By olud.ai editorial · built from our own tracking data, published daily
DeepSeek adjusted prices on several models today, cutting V4 Pro 0813 input by 41% while raising V4 Flash 0423 by 12%. The wider ecosystem gained 64,189 GitHub stars across 10,000+ tracked projects, with archify leading at +4,210 stars. This week 1,912 new projects and 100 new Hugging Face spaces entered the index.
A new open-source project, forkd, reached 2,803 stars by promising near-instant forking of AI agent microVMs. The Hugging Face community highlighted a research paper arguing that scaling world models needs a recursive data engine with grounded rewards, not just more video data.
Releases of the day
Pydantic AI v2.36.0 added --mcp-config support and tool-call streaming, improving CLI agent integration. KoboldCpp v1.120 fixed a bug in assistant generation prefills and added DirectIO support. Phoenix MCP client updated dependencies to v7.6.0. Stability Matrix v2.16.3 introduced a CivitAI Workflow Browser that imports ComfyUI workflows directly from the model repository. All four releases are minor patches that incrementally improve tooling and compatibility.
DeepSeek lowered V4 Pro 0813 input by 41% to $0.66 per 1M tokens, and V4 Pro 0423 by 15% to $0.74. V4 Flash 0423 rose 12% to $0.09, and V3.2 rose 4% to $0.27. Alibaba raised Qwen3 VL 30B A3B by 15% to $0.15 and Qwen3.5-122B-A10B by 12% to $0.29. Moonshot AI cut Kimi K3 by 15% to $2.55, and Meta cut Muse Glimmer 30B by 14% to $0.30. The moves show a mixed directional signal across providers.
▼ 41%DeepSeek V4 Pro 0813DeepSeek · input $1.12 → $0.66 per 1M tokens
Emerging projects
Forkd is a Rust-based project that forks AI agent microVMs in roughly 100ms for 100 children and 150ms for a live branch. It uses KVM isolation and copy-on-write snapshots, targeting developers who need fast, lightweight sandboxing for agent loops.
A paper titled "Agentic Game Development as a Verifiable Trajectory Data Engine for Scaling World Models" proposes using game environments with built-in reward signals to generate training data for world models, rather than relying solely on crawled video. The approach treats agent interactions as a verifiable data engine that can scale recursively.
Today's pricing moves from DeepSeek suggest a strategic rebalancing between older and newer model tiers, while the emergence of forkd points to growing demand for fast agent isolation. The research pick reinforces a shift toward grounded, interactive data generation for world model training.
Source: olud.ai tracking of 10,000+ open-source AI projects, 300+ models and live provider pricing. Figures are measured, not estimated. All releases · Live pricing · Latest in AI
