LongCat-2.0: a cheap MIT-licensed coder is topping OpenRouter
Meituan's 1.6T open MoE undercuts GPT-5.5 by up to 10x, beats it on SWE-bench, and trained on 50K Chinese chips — plus Claude Code's spend cap finally bites.
Copy markdown
Call it today, or self-host the weights
LongCat-2.0 is a 1.6-trillion-parameter MoE (~48B active per token) with a native 1M-token context, released by Meituan under a permissive MIT license. The weights are live on Hugging Face as INT8 safetensors, and it serves OpenAI- and Anthropic-compatible endpoints, so it drops into Claude Code or your existing agent stack with a base-URL swap.
$0.75 in, $2.95 out — and it hid at #1 as 'Owl Alpha'
On OpenRouter it runs $0.75/$2.95 per million tokens (launch promo $0.30/$1.20, cached reads free), undercutting GPT-5.5's $5/$30 and Claude Sonnet 5's $2/$10. It quietly became the platform's most-used coding model under the codename 'Owl Alpha' for two months before Meituan claimed it — near-frontier agentic coding at a fraction of the token cost.
It beats GPT-5.5 on SWE-bench Pro
LongCat-2.0 scores 59.5 on SWE-bench Pro (vs GPT-5.5's 58.6) and 73.2 on FORTE, tying Claude Opus 4.6 — coding-agent benchmarks, not a marketing chart. If you've been paying frontier prices for agentic edits, this is a cheaper baseline worth A/B-testing this week.
Trained on 50,000 Chinese ASICs, no Nvidia
Meituan says the model was trained end-to-end on roughly 50,000 domestic AI ASICs across 35T+ tokens. The builder takeaway isn't geopolitics — it's supply: cheap, open, capable models are now shipping off a decoupled chip stack, so expect more MIT-licensed, frontier-adjacent weights, and faster.
Claude Code's spend cap finally halts background agents
Claude Code 2.1.217 (Jul 21) fixes `--max-budget-usd` so it actually stops background subagents once the cap is hit — before, they kept spending past it. The same release adds :emoji: shortcode autocomplete and patches a memory leak that kept full, untruncated MCP tool outputs in memory.