Claude Code "limit reached" on a paid plan, rarely a billing bug:
• Two caps; the weekly one is invisible till it fires at 2% session
• Opus default burns several times Sonnet/turn
• MCP servers eat 33% of context at startup
7 fixes → https://t.co/GdYULsZzTX
Doubao 2.1 Series — 4 models, now live on Ofox
🔸 Doubao-Seed-2.1 Pro — deep-thinking flagship, rivals GPT-5.5 / Opus 4.7
🔸 Doubao-Seed-2.1 Turbo — scaled production at half the price
🔸 Doubao-Seed-Character — roleplay & lifelike dialogue
🔸 Doubao-Seed-Evolving — Agent/Coding, weekly updates
https://t.co/KIsmamr0EZ
"Claude in Slack" is dead. Claude Tag replaces it June 23:
• Shared @Claude per channel, on Opus 4.8
• Ambient mode posts unprompted, chases stale threads
• Spend caps decline over-cap work, no cutoff
• Old app retired Aug 3
Setup + gotchas → https://t.co/gKwJ2Xgar3
GLM 5.2 (753B, MIT) runs on a Mac you might already own:
• 2-bit UD-IQ2_M fits a 256GB Mac Studio, ~240GB, 3-9 tok/s
• 4-bit wants 512GB
• Under 256GB? No quant saves you — go hosted
• Memory-bound: a Mac beats a 4090 box
Quant picks + flags → https://t.co/3Pgrsx5QQP
GLM-5.2 vs GPT-5.5, same coding workload, same key:
• $2.40 vs $13.33 per 1M tokens — 5.56x gap
• 6.82x cheaper on output tokens
• 100K req/day: $21.6K vs $120K/mo
• Caching widens the gap, not closes it
Per-token math → https://t.co/CXNKovUc1k
GLM-5.2 vs GPT-5.5, same coding workload, same endpoint:
• $2.40 vs $13.33 per 1M tokens blended — 5.56x gap
• 6.82x cheaper on pure output tokens
• 100K req/day: $21.6K/mo vs $120K/mo
• Caching widens the gap, not closes it
• Both on one OpenAI-compatible key
Full per-token math → https://t.co/6bFfD3fGDq
GLM 5.2's MIT weights are free. The hardware isn't.
• 753B params: 8x H200 (FP8) or 4x H100 (GGUF)
• Beats Opus 4.8 on Terminal-Bench 2.1 (82.7 vs 78.9)
• Self-host breaks even past 10k prompts/day vs $30/mo hosted
Setup + cost math → https://t.co/KuZg5YXija
One Codex agent loop can drain a $20 Plus weekly cap in 3 hours:
• 1 heavy run ≈ 50 credits; weekly budget ~250-300 → 6 runs gone
• Bug #19215: /status reads 50%+ but Codex still rejects requests
Fix: point OPENAI_BASE_URL at a metered API → https://t.co/QnawhzC4vl
MiniMax M3 costs 1/10th of Opus 4.8 — its "beats GPT-5.5" line skips the newer flagship:
• 59% vs 69.2% SWE-Bench Pro (M3 benched vs old Opus 4.7)
• $0.60/$2.40 vs $5/$25 per M tokens
• 5 devs, default routing: $5.9K vs $55K/mo
Full breakdown → https://t.co/Pe475WAjDa
DeepSeek V4 Pro's "$0.28/M" is wrong twice over:
• $0.28 is Flash output — Pro is $0.87/M
• Cache miss $0.435 vs hit $0.003625 = 120x gap
• Our refactor: V4 Pro 1,556 tokens vs GPT-5.5 capped at 8,192
Verbosity, not sticker, sets the bill → https://t.co/kTUi3icEdA
Claude Fable 5 just dropped — Anthropic's most capable public model.
1M context window. 128K max output. Deep thinking always on.
Built for long-horizon agentic work. Now live on OFox.
/clear wipes your chat. The broken config? Untouched.
Claude Code v2.1.169's --safe-mode kills all 5 in one flag:
• CLAUDE.md (project + parent + global)
• plugins + user skills
• every hook
• all MCP servers
Auth + base URL stay live → https://t.co/b0bt6yAtWT