AI News · October 29, 2026

DeepSeek V4 Flash price jumps 80%, GLM 5.3 Flash drops 80%

84 new projects

DeepSeek raised the input price of its V4 Flash 0423 model by 80% to $0.09 per 1M tokens, while Z.AI cut the price of GLM 5.3 Flash by 80% to $0.03 per 1M tokens. These are the largest single-day price moves on the platform, with several other models also seeing adjustments. The index added 2,217 new open-source projects and 100 new Hugging Face spaces this week.

The research community published a new method for training diffusion language models with reinforcement learning, gaining 256 upvotes on Hugging Face. Meanwhile, eight emerging projects entered the tracker today, led by a Kotlin-based agent that can control any app without integrations and a Rust filesystem tool for agent workspaces.

On the release front, six projects shipped minor updates. Open WebUI added an analytics session explorer for administrators, and CVAT introduced the ability to create and delete replica jobs. CrewAI expanded its skill search and deployment options.

Releases of the day

n8n 2.45.6 fixed a bug in queue mode execution timing and adjusted sidebar icon sizes. Open WebUI v0.12.5 added a Sessions tab in Admin Analytics that lists each chat with its tokens, responses, errors and time span. Lobe Chat v2.2.21 was published automatically from a pull request. CrewAI 1.15.33 added skill search for listing organization skills, support for passing --inputs to Python flows, and a deploy create --zip command. CVAT v2.81.0 added API endpoints to create and delete replica jobs. LanceDB v0.41.0 changed SentenceTransformerEmbeddings.trust_remote_code to default to False and updated the underlying Lance engine to 14.0.2.

Pricing moves

DeepSeek V4 Flash 0423 input price rose 80% to $0.09 per 1M tokens. DeepSeek V4 Pro 0423 input price increased 1% to $0.89. DeepSeek V4 Pro 0813 input price dropped 85% to $0.10. Z.AI adjusted three models: GLM 5.3 Flash input price fell 80% to $0.03, GLM 5.2 input price fell 67% to $0.01, and GLM 5.1 input price rose 44% to $1.40. Alibaba raised Qwen3.6 27B input price by 41% to $0.45. Moonshot AI raised Kimi K3 input price by 6% to $0.19.

▲ 80%DeepSeek V4 Flash 0423DeepSeek · input $0.05 → $0.09 per 1M tokens
▼ 67%GLM 5.2Z.AI · input $0.03 → $0.01 per 1M tokens
▲ 1%DeepSeek V4 Pro 0423DeepSeek · input $0.88 → $0.89 per 1M tokens
▲ 44%GLM 5.1Z.AI · input $0.97 → $1.4 per 1M tokens
▲ 41%Qwen3.6 27BAlibaba · input $0.32 → $0.45 per 1M tokens
▲ 6%Kimi K3Moonshot AI · input $0.18 → $0.19 per 1M tokens
▼ 85%DeepSeek V4 Pro 0813DeepSeek · input $0.66 → $0.1 per 1M tokens
▼ 80%GLM 5.3 FlashZ.AI · input $0.15 → $0.03 per 1M tokens

Emerging projects

Eight new projects entered the index today. andee (Kotlin, 3,215 stars) moves into a phone or Mac and uses any app without integrations or root. layerfs (Rust, 2,377 stars) gives every agent a complete, writable filesystem workspace over shared committed history. mcd-saver-strategist (Python, 1,545 stars) optimizes McDonald's meal combinations using coupons. vibe-astock (TypeScript, 879 stars) is a Chinese A-share stock analysis workbench. pgx (Rust, 498 stars) provides lightweight isolated Postgres databases for AI agents. ai-draw (Python, 398 stars) creates and edits drawings through conversation. respan-gateway (Rust, 352 stars) offers a single endpoint for every LLM vendor. durable-harness (TypeScript, 255 stars) builds durable coding-agent harnesses on a workflow SDK.

Research pick

A paper titled "DTCA: Denoising Token Credit Assignment for Reinforcement Learning in Diffusion Language Models" proposes a method to apply reinforcement learning to diffusion language models. These models generate text by iteratively denoising, which allows bidirectional context and parallel token refinement. The new technique assigns credit to individual tokens during denoising, improving the model's reasoning capabilities.

The bigger picture

Today's facts show a market in flux: major price swings from DeepSeek and Z.AI signal aggressive repositioning, while the surge in new projects points to growing interest in agent infrastructure, app control, and specialized tools. The research advance in diffusion language models suggests the next frontier for RL-based reasoning may be outside the transformer paradigm.

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