Skip to content

Use case

The best API for chatbots and support

Chat costs grow faster than traffic, because each turn resends the whole conversation so far. A 20-turn conversation sends roughly ten times the tokens of a 2-turn one for the same number of replies — which is why per-conversation cost, not per-message cost, is the number worth modelling.

Every turn resends the history. Costs scale with the square of conversation length, not linearly with messages, and a support bot that averages 12 turns costs far more than one averaging 4 even at identical volume.

The two effective levers are cached input on the fixed system prompt and a summarisation or truncation policy on older turns. Choosing a marginally cheaper model is usually a smaller win than capping how much history each turn carries.

What actually matters here

Cached-input rate
The system prompt and persona are identical on every turn of every conversation — the clearest cache win of any workload.
Output rate
Unlike agents, chat produces substantial output. Here the output rate genuinely matters.
Latency
A human is waiting. Time to first token is felt directly, in a way it is not for a background job.

What it costs, at 50,000 conversations/month, ~8 turns each

Worked from this catalogue’s published rates at 600M input and 40M output tokens a month. Your figure will differ; the arithmetic will not.

ModelInputOutputPer monthAt official rates
DeepSeek V4 Flash36%$0.090/M$0.18/M$61.20$95.20
GLM-5.242%$0.82/M$2.55/M$594.00$1,016.00
Kimi K2.642%$0.55/M$2.30/M$422.00$730.00

Excludes cached-input savings, which on a repeated-prefix workload typically reduce the input column substantially. Cached rates are published per model on the pricing page.

Questions

Does truncating history hurt quality?
Less than most teams expect for support workloads, where the useful context is usually the last few turns plus a summary. It is worth measuring on your own transcripts, because the cost difference is large and the quality difference often is not.