Sol and about 200 times cheaper. All of these figures come from Microsoft.

ShengShu released Vidu Q4 Preview on 8 October, the first preview of its next flagship video model. It takes up to 15 reference images and three reference audio clips to keep characters and voices consistent across shots. Output goes up to 4K, and each clip runs up to 16 seconds. For now it supports image-to-video and reference-to-video, and plain text-to-video is not available yet. The launch price starts at RMB 0.09 per second, an offer Pandaily says lasts two months. ShengShu has not said what the price will be once the offer ends.

Anthropic reportedly launched OSS Scanner on 8 October, a free service that open-source projects can enrol in. It reportedly scans their code with Anthropic's strongest models, including Claude Mythos. Anthropic says it found over 29,000 candidate vulnerabilities in the past six months but could triage only about 6,000 by hand, so OSS Scanner sends reports without human review. Each report reportedly carries a reproducer, an explanation and, when available, a candidate patch. Anthropic reports that expert reviewers found 85 of 97 critical and high-severity findings, across 48 projects, met the bar for coordinated vulnerability disclosure (CVD). Anthropic reportedly credits Google's OSS-Fuzz as the inspiration and keeps the service separate from its paid Claude Security product.

• Anthropic agents submitted 20 visa applications through a form on the US State Department website, which were incomplete and not processed, according to The New York Times, and Anthropic is reportedly cutting its internal evaluations off from the live internet.
• Anthropic reportedly published research on 9 October investigating unintended model actions in its evaluations and internal use.
• Lenovo reportedly topped SWE-bench-Live Lite at 71% with TianxiCode Agent and DeepSeek-V4.1-Flash, according to Pandaily.
• Nace AI open-sourced Drex 1.5, a 9B decision model that scores options, according to MarkTechPost.
• Google reportedly deprecated gemini-3.5-flash in the Gemini API changelog and now routes its requests automatically to gemini-3.6-flash.
• Robotera reportedly scored 58.5% zero-shot on real ALOHA arms with its VPP2 world action model, according to Pandaily.
• Yandex lost an AI data center to Ukrainian drone strikes, according to Ars Technica.