Claude Opus 5 is kinda insane

2026-07-25 05:27

Claude Opus 5 is kinda insane Near-Fable-5 level intelligence at half the price. 43.3% on Frontier-Bench — more than 2x Opus 4.8, and cheaper. On CursorBench (high), it's within 0.5% of Fable 5. And it's getting more AGI by the day — takes initiative like a real person, reflects and verifies its own work. Other models just patch bugs. This one digs out the root cause. For long tasks it takes over your dev environment and runs monitoring itself — only pings you when it genuinely can't decide.

grok-4.6 in 2 weeks, 4.7 in 4

. Musk says they're gunning for Kimi K3.

2026-07-26

Unlimited grok-4.5 + deepseek-v4-flashLately I've seen tons of

relay services offering grok-4.5 as low as 0.1 yuan per 1M tokens. Dug in — the trick is dead simple.At its core it's an account pool polling free grok accounts for their quota. No public number on free grok-4.5 limits, but what I found is ~1M tokens/day. In practice on pi it felt like less.Grok accounts don't currently care about IP, but I'd still route through a proxy pool — one account, one IP, safer. Those same IPs can also poll free models on opencode and mimo code (deepseek-v4-flash, mimo-v2.5-pro).I wired up a 20-IP pool. opencode + mimo code free models easily cover a general agent's daily tasks.For the account/proxy pool tooling, use 9router — it lets you bundle different models into one custom route. E.g. free models from different providers under different names, merged into a single "free" endpoint.Its token-saver is also great: cuts token use across tools, language, and code.9router seems to have a bug. Over public internet it fails constantly; local deploy or Tailscale works rock solid.IPs: webshare. Free tier gives 10 IPs, 1GB/month — more than enough if you budget it.9router: github.com/decolua/9routerFree proxies: webshare.io/?referral_code=pauzhzfueq8v

2026-07-25

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