Perplexity's open 27B decision model beats Jev; API is near-free
An approve/reject endpoint at $0.04/M input, free output — plus Comfy's agent that builds ComfyUI graphs and fal's H3 person-swap for video.
Copy markdown
A 27B that beats Jev, with free output
Perplexity's pplx-decider-v1-27b scores 85.71% across 11 decision benchmarks to Jev's 84.51%, and the hosted Decisions API bills $0.04 per million input tokens with output free — a near-zero way to wire approve/reject or routing calls into an agent instead of paying full-model rates to classify. The weights (a ~26B Qwen3.8 fine-tune, ~49 GiB in BF16) are on Hugging Face under Apache-2.0, multimodal over text and images at ~250K context, so you can self-host the same model.
Comfy Agent builds the ComfyUI graph for you
Comfy Agent is live in Comfy Cloud: describe a task in plain language and it plans, builds, edits, runs, and debugs the workflow on the canvas with you, powered by Anthropic frontier models and running up to five chats in parallel. For anyone who bounced off ComfyUI's node spaghetti, it turns graph-wiring into a conversation — local-GPU support is listed as coming soon.
fal's H3 Max drops a person into any clip
fal's new H3 Max 'Recast' (video-to-video) swaps a person into an existing clip while preserving the original motion, gestures, and timing, taking up to four reference photos on 5–30s videos. It runs $0.30/sec at 768p and $0.45/sec at 1080p — cheap enough to prototype a face-consistent character across shots in an afternoon.