By olud.ai editorial · built from our own tracking data, published daily
DeepSeek V4 Pro input price rose 86% to $1.17 per 1M tokens, the largest single-day increase among tracked models. GLM 5.2 from Z.AI went up 22% to $0.49, while OpenAI cut gpt-oss-120b by 25% to $0.03 per 1M tokens.
The ecosystem added 270,384 GitHub stars across 10,000+ tracked open-source AI projects today, with 73 new projects entering the index. The top gainer "skills" collected +7,887 stars. This week 1,909 new projects entered the index alongside 100 new Hugging Face spaces.
Multiple minor releases shipped across agent frameworks and tooling, while llama.cpp addressed a critical GPU kernel issue affecting AMD RDNA3.5 hardware.
Releases of the day
llama.cpp b10405 removed -funsafe-math-optimizations from the ggml-hip build, which enabled -fassociative-math and could flip greedy argmax on RDNA3.5 GPUs. The fix prevents non-deterministic token selection during multi-token prediction on AMD hardware. AutoGPT beta-v0.7.1 added Tavily provider blocks for search, extract, crawl and map operations, plus Claude Sonnet 5 support with the corresponding tokenizer. OpenHands v1.13.0 introduced client-side conversation archiving. Cline v4.1.9 switched diff block text to use the editor's foreground color for better theme compatibility. Pydantic AI v2.29.0 added support for FastMCP 4 and MCP SDK v2 alongside FastMCP 3, plus Azure integrations. TRL v1.10.0 promoted DistillationTrainer to a stable trainer after a 30-PR refactor.
DeepSeek V4 Pro input price rose 86% to $1.17 per 1M tokens, the steepest increase today. GLM 5.2 from Z.AI went up 22% to $0.49, GLM 4.6 up 10% to $0.55, and Qwen3 VL 235B A22B Instruct up 24% to $0.26. Kimi K2.7 Code from Moonshot AI dropped 4% to $0.67. OpenAI cut gpt-oss-120b input price 25% to $0.03 per 1M tokens, the lowest price point among today's moves.
OpenART introduces a method for red-teaming AI agents by evolving their operating environments. Unlike standard language model testing, agents act in persistent worlds where early changes cascade into later decisions. The approach generates diverse adversarial scenarios to uncover failure modes that static tests miss.
The price landscape is fragmenting sharply, with some models doubling in cost while others become nearly free. The llama.cpp fix and the OpenART research both point to a maturing focus on reliability and safety testing for agentic systems.
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
