Novita AI hikes Deepseek V4 Flash 0731 pricing and prunes its media catalog

Novita AI’s serverless rate card saw its sharpest single-SKU repricing since the July 2026 Flux.1 Kontext Pro 10x jump. Deepseek V4 Flash 0731 — a dated variant sitting alongside the unchanged base “Deepseek V4 Flash” model — rose from $0.14/M input and $0.28/M output to $0.44/M input and $1.32/M output (input +214%, output +371%), while its cache-read rate held at $0.028/M. A new flagship dated SKU, DeepSeek V4 Pro 0813, was added at $1.32/M input ($0.132/M cache read) and $3.96/M output.

On the GPU side, the self-serve /en/gpus lineup swapped RTX 4090 24GB “High frequency” ($0.69/hr on-demand, $0.35 spot) for the return of RTX 6000 Ada 48GB ($0.77/hr on-demand, $0.39 spot) — the same GPU that was pulled from this exact page on 2026-08-04, back three weeks later at an unchanged rate. Dedicated-endpoint, bare-metal, and Agent Sandbox pricing were confirmed unchanged via a re-capture of the Dedicated Endpoints tab.

The model catalog also contracted sharply on the media side: the Image category fell from 14 to 5 listed SKUs as Z Image Turbo, Seedream 4.0, Image Eraser, Image Remove Background, Image Upscaler, and the FLUX 2 Dev/Flex/Pro family were all delisted, leaving only the Flux.1 Kontext and Qwen-Image lines. The Video category dropped Veo 3.1, the entire Vidu Q2/Q3 lineup, older Kling versions (V1.6, V2.5 Turbo, V2.6 Pro), Wan 2.1/2.2, and Seedance 1.5 Pro, consolidating around the newer Kling v3.0 and Wan 2.5–2.7 families. Separately, the /en/models catalog counter (All Models tab) now reads 144, versus the 174 figure this page logged for the 2026-08-14 capture — a discrepancy this capture could not reconcile and flags for the next research pass.