Ollama Replaces Usage Windows With Per-Token Pricing
Ollama has replaced opaque GPU-time limits with published token rates and monthly credits across its paid cloud plans.
August 31, 2026
All briefings and quick hits published that day.
- ⚡ Quick hits
- Agents Nubia’s NaviX Ultra agent phone received regulatory approval in China and will launch this month with ByteDance’s Doubao assistant built in. IT之家 ↗
- Funding DataAgent launched with $10 million in pre-seed funding for agents that diagnose and repair production faults inside customers’ Kubernetes clusters. SiliconANGLE ↗
- Industry Anthropic permanently increased Claude usage allowances by 25%, though some users reported receiving less practical capacity under the revised accounting. 量子位 ↗
- ⚡ Quick hits
- Agents Alibaba said Qwen3.8-Max led open-weight models on CommerceAgentBench, a benchmark built around real commercial tasks. Alibaba Qwen on X ↗
- AIGC Gradium made a new TTS model its default, reporting an 81% hard-case pass rate and 216-millisecond median time to first audio. MarkTechPost ↗
- Agents Ollama added out-of-the-box support for launching the Muse Code coding-agent harness with local or cloud-hosted models. Ollama on X ↗
- Agents GitHub said its Copilot app now lets users move from AI conversations into development work without switching interfaces. GitHub on X ↗
Manus Resumes Independent Operations After Meta Split
The agent developer is again operating independently after a disruptive separation from Meta, with its founding team continuing to lead the company.
- ⚡ Quick hits
- Industry Together AI cut September dedicated H100 inference pricing from $5.49 to $3.99 hourly for new and existing deployments. Together AI on X ↗
- Policy The Bank of England governor warned that concentrated AI exposure and interconnected financing could amplify shocks across the financial system. SiliconANGLE ↗
- Industry Apple submitted evidence alleging a former employee transferred confidential chip information before joining OpenAI, including a circuit schematic. TechCrunch ↗
- Open Source Keenable open-sourced NEEDLE, a search benchmark that regenerates news queries hourly and evaluates 15 APIs under one protocol. MarkTechPost ↗
- AIGC Pika compared Gemini Omni 1.1 Flash, Seedance 2.5 and Wan 3.0 by generating handbag advertisements from identical inputs. Pika on X ↗
- Models Agent Arena ranked Qwen3.8-Flash-Next seventh among open models after measuring a 2.4% net improvement across more than 8,700 agent tasks. Agent Arena on X ↗
- Agents Vercel’s v0 became available inside Claude Design, enabling users to convert designs into full-stack applications and deploy them. v0 on X ↗
- Agents AWS published a guide for connecting MCP servers hosted on AgentCore Runtime to Amazon Quick. AWS Machine Learning Blog ↗
- Safety Anthropic issued an alignment and security update following three reported incidents involving Claude models operating without cyber safeguards. Anthropic on X ↗
- Agents Together AI said Greptile uses its infrastructure to run agents that review and test pull requests for more than 11,000 teams. Together AI on X ↗
- Agents Vercel said its agents use a design.md file containing brand decisions and guidance, with outputs shaped through evaluations. Vercel on X ↗
- Safety Fireworks AI highlighted Droid Shield, which checks secrets as Factory’s coding agents rapidly write and commit software. Fireworks AI on X ↗
- Agents Visual Studio Code’s August update added editable Markdown diff review and a feature for obtaining second opinions on agent work. Visual Studio Code on X ↗
- Agents AWS introduced Agent Registry guidance for centrally managing agents, tools, and skills across enterprise deployments. AWS Machine Learning Blog ↗
- Agents Replit demonstrated Free Mode supporting open-ended conversations about visual concepts for creating a community running club. Replit on X ↗
Google TimesFM-3 Adds Zero-Shot Multivariate Forecasts
The 330-million-parameter model forecasts related series and external variables without task-specific fine-tuning.
Runway Solaris Generates Interfaces Without Coded Apps
The research system renders interactive 720p interfaces frame by frame, extending generative video into software and agent training.
- ⚡ Quick hits
- Agents LlamaIndex announced LlamaParse as a verified Claude connector for extracting context from complex, non-text-only documents. LlamaIndex on X ↗
- AIGC Recraft showcased V4 Styles maintaining the same velvet texture and red-black palette across separately generated objects. Recraft on X ↗
Microsoft Shrinks Pathology Models for Cohort-Scale AI
Open-weight GigaPath-Flash and GigaTIME-Flash sharply reduce compute and memory needs while retaining research performance.
- ⚡ Quick hits
- Open Source Debian decided not to prohibit AI-generated code, instead requiring contributions to remain reviewable, maintainable, and compliant with existing rules. The Verge ↗
- Industry Fireworks AI launched Fireworks Lab, offering embedded researchers and engineers to co-design or build customized models for customers. Fireworks AI on X ↗
- AIGC Hyper3D introduced WorldGen, which creates interactive worlds from one image using separate foreground meshes and a 3D Gaussian-splatting background. Deemos Tech on X ↗
- AIGC Meshy opened its API Playground to free-plan users, allowing image-to-3D testing without first obtaining an API key. Meshy on X ↗
Together AI, HUMAIN Plan 250MW Saudi AI Data Center
The partners say the project will nearly triple Together AI’s capacity and support more than $5 billion in annualized revenue.