AI News · August 22, 2026

SGLang v0.5.18 ships with 710 PRs from 212 contributors

149,514 stars today52 new projects+5,661 public-apis
By olud.ai editorial · built from our own tracking data, published daily

The open-source AI ecosystem gained 149,514 GitHub stars across 10,000+ tracked projects today. 52 new projects entered the index. The top gainer was "public-apis" with +5,661 stars.

This week, 2,082 new projects entered the index alongside 100 new Hugging Face spaces. Several model prices moved, with DeepSeek V4 Flash 0731 dropping 43% and GPT-5.6 Sol falling 20%.

Releases of the day

SGLang v0.5.18 arrived with 710 pull requests from 212 contributors. The release adds new model support, detailed in the project cookbook. Phoenix arize-phoenix-client-v3.3.0 enables phoenix-gql mutations by default with approval in manual mode. Nexa SDK v0.5.0 introduces a Microsoft-signed HTP catalog for llama.cpp on Windows on Snapdragon, removing the need for certificate imports.

SGLang v0.5.18 Highlights · *710 PRs from 212 contributors.* · New models in this release (see the cookbook for all supported models): · | Model | Type | PRs | Cookbook | |---|---|---|---| |... Phoenix arize-phoenix-client-v3.3.0 3.3.0 (2026-08-22) · Features · * pxi: enable phoenix-gql mutations by default with approval in manual mode (#15415) (391d0dd) NNexa SDK v0.5.0 Hexagon HTP · Microsoft-signed HTP catalog (llama.cpp @ 873e5d). No cert import required on Windows on Snapdragon. · What's Changed * ci(bench): add gemma-4 GPU cells and a...

Pricing moves

Input prices moved in both directions. DeepSeek V4 Flash 0731 dropped 43% to $0.08 per 1M tokens. GPT-5.6 Sol and GPT-5.6 Sol Pro both fell 20% to $2 per 1M tokens. Mistral Small 3.2 24B decreased 11% to $0.08 per 1M tokens. Kimi K2.7 Code fell 6% to $0.67 per 1M tokens. On the rise, MiniMax M2.5 increased 23% to $0.27 per 1M tokens. Gemma 4 31B rose 11% to $0.1 per 1M tokens. Kimi K2.6 went up 7% to $0.58 per 1M tokens.

▲ 7%Kimi K2.6Moonshot AI · input $0.54 → $0.58 per 1M tokens
▲ 11%Gemma 4 31BGoogle · input $0.09 → $0.1 per 1M tokens
▼ 6%Kimi K2.7 CodeMoonshot AI · input $0.71 → $0.67 per 1M tokens
▲ 23%MiniMax M2.5MiniMax · input $0.22 → $0.27 per 1M tokens
▼ 11%Mistral Small 3.2 24BMistral AI · input $0.09 → $0.08 per 1M tokens
▼ 20%GPT-5.6 SolOpenAI · input $2.5 → $2 per 1M tokens
▼ 20%GPT-5.6 Sol ProOpenAI · input $2.5 → $2 per 1M tokens
▼ 43%DeepSeek V4 Flash 0731DeepSeek · input $0.14 → $0.08 per 1M tokens

Research pick

EnvHarness proposes a method for generating training environments for LLM agents. Current environments are hand-built and static, unable to adapt to an agent's weaknesses. The paper suggests generating environments that evolve as the agent improves, addressing a key limitation in agent learning.

EnvHarness: Awakening Static Worlds for Agent LearningChengsong Huang, Zifeng Wang, Rujun Han et al. · 242 community upvotes · LLM agents learn by interacting with environments, yet these environments are hand-built and static: blind to an agent's weaknesses, and quickly left behind as it improves. While recent environment generation methods attempt to address this

The bigger picture

Today's facts show a market where major model prices are falling while smaller providers raise theirs, and infrastructure releases like SGLang continue to accumulate significant community contributions.

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

← All AI news