Promptfoo defaults to GPT-5.6+, TRL trains beyond 1M tokens
57,684 stars today69 new projects+3,433 archify
By olud.ai editorial · built from our own tracking data, published daily
Promptfoo released version 0.123.0 with a breaking change that defaults GPT-5.6 and later models to the Responses API. The update also adds native audio grading with llm-rubric. TRL v1.13.0 introduces long-context training beyond 1 million tokens, with a runnable example that trains a book-length sequence per step on a single 8xH100 node.
The ecosystem gained 57,684 GitHub stars across the 10,000+ tracked open-source AI projects today. 69 new projects entered the index. This week 2,287 new projects were added, along with 100 new Hugging Face spaces.
The top gainer today is "archify" with +3,433 stars. Price moves were mixed across eight models, with several significant drops and two notable increases.
Releases of the day
n8n 2.38.6 improves performance by bounding peak memory during source control push. Cline desktop-v0.0.25 fixes the ChatGPT subscription model picker to show only models the user's plan can actually use. Chatbox v1.23.2 adds OAuth support and ships builds for macOS, Windows, Linux, and Android. Promptfoo 0.123.0 breaks existing configurations by defaulting GPT-5.6+ to the Responses API and adds native audio grading. TRL v1.13.0 enables training beyond 1 million tokens with a new long-context guide and example.
DeepSeek V4 Flash 0423 input price dropped 12% to $0.07 per 1M tokens. GLM 5.2 from Z.AI fell 71% to $0.28. DeepSeek V4 Pro 0423 fell 8% to $0.87. Qwen3 235B A22B Instruct from Alibaba rose 144% to $0.22. Qwen3 Next 80B A3B Instruct dropped 10% to $0.09. GLM 4.6 fell 22% to $0.43. MiniMax M1 rose 38% to $0.55. Granite 4.2 8B from IBM fell 40% to $0.06.
Show-Harness, a paper by Yanzhe Chen, Zechen Bai, Zhijun Cao and others, presents an embodied harness that lets vision-language models control robots directly. The approach uses foundation VLMs to play robots without additional training, translating broad world knowledge into physical action.
Today's releases push both evaluation and training boundaries — Promptfoo adapts to new API defaults while TRL extends context length. Price cuts dominate, especially on GLM and Granite models, suggesting continued competition in the inference market.
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
