I'm on the $1.5/mo plan for 50 IPs.
pointed them at opencode running deepseek-v4-flash as the Hermes. been using it for real and it's actually solid, everyday stuff is no problem and it keeps up fine even when I pile on the work.
I'm on the $1.5/mo plan for 50 IPs.
More posts
Finally cracked how SuperGrok billing works.
Your weekly quota is fixed and shared across XAI, Build, and Chat. The difference comes down to multiplier rates and caching. XAI burns way more credits per call than Build. Also if you're routing Build through 9router like I was, you get zero cache hits. Only the official Build CLI actually caches. Someone tested it: the official Build CLI pulls at least 1B tokens per week on SuperGrok's grok-4.5. Build CLI went open source recently too, honestly pretty solid as a daily coding agent.
Vivix just shipped A1 and it's got something
. Full-duplex, full-body digital human. Siri-style turn-taking is half-duplex: you talk, then it talks. A1 runs full-duplex, listening while it replies, talking over you and cutting in like a real person. The full-body part is what got me. This thing walks, turns, crouches to grab objects, holds them up, swaps products mid-demo. First time I've seen that in a digital human. For live commerce it's basically a host that picks up your product, demos it, and changes outfits to match your brand, 24/7, no cuts. They ship a standard API, so you can drop it into anything. Point it at your product and it's a support agent. Drop it in a game and it's a companion. Put it on a stream and it's your host. The use cases pretty much write themselves.
Meituan shipped LongCat-2.5
. 1.6T params, ~48B active, native multimodal, aimed at long-horizon agents, trained start to finish on Chinese hardware. Images get quantized into multi-level discrete tokens, and that's where Visual Coding and 3D pick up. On the RL side, failed trajectories get turned back into verifiable training tasks. W4A4 inference, 43% faster prefill. Meituan's own evals put it between Claude Opus 4.8 and Opus 5.
SenseTime open-sourced SenseNova Loop-9B-Preview.
It stacks layers instead of parameters. The decoder has 32 layers, but layers 4 through 27 run twice with shared weights, so a forward pass does 56 layers of work while the checkpoint stays at 9B. Their own eval numbers: 71.38 on BIRD-Dev and 13.72 on AutomationBench, both above Qwen3.5-27B. Terminal Bench is still behind.
