▲ microsoft-decision-1, openai's decisions api, and drex 1.5 go open

Hey there, This issue features Microsoft-Decision-1, a hosted decision-scoring model that undercuts OpenAI's new Decisions API on price and claims 85 ms p50. It also covers Nace AI's open-weight Drex 1.5, Alibaba's Qwen-Image-2.1-Turbo, and Sakana AI's peer-review system that catches 73% of planted core-claim errors. Then 5 Signals: TypeSafe's $870M round, Nous Research's $90M, Muse on iPad, LTX-2.5, and the Decision Index leaderboard. Nebius Opens 2026 Physical AI Awards: Five $150K Compute Prizes Run with NVIDIA._sponsored

The 7 AI stories on X’s News tab drew about 35.6K posts combined. Google Cloud’s universal Gemini AI agent for businesses led with 17K posts, roughly 48% of the total. Microsoft’s Decision-1 launch followed at 9,929. Anthropic’s multi-agent workflows in Claude beta reached 3.9K. Add Prime Intellect’s 2,000-agent swarm at 1,413, and agent stories took 63% of all posts. Alibaba’s Qwen team placed 2 stories: Qwen-Image-2.1-Turbo at 2,065 and the Qwen 4 tease at 924. vLLM’s 7.8x throughput gain on NVIDIA Vera Rubin drew 382.

▲ Microsoft AI released Microsoft-Decision-1, a Qwen3.5-9B decision-scoring model for routing, classification, verification, and agent control. It returns a calibrated probability per fixed option, no text. Post-trained from Alibaba's Qwen3.5-9B. 32,768-token context. Microsoft reports 83.5% average across 36 benchmarks, against 81.9% for Quyet-1.0-Large. p50 latency is 85 ms, p95 125 ms. Decisions flip 1.3% of the time under paraphrase and option shuffling. Calibration is 92.2, just behind Quyet's 93.1. Price is $0.042 per 1M input tokens, output free, matching Jev and under OpenAI's $0.10. The latency comparison is contested: competitor timings use JevBench's adjusted median, which H2O.ai says roughly doubles measured time, and Decision-1 is not on the JevBench board. All benchmarks are Microsoft's. Hosted only on Foundry and OpenRouter, no weights. Text only, no rationales.

Also this week’s main signals

▲ Alibaba Qwen Releases Qwen-Image-2.1-Turbo, an 8-Step 7B Image Model. An accelerated checkpoint of Qwen-Image-2.1: 8 denoising steps instead of 40.

▲ The web is the one API most agents are missing. Databases, calendars and repos have APIs. The open web mostly doesn't. The TinyFish MCP server gives any MCP client four tools: TinySearch, TinyFetch (full pages as markdown, JavaScript included), TinyBrowser for logins and forms, and TinyAgent for multi-step jobs. [Sponsored]

▲ Sakana AI's LLM Peer Review System Catches 73% of Core-Claim Errors. A TMLR paper that grades AI reviewers on finding planted mistakes, not on matching human reviews.

▲ Nace AI Open-Sources Drex 1.5: A 9B Decision Model That Scores Options, Not Text The open one. On Decision Index 0.3.1 it scores 58.08, top under 10B, against 57.96 for closed Jev 1.13.0. That gap is inside the leaderboard's tie band, and Nace ran the kit itself.

▲ OpenAI Decisions API Hits Public Beta With 10x Faster Typed Answers POST /v1/decisions turns text and images into probabilities, choices, and scores. 1 model for now, gpt-6-luna, with a 1,050,000-token context. OpenAI claims about 10× the speed of the Responses API.

▲ Datalab released OmniExtractBench, an open benchmark for structured extraction from PDFs. It pools 620 documents from 4 sources, so no single vendor picked the set. 33 documents run past 100 pages.

▲ TypeSafe AI raised $870M led by a16z at a $7.5B valuation, 24 days after Jev launched.

▲ Nous Research raised $90M to launch Hermes for Businesses; it says Hermes has been cloned 24M+ times.

▲ Nebius Opens 2026 Physical AI Awards: Five $150K Compute Prizes Run with NVIDIA (sponsored).

▲ Meta's Muse is now on iPad with Asana, Figma, GitHub, and Zoom connectors, weeks after topping iPhone.

▲ Lightricks' LTX-2.5 is a 22B video-and-audio model with multishot generation, free commercial use under $10M revenue.

▲ Decision Index 0.3.1 is the public scoreboard this week's decision models cite; check it before trusting vendor runs.