Three $10 plans compared.

2026-11-03 13:13

Three $10 plans compared. Last post I mentioned broker sites for flagship models, where you can get pricing down to about a tenth of list. Still too rich for subagents or anything you run all day. For normal use, unless you're doing development or research, I keep landing on the ~30 tier of Artificial Analysis: deepseek-v4.1-flash. Step up to ~40, which is Kimi K3.1 Pro, and the price goes 10x. For DS itself, these three subscriptions are the cheapest route today. OpenDesign Go says $120 of usage, OpenCode Go and Command Code GOAT say $100. Those aren't the same number though. OpenDesign's $120 is one pool shared by 12 models, with a $10 cap per 5 hours. The other two give DS its own allowance and $20 per 5 hours.

SpaceX quietly trimmed Grok Bot's usage limits this weekend.

A lot of users woke up to find their weekly quota had shot from about 50% to a flat 100%, and they hadn't even touched it. Musk confirmed on X it wasn't a glitch. His words: serving Grok Bot "was unsustainably high relative to revenue." So the thing really does cost a fortune to run. That said, I'm seeing the limits roll back to the earlier version. Yesterday it showed 50% used; today I'm at 30%.

2026-11-03

OpenRouter, Carpooling edition.

wokey is a marketplace for LLM quota. It pairs people with unused plan credits to people who need more, and takes a 20% cut. Sellers keep 80% of what their credits earn; buyers save up to 90% of the official API price. I use it mostly for flagship models. The models check out as the real thing, cost less than going direct, and I skip the subscription jugglingling. If all you call are cheap models, a router still beats it. 🔗 https://wokey.ai/r/GKT1C78R

2026-11-03

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.

2026-11-02

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.

2026-11-02