LLM API monthly cost scenarios

1. Estimation method (shared by all three scenarios)

Scenario 1: Chatbot / online customer service

Scenario assumptions: each turn carries the system prompt and recent history (about 1,200 tokens), with short replies (about 300 tokens); customer-service apps are call-heavy, so 2,000 calls/day. The system prompt is stable, so the cache hit rate is 40%. Cost structure: call count is the amplifier — lightweight low-price tiers win.

Assumptions: 2000 calls/day · 1200 input / 300 output tokens per call · 40% cache hit rate · off-peak rates · 30-day month. Monthly cost = calls × 30 × (input ÷ 1,000,000 × effective input price + output ÷ 1,000,000 × output price); tiered models pick the tier by the 1200-token input, and tiers with unpublished rates are honestly marked "cannot estimate".

Scenario 2: RAG knowledge-base Q&A

Scenario assumptions: retrieved knowledge chunks are injected with the prompt, so each call inputs about 10,000 tokens; answers are short (about 500 tokens); 300 Q&A/day. The system prompt and knowledge-base prefix are stable, so the cache hit rate is 60%. Cost structure: input dominates — the cache price and hit rate are the decisive variables.

Assumptions: 300 calls/day · 10000 input / 500 output tokens per call · 60% cache hit rate · off-peak rates · 30-day month. Monthly cost = calls × 30 × (input ÷ 1,000,000 × effective input price + output ÷ 1,000,000 × output price); tiered models pick the tier by the 10000-token input, and tiers with unpublished rates are honestly marked "cannot estimate".

Scenario 3: Coding agent

Scenario assumptions: each task carries code context and tool output (about 6,000 tokens) and produces a patch plus explanation (about 1,500 tokens); 100 tasks/day. Context varies a lot between tasks, so the cache hit rate is conservatively 20%. Cost structure: output carries the weight — output unit prices separate the field.

Assumptions: 100 calls/day · 6000 input / 1500 output tokens per call · 20% cache hit rate · off-peak rates · 30-day month. Monthly cost = calls × 30 × (input ÷ 1,000,000 × effective input price + output ÷ 1,000,000 × output price); tiered models pick the tier by the 6000-token input, and tiers with unpublished rates are honestly marked "cannot estimate".

2. Comparing scenarios (factual observations)

3. Sensitivity example: how cache hit rate changes the RAG bill

Using scenario 2 (RAG, 300 calls/day, 10,000 input / 500 output tokens, off-peak rates), hold every other assumption fixed and vary only the cache hit rate to see how each model's estimate moves on the same basis. The table is generated by the build script with the same formula — it is an illustration of the order of magnitude of this single variable, not an industry survey or user statistics.

Reading the table: the lower a model's cache price (e.g. DeepSeek V4.1 Flash off-peak cache price ¥0.02/1M), the bigger the drop as hit rate rises. Models without a published cache price (e.g. MiniMax M3) are billed at the standard input price, so hit rate changes nothing — which is exactly why "check whether the vendor publishes a cache price" matters. USD models keep their $ list price and are converted at 1 USD = 6.718 CNY, CFETS central parity rate 2026-10-23.

4. Data sources and limitations

5. Next steps