GLM 5.2 price skyrockets 1450% as Llama models see deep cuts
512 new projects
By olud.ai editorial · built from our own tracking data, published daily
GLM 5.2 input price jumped 1450% to $0.62 per 1M tokens on Z.AI. The increase is the largest single price move this week. Other providers moved in the opposite direction, with several models seeing steep discounts.
This week 3,864 new projects entered the index and 100 new Hugging Face spaces were added. Among the price declines, Llama 3.3 70B Instruct input price dropped 83% to $0.1 per 1M tokens, and GLM 5.3 Flash input price fell 80% to $0.03 per 1M tokens. DeepSeek V4 Pro, Gemma 4 26B A4B, Qwen3 Next 80B A3B Instruct, and Llama 4 Scout also saw reductions ranging from 10% to 44%. Kimi K3 input price edged up 3% to $0.3 per 1M tokens.
Four minor releases shipped today. Cline desktop-v0.0.48 now supports GPT-6 models on OpenAI and Azure OpenAI-compatible endpoints. Morphic v1.8.0 added account usage budgets and fixed a privacy issue with source favicons. Nexa SDK v0.9.1 brought Hexagon HTP support with a Microsoft-signed HTP catalog, removing the need for certificate import on Windows on Snapdragon. Nav2 1.5.3 released with a lyrical sync update.
Releases of the day
Cline desktop-v0.0.48 fixed a request rejection bug that blocked GPT-6 models such as gpt-6-sol and gpt-6.1-sol. Previously every request was rejected because Cline sent max_tokens instead of the expected parameter. Users on OpenAI and Azure OpenAI endpoints can now use these models without workarounds. Morphic v1.8.0 introduced account usage budgets for controlling spending, and now serves source favicons from its own origin to improve privacy. Both changes are aimed at production deployments. Nexa SDK v0.9.1 adds Hexagon HTP acceleration on Snapdragon hardware. The Microsoft-signed HTP catalog means users on Windows no longer need to import a certificate manually, simplifying setup. The update is paired with a llama.cpp commit. Nav2 1.5.3 is a smaller release focused on synchronizing lyrical components. The cumulative effect of these releases is improved hardware support, privacy, and budget control.
GLM 5.2 input price rose 1450% to $0.62 per 1M tokens, making it the most expensive model on Z.AI. Kimi K3 input price increased 3% to $0.3 per 1M tokens. On the downside, Llama 3.3 70B Instruct input price fell 83% to $0.1 per 1M tokens, and Llama 4 Scout dropped 44% to $0.1 per 1M tokens. GLM 5.3 Flash input price fell 80% to $0.03 per 1M tokens. DeepSeek V4 Pro 0423 input price fell 19% to $0.29 per 1M tokens, Gemma 4 26B A4B fell 12% to $0.07 per 1M tokens, and Qwen3 Next 80B A3B Instruct fell 10% to $0.09 per 1M tokens. Users of GLM 5.2 face a sharp cost increase, while Llama and GLM 5.3 Flash users see significant savings.
Waypoint-OPD, posted on Hugging Face with 660 upvotes, proposes a method to transfer policy shifts induced by reinforcement learning from a teacher to a student model. The technique reuses the difference in the teacher's log probabilities before and after RL training as supervision on trajectories generated by the student. A gating mechanism controls when to apply this distillation, aiming to preserve the teacher's improved behaviour while avoiding harmful shifts. The work comes from Bingxi Hou, Guochao Jiang, Guofeng Quan and others.
The dramatic price increase for GLM 5.2 contrasts with broad price cuts across other models, suggesting providers are still adjusting pricing strategies. The influx of 3,864 new projects this week, coupled with minor releases improving hardware support and privacy, points to steady ecosystem growth despite pricing volatility.
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
