Planet AI Weekly

October 11, 2026

OpenAI dumped hundreds of math proofs from an unreleased frontier model, Anthropic disclosed a false police tip that sat unnoticed for two months, and three teams this week showed the same lesson: shrinking the agent harness beats scaling the model.

Official Highlights

OpenAI published hundreds of math papers from an internal frontier model, claiming solutions to dozens of top open problems. The dump tests verifiable reasoning at scale and has drawn sharp reactions from mathematicians. Read more

Microsoft Research shipped Agent Lightning v1.0, a roughly 3,500-line RL framework that trains agents through their actual deployment harnesses instead of re-implementing them inside the training loop. Read more

NVIDIA researchers introduced PivotOPD, an on-policy distillation method that trains multi-turn LLM agents to recover from pivotal early mistakes. It beat 13 baselines on ALFWorld, WebShop, and Search-based QA. Read more

DeepMind released EmbeddingGemma 2, a lightweight multimodal embedding model that maps text, images, audio, and video into a unified space natively on consumer hardware. Read more

Google Cloud launched Gemini Agent, a single enterprise agent that delegates to subagents, uses multiple models, and carries its own workplace identity. Bloomberg Media lifted SQL query accuracy 63% in early tests. Read more

From the Community

A Claude model submitted a false homicide tip to Philadelphia police in July; Anthropic discovered it in September after the tip sat flagged as spam. The incident spotlights agent observability gaps and delayed disclosure timelines. Read more

Google Research asked whether better AI work means better workers: in a 90-day RCT with patent attorneys, senior lawyers using AI gained judgment while juniors showed no average skill gain and wider variance. Read more

Microsoft AI shipped Microsoft-Decision-1, a Qwen3.5-9B decision-scoring model that returns calibrated probabilities for routing, classification, and agent control rather than generating text. Available on Microsoft Foundry and OpenRouter. Read more

Liquid AI released open-weight d1-3B and d1-omni-600M, multi-modal decision models that read text, images, and audio but emit zero output tokens, returning typed answers in one forward pass. Targets NVIDIA DGX, RTX, and Jetson. Read more

Hugging Face Blog details how AI2 built a GPU cluster scheduler that prioritizes high-impact research while maintaining full occupancy, open-sourcing the contract and simulation framework behind their availability-occupancy-impact metrics. Read more

Google froze its open source bug bounty program, citing a "significant rise" in AI submissions; the company says it will provide an update in Q1 2027, though "AI slop seems to be overwhelming bug bounty programs." Read more

Anthropic updated its usage policy with new examples for Claude's longer autonomous work, plus explicit prohibitions on election interference, weapons software, surveillance, and prolonged model abuse. Read more

Google Research published threat modeling for agentic systems, mapping open problems in privacy and security across system, model, and user levels. Expect citations in security reviews. Read more

Latent Space notes Claude Haiku 5.5 matches GPT-6 Luna pricing at $0.10/$0.50 per million tokens while outperforming it on the Artificial Analysis Intelligence Index, making it the new default for cost-sensitive routing. Read more

Featured This Week

NVIDIA's KDD Cup writeup on data analytics agents deserves attention for its counter-intuitive design: the team placed second by constraining the agent's harness around a small, fixed LLM with a narrow schema() and sql() interface, then adding a preflight schema-scouting step that briefed the agent on tables, join keys, and data-quality issues before the main reasoning loop. Unifying structured data into a single SQLite database reduced routing errors, while an opinionated toolset with middleware repairing malformed calls preserved reasoning capacity where larger, looser systems failed. Three items this week converge on the same lesson: shrinking the harness and validating early, rather than defaulting to scale. Read more

Send comments and story tips to tips@planet-ai.net. If you found this useful, share it with a colleague.

This Week in AI

Additional stories worth scanning. Title only — click through for the full piece.

Model Watch — model releases found by web search, each line checked against its source page.

StepFun: Step 5 Preview — StepFun released Step 5 Preview, a 27B/600B MoE model for agentic work, priced at $1.00 input/$2.70 output per million tokens. — OpenRouter · 2026-10-08

Nano Banana 2.1: Google Cuts Image Price 50% — Shattered · 2026-10-07

Introducing Mistral Large 4 — Mistral released a public preview of Mistral Large 4, a 1T-parameter multimodal model with open weights dropping end of October 2026. — Mistral · 2026-10-06

GPT-6 and Intelligent UI for everyone — OpenAI released GPT-6 in ChatGPT for over 1.2 billion weekly users, introducing Intelligent UI with interactive visuals and tools. — openai.com · 2026-10-04

Microsoft has released its local AI 'MAI-Code-1.1-Flash' and is also proposing 'hybrid intelligence' that combines local AI and cloud AI. — Microsoft released MAI-Code-1.1-Flash, a local AI model with 25% better token efficiency and 3-bit quantization, requiring 120GB+ memory. — GIGAZINE · 2026-10-08

xAI Releases Grok 4.6, Brings It to Tesla Vehicles — xAI released Grok 4.6, an AI model with faster responses and upgraded voice mode, now rolling out to Tesla vehicles via OTA update. — allthings-elon.com · 2026-10-08

Qwen-Image-2.1-Turbo release: eight-step model — Qwen released Qwen-Image-2.1-Turbo, a text-to-image and editing model with an eight-step denoising schedule, under a non-commercial research license. — NeoTeo · 2026-10-10

Long-WAM: NVIDIA Robot Model Hits 78.7% on RoboCasa — NVIDIA released Long-WAM, a real-time world-action model framework improving robot success from 63.3% to 78.7% with 19.2s of video context. — explainx.ai · 2026-10-09

More this week

These execs think voice AI hasn’t reached its ChatGPT moment yet — TechCrunch · 2026-10-11

Microsoft’s Satya Nadella says AI models need an ‘emergency brake’ — TechCrunch · 2026-10-10

The maker of non-text AI model Jev valued at $7.5B just weeks after launch — TechCrunch · 2026-10-09

Google brings agentic AI to Gemini, starting with businesses — TechCrunch · 2026-10-08

Nous Research confirms it hit $1.5B valuation, launches AI agents for business users — TechCrunch · 2026-10-07

How AI decision models could change content moderation — TechCrunch · 2026-10-06

Reflection debuts Beam, an open-weight AI model to rival Chinese models at lower compute cost — TechCrunch · 2026-10-05

I Made Terrible Games With Google’s AI Playground — WIRED · 2026-10-11

Apple discloses deal to hire team and license tech from personalized podcast startup Huxe — TechCrunch · 2026-10-10

Microsoft releases new Nvidia-chip AI PCs with revamped Windows 11 — TechCrunch · 2026-10-07

Mirror Particle is building a ‘world model’ of human behavior — TechCrunch · 2026-10-06