Paperclip first stable canary release, five model prices adjusted
149,514 stars today135 new projects+5,661 public-apis
By olud.ai editorial · built from our own tracking data, published daily
Paperclip v2026.817.0 dropped as the first stable release since July and the first stable published through the new canary pipeline. The project marks a minor release milestone after nearly four weeks.
Across the tracked ecosystem 149,514 GitHub stars accumulated today across 10,000-plus open-source AI projects, with 135 new projects entering the index. The week has added 2,536 new projects and 100 new Hugging Face spaces. The public-apis repository led the day with plus 5,661 stars.
Pydantic AI v2.31.1 arrived with a bug fix that denies native structured output for Claude Sonnet 5 and Fable 5 on Bedrock. Docling v2.120.3 resolved an ODF image loading bug and added coordinate padding for OCR. Skyvern v1.0.50 bumped a patch increment.
DeepSeek adjusted two V4 models in opposite directions. V4 Flash 0423 input rose 17 percent to $0.07 per million tokens while V4 Pro 0423 input fell 44 percent to $0.66. GLM 5.2 input jumped 48 percent to $0.46, while GLM 4.6 input dropped 9 percent to $0.50. Google cut Gemma 4 26B A4B input 42 percent to $0.07. Alibaba lowered Qwen3.6 27B input 3 percent to $0.29 and Qwen3 VL 235B A22B Instruct input 19 percent to $0.21. Kimi K2.6 input rose 4 percent to $0.56.
Eight new projects entered the index today. Zero (Go, 1,546 stars) is a coding agent that answers to the user on their machine with their model and rules. Munder-difflin (TypeScript, 1,465 stars) is a local multi-agent harness. Agent-vision-toolkit (Python, 995 stars) provides vision tools for text-only models. Ima2-gen (TypeScript, 696 stars) offers a local-first visual generation runtime. Awesome-ai-for-economists (566 stars) curates AI resources for economics. KADATH (Python, 344 stars) breeds and improves agents across reproducible epochs. Juror (TypeScript, 174 stars) runs cheaper code reviews inside GitHub Actions. Goal-flow (Python, 132 stars) combines workflow graphs and agent loops on LangGraph.
HarnessEval-W proposes that world model benchmarks should output reasoning alongside scores, not just a single number. The work argues that evaluating visual rollouts requires judging whether physics, geometry and temporal consistency hold, making the justification behind the score part of the evaluation itself.
Price volatility now runs in both directions across the same model families, suggesting providers are testing demand elasticity more aggressively than usual. The research shift toward evaluative transparency signals that pure benchmark scores are losing trust.
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
