Subscription math: Opus on Max costs ~4x DeepSeek tokens but wins on complexity, GPU payback ~10.8 months
teortaxesTex · x · 2026-10-30
teortaxesTex runs subscription economics with a correction from ChatGPT Pro: a given workload consumes 1/4 of the $200 ChatGPT Pro monthly capacity, and Claude Opus 5.5 tokens on Max cost 4x more than DeepSeek V4.1F's OpenCode tokens — yet Opus completes tasks with far fewer tokens (exponentially fewer on hard tasks), making it the better deal past some complexity threshold. At open-market GPU prices, serving this way would pay back GPUs in 10.8 months for DeepSeek, though realistically years given actual costs; Liang estimates 5-year Nvidia GPU depreciation. The author also disputes the "Chinese price wars force distillation" narrative and speculates Anthropic's Max-subscription economics are similarly involuted.
Related event: Debate rages over Chinese model pricing and distillation shortcuts(2 posts)→
More from Infra
- Pika H3 Flex ported to Mac via MLX: 5s video in ~5 minutes on M4 Max — Opening-Ad5541 · 2026-10-30
- FlashAttention-2 v2.9.2 Windows wheels released for PyTorch 2.15.0+cu134 — Zestyclose_Bake3680 · 2026-10-30
- AgentX benchmark: lossless-CSF hits 11.4M KV cache tokens, 8.7x DeepSeek V4.1 Flash baseline — TheZachMueller · 2026-10-30
- 'Girls only want one thing': an $496,639 8x NVIDIA B300 deep learning server — Kyrannio · 2026-10-30
- SAM3 now runs offline on your GPU: ONNX export hits 0.37s on M5, 0.08s on RTX 3080 Ti — CSProfKGD · 2026-10-30
- Same model, different hardware: dev finds A100s vs B300s produce different outputs — Kyrannio · 2026-10-30