Why Claude Code Beats DeepSeek After the Price Hike: Cost‑Per‑Task Analysis

DeepSeek V4‑Pro pricing and benchmark

DeepSeek V4‑Pro launched with a 50 % off‑peak discount effective 2026‑08‑17 00:00 UTC+8. Under the same benchmark suite, DeepSeek V4‑Flash averaged $0.03 per task, Claude Fable 5 $3.15 per task, and GPT‑5.6 $1.86 per task.

Token price ≠ task price

Choosing a model solely by input‑ and output‑token rates ignores inference depth, cache hit rate, output length, and retry failures. A low token price can become more expensive if more reasoning steps or retries are needed; a higher token price may be cheaper when the model gets the job right on the first try.

Cost per Task aggregates token usage, inference steps, cache efficiency, output length, and failure retries into a single budget figure, which is more relevant for agents, customer‑service bots, code‑review tools, or data‑cleaning pipelines.

Peak‑off‑peak pricing strategy

Peak pricing (Beijing 09:00‑12:00, 14:00‑18:00) reaches 27 CNY per million tokens; off‑peak periods are half price. Moving batch generation, offline labeling, data cleaning, and testing to off‑peak slots can cut costs by up to 50 %. Real‑time online services remain subject to peak pricing.

Claude Code and Code80

Claude Pro ($20 / month) is intended for web users; API calls are billed per token and can cost hundreds of dollars for heavy usage. Code80 provides a proxy endpoint that uses a real Anthropic subscription, eliminating the need for an overseas credit card or stable network, and charges strictly by usage.

Installation and usage example:

# Install Claude Code
npm install -g @anthropic-ai/claude-code

# Run with Code80 endpoint
ANTHROPIC_BASE_URL=https://api.code80.vip claude

Other tools (Cursor, Cline, custom scripts) only need the base_url parameter changed to the Code80 endpoint.

For more information, see https://code.ai80.vip.