Pro drops to 20 downloads a month, and it applies to songs you already made.
Suno is introducing download limits starting September 3. Free accounts get 7 lifetime downloads, Pro gets 20 per month, Premier gets 60. Streaming and sharing stay unlimited, and paid subscribers keep commercial rights. Existing songs remain playable, but downloads count against the quota, even for old tracks.
AI-generated micro-dramas now dominate Chinese platforms: 89 of the top 100 animated Douyin videos in May were AI, and 95% of the 128,000 micro-dramas produced in Q1 2026 were AI, per FT and NHK data. Production economics explain the shift: Wuhan-based Manghe creates AI dramas at one-third the cost and one-fifth the time of live-action with a 6-8 person team, while Zeelin claims 10% of live-action costs and produced 8,000 minutes in May alone. Background actors near Hengdian report lost work as the trend accelerates.
Google Flow's Gemini Omni 1.1 Flash update adds first and last frame interpolation for precise camera moves and looping, scene extension that analyzes up to 10 seconds of source footage, and video references up to 3 seconds for character consistency. A new 360p draft mode renders up to 60% faster at one-third the cost of 720p, letting creators iterate before committing to 1080p or 4K. The update is incremental but practical for shipping finished AI video.
A single watermark or credential check is insufficient for AI-media provenance. The recommended trail combines SynthID watermarks, C2PA Content Credentials, and a production log, since each answers a different question. Limits: SynthID non-detection doesn't rule out other AI systems, C2PA validates structure not truth, and embedded manifests can be stripped; the practical workflow uses a four-folder structure, naming conventions, and hash tracking that survives platform recompression.
One analysis argues YouTube's ad-based model rewards attention over quality, which is why it won't crack down on AI slop as aggressively as Spotify or Apple Music. The piece points to YouTube's new policy of counting a view at playback start, which incentivizes thumbnail-click optimization, exactly what AI slop excels at. YouTube does enforce its Inauthentic Content policy, but the author frames it as after-the-fact enforcement rather than prevention.
Fizgig v5.0.0 enables full fine-tuning of MiniMax H3 and Krea 2 base models on consumer GPUs with as little as 16GB VRAM, using a rotating training window and 4-bit quantization. Measured VRAM peaks stay under 12.3GB for H3, and a built-in utility converts fine-tuned checkpoints back to shareable LoRAs.
Sopro v2 Turbo is a 120M-parameter open-source TTS model that clones a voice from 5-20 seconds of audio and streams with ~300ms time-to-first-audio on a laptop CPU. It runs in the browser via ONNX and supports English, Portuguese, French, and German. No watermarking is included, so the developer asks users to avoid impersonation.
FLUX.3 and MiniMax H3 now offer 2K resolution with native audio and open weights, making high-fidelity generation more accessible than Seedance's closed system. The comparison tested surfing and dialogue scenes, with the host walking through AI Arena leaderboard data and each model's architecture. Seedance still holds advantages in some areas, but the open-weight options are now cost-efficient alternatives for creators and developers.