DeepSeek: Free Chat, API Costs and Limits Explained
DeepSeek offers free chat and paid API access. We explain the difference between chat, API, and self-hosted models, plus how to calculate actual costs.
👁 116 reads
DeepSeek is both a company developing AI models and the name of the service where you can interact with them. The free DeepSeek chat, paid API, and self-hosted model represent different ways to use this technology.
The Official Chat Is Enough for Basic Conversations
We’re describing the producer’s access and documentation here, not the results of our own comparative model test. On the official DeepSeek website you’ll find a link to the free chat. The chat runs at chat.deepseek.com. We’d start checking out the service from this point, rather than searching for apps based only on a similar name.
For your first try, prepare a short, non-sensitive piece of material: a note, product description, or fragment of your own text. Write what you expect in your language and specify the language for the response. Fluent output alone doesn’t confirm factual accuracy. If the model expands an abbreviation differently than you intended, add its meaning and ask for a new response.
For someone starting out, a good task matters more than the variant name. In our guide what is a prompt we show how to separate source material from instructions. This principle also helps when you switch model providers.
Chat, API and Self-Hosted Models Have Different Costs
| Access Method | What It’s For | What to Check |
|---|---|---|
| Web chat | Manual conversation and work on individual tasks | Available features and limit messages |
| Producer API | Connecting the model to an application or automation | Model, input and output rates, limits |
| Self-hosted open model | Infrastructure control and processing method | License for specific release, hardware and maintenance |
Free chat access doesn’t mean free API. Downloading model weights doesn’t eliminate the cost of memory, computation, and implementation work. Meanwhile, a reseller service may have different rates than DeepSeek’s direct API. When comparing prices, we therefore record the provider as well, not just the model name.
DeepSeek documentation currently separates the identifiers deepseek-v4-flash, deepseek-v4-pro, and the experimental deepseek-v4-flash-vision-exp. This is the documentation status checked on September 9, 2026. We don’t transfer features of the image-handling variant to every model in this family. The name visible in your integration must match the function you need.
You Can Only Calculate Task Cost After Separating Tokens
We don’t give a single rate “for DeepSeek” because such a number hides the model, input, output, and billing conditions. Open the current price list from the producer’s documentation and record separately the cost of tokens sent and generated. If there’s a separate rate for cache hits, include it only in the portion of traffic it actually applies to.
We calculated a simple planning scenario without pretending to measure usage. We assume 240 tasks, each with 3,000 input tokens and 600 output tokens. This gives 720,000 input tokens and 144,000 output tokens. When you mark the prices per million as I and O, the cost is 0.72 × I + 0.144 × O. In this example, the generated text is one-sixth of the total token count, but it doesn’t have to be one-sixth of the bill.
This is a difference you can easily miss when focusing on advertised input cost. Add in retries after errors and additional application steps. If you need a comparison of model families, check the AI models section, but perform the calculation for your own task and one clearly defined provider.
We Compare DeepSeek and ChatGPT on the Same Data
We have no basis here to declare DeepSeek better than every ChatGPT model or vice versa. We suggest your own scorecard: alignment with source material, instruction execution, format correctness, and manual correction time. Each tool gets the same text and the same expected result. The model name and available tools should appear with the result.
In a data extraction task, establish the reference answer before running the test. If a note doesn’t contain a delivery date, the correct result is “no data.” A plausible-looking date that was added remains an error. The method of evaluating results is as important as the model benchmarks themselves.
Where Trust in the Answer Alone Fails
A long explanation doesn’t guarantee a correct conclusion. The chat may skip a caveat, add a detail, or confuse a product with a similar name. We treat the instruction “don’t make things up” as an instruction whose execution must be verified. For facts, ask for the fragment of material that supports the answer, then find it yourself.
Also don’t equate self-hosting a model with security for the entire application. A plugin, login system, and external tool can process data independently of the model. Before submitting company material, determine where the content will go and what rules apply in the chosen service. Use placeholder data for functional tests.
Google · Your sourcesQuantslant higher in your Google - one clickAdd to preferred sources →Start with One Task with a Verifiable Answer
Go through the official site, paste a short text, and ask for a table with information present in the material. In the instruction add “if the information isn’t there, write no data.” Compare each row with the original, correct the instruction, and repeat the test. Only after such verification should you move the task to API or larger-scale automation.
More on this from QUANTSLANT
guides·AI Agents: What They Do, Where They Fail, and How to Test ThemAn AI agent chooses its own steps and uses tools to complete tasks. See examples, error accumulation math, and how to test your first process.