AI News · September 7, 2026

Input prices fall for DeepSeek and Qwen as Optuna v5 lands

57,684 stars today77 new projects+3,433 archify

DeepSeek cut input prices on two models and Alibaba lowered Qwen input pricing. DeepSeek V4 Pro 0423 input fell 23% to $0.63 per 1M tokens. DeepSeek V4 Flash 0731 fell 29% to $0.05 per 1M tokens. Qwen3.5 397B A17B fell 29% to $0.39 per 1M tokens.

Across the tracked ecosystem, 57,684 GitHub stars were gained today across 10,000+ open-source AI projects. Seventy-seven new projects entered the index.

This week, 2,160 new projects joined the index, along with 100 new Hugging Face spaces. Three minor releases landed today: Optuna v5.0.0, Vespa v8.751.13 and AG2 v1.0.4.

Releases of the day

Optuna v5.0.0 brings the first major update to its default sampler, which affects hyperparameter optimization workflows for many projects. Vespa v8.751.13 is the command-line tool for Vespa.ai, usable on instances locally, remotely or in the cloud, with web service APIs still preferred in production. AG2 v1.0.4 aligns with Anthropic 1.x and OpenAI 3.x SDKs, expands TealTiger governance with two additional policies, and enables ACP servers to resume sessions. These are minor releases, but the sampler update and SDK alignment can change how developers integrate the tools.

Pricing moves

All three price moves today are input price reductions. DeepSeek V4 Pro 0423 dropped 23% to $0.63 per 1M tokens. DeepSeek V4 Flash 0731 dropped 29% to $0.05 per 1M tokens. Qwen3.5 397B A17B dropped 29% to $0.39 per 1M tokens. The cuts lower the cost of input-heavy workloads for developers using these models.

▼ 23%DeepSeek V4 Pro 0423DeepSeek · input $0.82 → $0.63 per 1M tokens
▼ 29%Qwen3.5 397B A17BAlibaba · input $0.55 → $0.39 per 1M tokens
▼ 29%DeepSeek V4 Flash 0731DeepSeek · input $0.07 → $0.05 per 1M tokens

Emerging projects

Today's index added 77 new projects, and the weekly total stands at 2,160. Among tracked projects, archify leads GitHub today with +3,433 stars.

Research pick

The Hugging Face research pick, Iris: Climbing to the Search Frontier, presents two search agents: Iris-mini at 35B-A3B scale and Iris-pro at 397B-A17B scale. The authors detail the data pipeline and training recipe used to build them. Tasks are reverse-constructed from the hyperlink structure of a web corpus.

The bigger picture

Today's price cuts on DeepSeek and Qwen input tokens lower the barrier for large-scale inference, while Optuna's sampler update points to continued maturation of core ML tooling. The steady inflow of new projects suggests the ecosystem remains active.

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