GLM 5.2 price drops 95%, Nemotron 3 Super surges 275%
68,399 stars today63 new projects+9,914 rea
By olud.ai editorial · built from our own tracking data, published daily
GLM 5.2 from Z.AI saw its input price fall 95% to $0.03 per 1M tokens, the largest single-day price move among tracked models. At the other end, NVIDIA's Nemotron 3 Super input price jumped 275% to $0.3 per 1M tokens. The day's price swings reshuffle cost rankings for developers choosing between frontier and mid-tier models.
The ecosystem gained 68,399 GitHub stars across the 10,000+ tracked open-source AI projects so far today, with 63 new projects entering the index. The top gainer "rea" added 9,914 stars. This week 1,908 new projects entered the index, plus 100 new Hugging Face spaces, continuing a steady influx of experimental tools.
Releases of the day
Four projects shipped minor patches. OpenHands v1.30.1 refactored Markdown preview rendering into CSS. JAX v0.12.0.1 fixed a Windows bug where C++-generated LLVM IR was empty, now enabling vectorized CPU intrinsics on that platform. Weaviate v1.40.2 applied a security patch for eight eMudhra intermediate certificates. Unstructured 0.27.33 fixed chunking to avoid repeating very large split elements across every chunk and updated dependencies.
Six models changed input prices today. GLM 5.2 (Z.AI) dropped 95% to $0.03 per 1M tokens. GLM 5.1 (Z.AI) rose 44% to $1.4 per 1M tokens. DeepSeek V4 Pro 0423 increased 14% to $0.4 per 1M tokens. Alibaba lowered Qwen3 Next 80B A3B Instruct by 10% to $0.09 and Qwen3.5-9B by 20% to $0.08. Xiaomi cut MiMo-V2.5 by 14% to $0.12 and MiMo-V2.5-Pro by 32% to $0.3. NVIDIA raised Nemotron 3 Super by 275% to $0.3.
▼ 14%MiMo-V2.5Xiaomi · input $0.14 → $0.12 per 1M tokens
▼ 32%MiMo-V2.5-ProXiaomi · input $0.44 → $0.3 per 1M tokens
Emerging projects
Eight new projects entered the index today. folkbench (2,256 stars, Rust) is an open-source AI API relay benchmark and ranking tool with a switcher for Claude Code and Codex. new-term-radar (1,606 stars, Python) scans X, Hugging Face, GitHub and sitemaps daily for new AI terms. StartLux-Decision (1,114 stars, Python) offers open decision models from 0.8B to 27B that output calibrated probabilities. FDEOps (972 stars, JavaScript) packages forward-deployed engineering skills for AI coding agents. blender-plugin (568 stars, Python) brings AI images, video, 3D and PBR materials into Blender 5.0+ viewport with a local MCP server. AesCode (123 stars, Python) generates aesthetic HTML/CSS infographics from prompts using decoupled cross-modal rewards. ai-video-production-editor (135 stars, TypeScript) is an open-source video editor for filmmakers and AI artists built on React and Electron. cjkfold (203 stars, Python) folds look-alike CJK glyphs to improve text extraction from PDFs.
A Hugging Face community paper titled "Cache Engineering for Agents" by Tianyu Fu, Ruoxi Wang, Zihan Min and others addresses the growing cost of key-value cache I/O and storage as LLM agents handle longer trajectories. The authors argue that cache engineering, not just model improvements, is needed to keep agent workflows affordable.
Today's price moves show a market still in flux: aggressive cuts from Z.AI and Alibaba compete with sharp increases from NVIDIA and DeepSeek. Meanwhile, the surge in agent-focused tools and cache research signals that infrastructure, not just model capability, is becoming the next battleground for practical AI deployment.
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
