Who a DeepSeek chatbot is a good fit for
A DeepSeek chatbot makes the most sense when conversations are frequent, low-stakes and impersonal: product questions, opening hours, how-to answers from public documentation, content sites with a lot of curious readers. In those cases visitors rarely type anything personal, and a lower token price matters because volume adds up.
It makes less sense when visitors tend to share names, health details, account information or anything sector-regulated. For those agents, use another provider, and keep DeepSeek for the general ones. You can give different agents different providers, so this is a per-agent decision rather than an all-or-nothing one.
Read the data-location terms before you connect
As of writing, DeepSeek’s API is operated by a company based in China, and its privacy policy states that data is processed on servers there. Whether that is acceptable depends on your industry, your customers’ expectations and the laws that apply to you, for example data transfer rules in the EU and UK or sector rules.
- Read DeepSeek’s current privacy policy and terms before sending real visitor conversations.
- Disclose the provider in your own privacy policy.
- If visitors may share personal or sensitive information, consider another provider for that agent.
- Do not rely on this page for legal advice; ask your own adviser if you are unsure.
None of this means DeepSeek is unsuitable. It means you should choose it deliberately. Our Security and privacy page describes what SiteTrellis itself stores and sends.
DeepSeek models and estimated costs
Both DeepSeek rows are unverified. The model ids were confirmed from DeepSeek’s documentation, but the prices and context sizes come from third-party trackers and use a peak-hours rate to stay conservative. Confirm current rates, including any off-peak discount, on DeepSeek’s own pricing page.
| Model | Input, per 1M tokens | Output, per 1M tokens | Estimate per 1,000 conversations |
|---|---|---|---|
| DeepSeek V4.1 Flash * | $0.30 | $1.20 | $5.06 |
| DeepSeek V4 Pro * | $1.32 | $3.96 | $20.83 |
Estimates, not quotes: list prices from our model catalog in US dollars per million tokens, for 6 visitor messages per conversation, 1,500 tokens of prompt and knowledge sent every turn, and 180-token replies, with no prompt caching. Your provider's bill is authoritative. * Not yet confirmed against the provider's own pricing page.
On those catalog figures, DeepSeek V4.1 Flash comes to about $5.06 per 1,000 conversations and DeepSeek V4 Pro to about $20.83. Compare that with the small tiers on the Claude or OpenAI pages before concluding anything, and keep the estimates in perspective: at a few hundred conversations a month, the difference between providers is often small next to the platform plan. Use the calculator with your own numbers.
Check the tool-support marker in the dashboard if your agent captures leads or calls HTTP tools: DeepSeek models differ in how well they handle tool calls, and the playground will show you.
Set up a DeepSeek chatbot
Top up your balance
Sign in at platform.deepseek.com and open the top-up area. DeepSeek works from a prepaid balance, and an empty balance returns an insufficient balance error.
Create the key
Open API keys, create a new key named “SiteTrellis production” and copy it at once. It starts with
sk-and is shown only once.Add it to SiteTrellis
Under AI providers choose DeepSeek, name the key, paste it and confirm that the status becomes ok.
Pick the model
In the agent’s model section select the key and choose a model from the list.
Raise the step timeout a little
DeepSeek can be slow at busy times. Allow more seconds on this step (the builder allows up to 55), or put it behind a faster provider.
Test, then publish
Run real questions in the playground, including a few that need a tool, then publish.
The screens are covered in the Connect DeepSeek guide.
Cost control with a prepaid balance
Because the account is prepaid, the balance you load is your spending cap. Load a modest amount, watch the usage page for a week, and top up on purpose rather than through automatic top-ups. The flip side is that a zero balance makes the chat fail, so either keep an eye on it or add a second provider.
Inside SiteTrellis, a lower history window, shorter knowledge text and a monthly message cap on the agent all reduce tokens. The cost guide shows how conversation length drives the bill.
An example fallback chain for DeepSeek
| Step | Key and model | Why |
|---|---|---|
| 1 | DeepSeek key, DeepSeek V4.1 Flash, timeout about 15 seconds | Low list price for a public-FAQ agent that collects no personal data. |
| 2 | Anthropic key, Claude Haiku 5.5 | Catches a 402 empty balance, a 429 throttle, a 5xx busy reply or a slow peak-time response. |
If your agent may handle personal details, flip the order or leave DeepSeek out entirely. Remember the overall reply budget is about 50 seconds, so a long timeout on step one leaves less room for step two. Background in the fallback guide.
DeepSeek errors and honest limits
- 401: wrong or deleted key. Create a new one and rotate it in.
- 402 insufficient balance: top up, then press Test.
- 400 or 422: the model rejects a setting such as a temperature or tool format. Choose a model that supports your setup or reduce the tools.
- 429: DeepSeek throttles dynamically under load. Retry later or rely on a fallback.
- 500 or 503: busy service, more likely at peak times.
As with every agent here, knowledge is pasted text, replies are not streamed and answers can be wrong. See troubleshooting and the BYOK overview.