Sakana Fugu Max Targets Frontier AI at Lower Cost
Sakana AI released Fugu Max and Fugu Ultra v2, using model orchestration to pursue cheaper agent performance and higher autonomous-task scores.
September 10, 2026
All briefings and quick hits published that day.
- ⚡ Quick hits
- AIGC Hailuo AI promoted a September 24 creator event in Lisbon featuring films, creator talks, and new generative-video ideas. X / Hailuo AI ↗
- Research OpenDiscoveryTrace proposes process traces for evaluating AI-scientist workflows rather than judging only their final discoveries. arXiv ↗
- Agents Subagents vs Agent Skills studies how reusable knowledge execution affects long-horizon agentic tasks. arXiv ↗
- Research HuRo proposes robotizing human videos to create scalable training data for vision-language-action pretraining. arXiv ↗
- Research ReactHuman introduces a physics-grounded benchmark for human-like reactive decisions in embodied multimodal language models. arXiv ↗
- Agents The Menu Is an Execution Prior models online agents with state-path tool menus to guide action selection. arXiv ↗
- ⚡ Quick hits
- AIGC Meshy showed an API workflow in which every asset in a project was generated through its service. X / MeshyAI ↗
- Industry Replit announced general availability of native Databricks deployment, carrying Databricks security and governance into AI-built applications. X / Replit ↗
- Open Source Ollama said ChatGPT users can select local or cloud models, with plugins, web search, and computer use working through its app. X / Ollama ↗
- Safety Fara Research updated its AI Security Leaderboard and reported zero universal jailbreaks across the latest OpenAI and Anthropic models. X / Fara Research ↗
- Agents MongoDB showcased a CrewAI-powered shopping agent that checks inventory and completes checkout inside an application. X / MongoDB ↗
- Industry MiniMax joined Nebius’s AI Builder Program, giving developers access to MiniMax alongside partner models, infrastructure, and tooling. X / MiniMax ↗
- Agents Salesforce previewed an Enterprise AI Harness to ground multiple models and agents in shared business context, with broad availability planned for early 2027. SiliconANGLE AI ↗
- Agents Artificial Analysis ranked Octen Search third on its Search Index, with 16.9-second task time and about $9 per 1,000 tasks. X / Artificial Analysis ↗
- Research Google Research introduced ToolGrad, a method that generates tool-use datasets with textual gradients to improve training efficiency. Google Research ↗
- AIGC Krea launched Flux Video Edit, enabling fast, precise replacement of objects and regions within generated or existing videos. X @krea_ai ↗
- Industry Together AI introduced preemptible GPU clusters priced at 50% of on-demand rates for evaluations, fine-tuning, and batch inference. X @togethercompute ↗
- Research NVIDIA detailed BioNeMo Inference Runtime, reporting 2.90x higher Boltz-2 folding throughput and 58.5K residues per GPU-hour on eight H100s. MarkTechPost ↗
- Models NVIDIA said Nemotron 3 Embed 8B ranked first on Q2D-Web’s combined nDCG@10 leaderboard across 190 million web documents. X @NVIDIAAI ↗
- Agents Cursor introduced Projects, persistent workspaces where agents share memory, artifacts, reminders, scheduled tasks, and automated issue handling. X @cursor_ai ↗
OpenAI Opens Agents API With Managed Sandboxes
OpenAI’s public-beta Agents API lets developers run Codex-style agents with hosted sandboxes, tools, files and configurable compute.
- ⚡ Quick hits
- Agents Amazon made Quick generally available on Windows and macOS, extending its agentic workplace assistant beyond browser-based access. AWS Machine Learning ↗
- Open Source Cognition acquired the Dioxus team, adding maintainers of a popular open-source Rust framework used extensively in Devin. X @cognition ↗
- AIGC HeyGen said its real-time AI experience framework, built with OpenAI, is now solved and open sourced for developers. X @HeyGen ↗
- Models Cohere released North Small Translate, supporting more than 50 languages and reporting an 83.6 average WMT score. X @cohere ↗
- AIGC Genspark introduced Gen-1 Slides, an open-weight-based model trained with Fireworks AI to power its automated presentation tools. X @genspark_ai ↗
Cognition Releases SWE-2 Coding Model at Lower Cost
Cognition’s SWE-2 reaches near-frontier coding scores while claiming sharply lower cost and fewer turns on long software tasks.
- ⚡ Quick hits
- Agents OpenAI launched a Data agent in ChatGPT Work that connects governed enterprise data to natural-language analysis, dashboards, and approved actions. OpenAI ↗
Universal Music and ElevenLabs Form AI Music Partnership
Universal Music and ElevenLabs announce a multi-year partnership to develop licensed AI music tools with artists, labels and publishers.
- ⚡ Quick hits
- Open Source Gradio introduced Workflow1111, letting users run eleven AI pipelines online without downloading large checkpoints or local environments. X @Gradio ↗
- Industry CoreWeave launched an engineering service to help enterprises implement physical AI systems, including robotics and industrial workloads. SiliconANGLE ↗