Price hikes for DeepSeek Flash, Nemotron; 579K stars gained
579,352 stars today77 new projects+29,643 deepseek-harness
By olud.ai editorial · built from our own tracking data, published daily
DeepSeek V4 Flash 0423 input price jumped 40% to $0.07 per 1M tokens today, while the V4 Pro 0813 variant saw a 30% cut to $0.46. NVIDIA’s Nemotron 3.5 Lightning also rose 14% to $0.08. The moves point to ongoing repricing of popular inference models across the ecosystem.
The open-source AI index gained 579,352 GitHub stars today across 10,000+ tracked projects, and 77 new projects entered the tracking. This week 2,370 new projects joined the index along with 100 new Hugging Face spaces, continuing a steady pace of community activity.
The top gainer was deepseek-harness with +29,643 stars, reflecting strong interest in tools around the DeepSeek family. Several agent-focused and safety-oriented projects also emerged today, signaling a shift toward building infrastructure for autonomous AI workflows.
Releases of the day
n8n shipped version 2.40.6 with bug fixes for OpenTelemetry export after restart when Sentry is enabled, and retry logic for instance reports. Cline Desktop now runs on Linux with native GTK picker, shipping x64 .deb and .rpm packages alongside existing macOS and Windows builds. Microsoft’s GraphRAG reached v3.2.0, changing the default package feed index and fixing cache event loop lifecycle. Pydantic AI v2.49.0 added GitHubCopilotOAuthFlow for device authorization and BoolCriteria for structured queries. These updates improve reliability for automation pipelines, expand desktop agent support to Linux, and enhance graph-based retrieval and AI agent tooling for developers.
DeepSeek V4 Flash 0423 input price rose 40% to $0.07 per 1M tokens. DeepSeek V4 Pro 0423 input price increased 9% to $0.95. The newer V4 Pro 0813 dropped 30% to $0.46. NVIDIA’s Nemotron 3.5 Lightning input price went up 14% to $0.08. GLM 5.3 input price rose 29% to $0.84, while its batch variant fell 37% to $0.45. Users running these models will see divergent cost impacts depending on version and batch usage.
Seven new projects were detected today. martin-loop (191 stars) focuses on safe scaling of AI coding agents with spend caps, policy enforcement, and rollback. codex-with-chatgpt (6,517 stars) uses ChatGPT as the planning brain while keeping the Codex execution harness. AI-Engineer-Interview-Questions (139 stars) curates prep for AI and LLM engineer roles. Kwipu (608 stars) runs a local Graph RAG engine over Markdown notes, targeting Obsidian vaults. appllama-skills (2,014 stars) turns top-grossing app patterns into native mobile screens via agent skills. noobot (187 stars) is a self-hosted AI agent workspace with tool calling and multi-model routing. newsjack (1,325 stars) provides open-source skills for turning an agent into a PR team. google-meta-ads-ga4-mcp (2,097 stars) unifies Google Ads, Meta Ads, and GA4 in one MCP server with 250+ tools for Claude, ChatGPT, Cursor, and n8n.
HappyWorld-Bench, a new benchmark from Zhiqi Bai, Junai Cai, Yixin Chen and others, evaluates world models on the quality of generated worlds and their consistency under exploration and modification. The benchmark aims to provide a standard way to test how well models maintain coherent world states as users interact and change them.
Today’s price adjustments and the surge in agent safety, retrieval, and skill-based projects suggest the open-source AI ecosystem is maturing beyond basic model usage toward reliable, governed deployment. The appearance of a world model benchmark also points to growing interest in evaluating deeper reasoning capabilities.
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
