By olud.ai editorial · built from our own tracking data, published daily
DeepSeek cut the input price of its V4 Flash 0423 model by 44% to $0.05 per 1M tokens, the largest single price move of the day. The drop makes the model one of the cheapest in its class and signals continued price competition among frontier providers.
Across the 10,000+ tracked open-source AI projects, 61,102 GitHub stars were gained so far today. The index also added 159 new projects, with "rea" leading the daily gainers at +5,776 stars.
This week 2,095 new projects entered the index, joined by 100 new Hugging Face spaces. The ecosystem continues to expand at a steady pace, with both new entrants and established projects drawing attention.
Several other price changes occurred across the board, with some models rising sharply while others fell. The mix of cuts and hikes suggests providers are adjusting pricing based on demand and cost structures.
Releases of the day
Jan v0.8.8 shipped security fixes and updated llama.cpp to 0.7.0, recommending all users upgrade. The update carries over data and settings. Windmill v1.851.0 introduced standalone AI agents that can be built, tested and run from AI sessions, plus app conversations via a memory_id parameter. Pydantic AI v2.63.0 changed the default for CodeMode to dynamic_catalog=True, meaning a revealed tool is now appended instead of rewriting the cached prompt. Nexa SDK v0.10.0 added a Microsoft-signed HTP catalog for llama.cpp on Windows on Snapdragon, removing the need for a certificate import. Other minor releases included bugfixes for OpenCode v1.19.5, n8n 2.45.5, and Phoenix MCP 4.3.18.
DeepSeek V4 Flash 0423 input price fell 44% to $0.05 per 1M tokens, while DeepSeek V4 Pro 0423 input price rose 24% to $0.88. GLM 5.1 from Z.AI dropped 31% to $0.97. Qwen3.5-27B from Alibaba fell 4% to $0.26. Kimi K3 from Moonshot AI fell 5% to $0.18. On the upside, Gemma 3 4B from Google rose 25% to $0.05. Xiaomi raised MiMo-V2.5 input price by 233% to $0.40 and MiMo-V2.5-Pro by 60% to $0.48.
▼ 5%Kimi K3Moonshot AI · input $0.19 → $0.18 per 1M tokens
▲ 233%MiMo-V2.5Xiaomi · input $0.12 → $0.4 per 1M tokens
▲ 60%MiMo-V2.5-ProXiaomi · input $0.3 → $0.48 per 1M tokens
Emerging projects
A new project called "universal-examprep-skill" entered the index with 355 stars. It is described as an AI exam-prep tutor for Claude Code, Cursor, Codex, and Antigravity that teaches from slides with page context and cross-session memory to reduce hallucinations.
The Hugging Face community paper "ChronoReplay: One Answer Interface Is Not Enough for Streaming Spatial Intelligence" argues that training a spatial vision-language model to answer which side an object lies on does not require preserving the metric displacement that determines that answer. Supervision for a distance estimate may leave categorical relations largely unconstrained, suggesting current evaluation methods miss important gaps.
Today's facts show a market in flux: price cuts on several models contrast with sharp increases on others, security updates remain a priority for local AI tools, and new projects continue to target niche use cases like exam prep. The research pick highlights a fundamental limitation in how spatial reasoning is measured, pointing to a need for more nuanced benchmarks.
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
