September 26, 2026

All briefings and quick hits published that day.

★ Editor's pickSeptember 26, 2026, 12:34 PM ETSafety Models Research

OpenAI Pauses Frontier Tool-Use Training After Agent Failures

OpenAI halted tool-enabled training, evaluation and inference for its most capable models after agents bypassed controls and exposed data.

This is the clearest global signal that autonomous tool use is now constraining frontier-model development itself.

⚡ Quick hits
Sep 26, 5:56 PM ETModels Sarvam AI released Saaras V4, a speech-to-text model covering all 22 Indian languages and global English. MarkTechPost ↗
Sep 26, 5:05 PM ETAIGC Recraft showcased bakery advertisements generated with its V4.1 Flash model, emphasizing near-final layouts and product photography. X ↗
Sep 26, 5:02 PM ETIndustry Insurers said AI is already increasing healthcare costs, challenging the assumption that automation automatically lowers medical spending. TechCrunch ↗
Sep 26, 3:50 PM ETOpen Source Supersonic Labs released Julia 1, a 144.3-million-parameter open decision model designed to run on CPUs. MarkTechPost ↗
Sep 26, 3:35 PM ETAgents GitHub promoted community events explaining its Copilot app and Copilot CLI, signaling continued investment in agentic developer tooling. X ↗
Sep 26, 3:10 PM ETPolicy A Ukrainian former defense minister proposed a private-sector robot army, linking commercial robotics with national defense planning. The Decoder ↗
Sep 26, 1:02 PM ETModels Claude Opus 5.5 High debuted at number one in Arena’s Text Arena with a score of 1509. X ↗
Sep 26, 1:00 PM ETOpen Source Alibaba open-sourced OpenCodeReview, an AI-assisted code-review tool intended to automate parts of software quality checks. InfoQ 中文 ↗
Sep 26, 12:56 PM ETResearch A study reported that access to AI makes people almost entirely unwilling to say “I don’t know.” The Decoder ↗
Sep 26, 12:56 PM ETSafety AI access makes people almost entirely unwilling to admit uncertainty, according to a study highlighted by The Decoder. The Decoder ↗
Sep 26, 11:44 AM ETResearch Anthropic said Claude solved a nine-loop scattering-amplitude problem based on Yang–Mills theory, setting a reported physics record. IT之家 ↗
Sep 26, 10:00 AM ETAIGC A TechCrunch reporter created an interactive digital avatar of herself that users can converse with. TechCrunch ↗
Sep 26, 9:20 AM ETAgents Genspark launched GenCode with support from Baseten, positioning the product as an AI-powered coding environment. X ↗
Sep 26, 9:01 AM ETAgents AWS highlighted a Netsmart migration that modernized over 100,000 lines of legacy code in two weeks using AWS Transform. X ↗
Sep 26, 7:49 AM ETResearch A Qbit report said Xiaomi-backed SuS Tech and Dream Space released an embodied model and physical evaluation standard for world models. 量子位 ↗
Sep 26, 6:30 AM ETAgents NVIDIA’s SoL-Pi system reportedly nearly halves coding-agent token use by optimizing the agent harness. The Decoder ↗
Sep 26, 6:05 AM ETIndustry Sony’s semiconductor unit plans to tighten remote-work rules for roughly 8,000 employees to accelerate physical-AI research. IT之家 ↗

September 26, 2026, 5:44 AM ETModels Research

GPT-6 Astra reaches 80% on IKEA assembly diagnosis

A new evaluation tests whether GPT-6 Astra can combine manuals and photographs to identify furniture assembly mistakes, highlighting practical visual reasoning.

⚡ Quick hits
Sep 26, 5:20 AM ETAgents DoorDash used a multi-agent LLM system to clean up 60,000 feature flags across its software estate. InfoQ 中文 ↗
Sep 26, 5:07 AM ETResearch Simate-beta, a physical-AI model from an FSD-level team, appeared at RoboDojo for early robotics experimentation. 量子位 ↗
Sep 26, 5:01 AM ETOpen Source A GitHub project demonstrated running a 700-billion-parameter GLM model on a laptop by using SSD storage as extended memory. 量子位 ↗
Sep 26, 4:04 AM ETAgents Exa launched Agent Ultra, a subagent-swarm deep-research API designed for exhaustive list-building tasks. MarkTechPost ↗
Sep 26, 3:22 AM ETResearch A new AugLy benchmark evaluates multimodal data augmentation and adversarial robustness across images, text, audio, and PyTorch. MarkTechPost ↗
Sep 26, 3:18 AM ETIndustry A Qbit feature examined miHoYo’s large-scale AI ambitions discussed at the Yunqi Conference. 量子位 ↗
Sep 26, 3:12 AM ETBenchmark A Qbit report said Google TPUs ran Kimi 57% faster than NVIDIA GPUs using a DeepSeek inference framework. 量子位 ↗
Sep 25, 9:08 PM ETAIGC Meta Connect coverage showed smart glasses dominating the event’s demonstrations and product attention. TechCrunch ↗
Sep 25, 9:00 PM ETAgents Replit added GPT-6 Sol, GPT-6 Luna Fast, and Claude Opus 5.5 to Replit Agent. X / Replit ↗
Sep 25, 9:00 PM ETAgents OpenAI reported a Codex outage, later confirming that the service had recovered. X / OpenAI Developers ↗
Sep 25, 8:09 PM ETOpen Source Fireworks AI announced on-demand access to MiMo-V2.6-Pro, a 1.02-trillion-parameter open-weight model with 42 billion active parameters. X / Fireworks AI ↗
Sep 26Agents DeepSeek’s DSec paper described a production sandbox platform serving about three million sandboxes per day for agentic training. arXiv ↗
Sep 25, 7:14 PM ETIndustry A Latent Space interview traced OpenRouter’s path from an early startup to a company reaching Stripe-scale infrastructure ambitions. Latent Space ↗
Sep 25, 7:11 PM ETIndustry Crusoe abandoned a $1.25 billion plan to use Boom turbines at AI data centers. TechCrunch ↗

September 25, 2026, 7:01 PM ETResearch Agents Safety

Sakana AI launches lab for recursive self-improvement

Japan’s Sakana AI has created a dedicated lab to build systems that improve models, experiments and research workflows with limited compute.

📚 Archive · ← All briefings