How to Get a DeepSeek API Key in 2026

What Is the DeepSeek API?

The DeepSeek API lets your app send requests to hosted DeepSeek models and receive responses in code. You can use it to turn a bug report into an issue ticket or add a coding assistant to your own product. The API uses a key and bills model usage by token. Developers who already use the OpenAI SDK can adapt their client using the DeepSeek API quickstart.

DeepSeek began in Hangzhou in 2023. Its founder, Liang Wenfeng, also founded the quantitative hedge fund High-Flyer, which backed the AI startup. DeepSeek released V3, a general-purpose language model, in December 2024. R1 followed in January 2025. The release of R1’s model weights and its focus on reasoning helped bring DeepSeek to a wider audience. Read the Congressional Research Service’s overview of DeepSeek and DeepSeek’s R1 release notes.

If you’ve seen R1 in an older tutorial, check the model ID before copying its code. The example below starts with DeepSeek V4.1 Flash, using the deepseek-flash model ID on GPT Proto. See the guide to DeepSeek Flash model IDs if you’re migrating an older integration.

How to Get a DeepSeek API Key on GPT Proto

1. Create a GPT Proto account and add balance. Check the Flash model page for current input, cached-input, and output rates before topping up.

2. Open Quick Start. On the DeepSeek Flash model page, click Try this model to open the setup panel. Choose Create API Key, or use Select Key if you already have a GPT Proto key.

3. Choose how to connect. For an existing project, click Copy Markdown for AI and paste the guide into your coding assistant. Ask it to add a server-side call using the model ID and endpoint in the guide. Set GPTPROTO_API_KEY in your local environment or server secret store and keep the value out of your code and assistant chat.

Make Your First DeepSeek API Call

Once GPTPROTO_API_KEY is set in your terminal, send a short request. This example uses the Chat Completions endpoint shown in the DeepSeek Flash Quick Start.

curl --fail-with-body --silent --show-error \
  https://gptproto.com/v1/chat/completions \
  -H "Authorization: Bearer ${GPTPROTO_API_KEY}" \
  -H 'Content-Type: application/json' \
  --data '{
    "model": "deepseek-flash",
    "messages": [{"role": "user", "content": "Reply with API connected."}]
  }'

Look for the reply in choices[0].message.content. If the response includes a usage object, check it before estimating the cost of a larger request.

Turn a Bug Report into a Ticket

If you maintain a small SaaS app, try Flash on a support message before wiring it into your issue tracker. Replace the test message with this prompt and sample report.

Turn this report into a JSON object with summary, reproduction_steps, acceptance_criteria, and missing_details. Base the steps on the report. Put any missing environment details or prerequisites in missing_details.

Sample report: On Safari, I uploaded a CSV and clicked Import. The spinner disappeared and the page showed zero imported rows. The same file worked in Chrome.

Check that the ticket keeps the Safari and Chrome results separate. The acceptance criteria should describe what a successful import would look like. The browser version and CSV contents belong in missing_details. If you automate this, validate the JSON and route the draft to a human reviewer before creating the ticket.

Why Use GPT Proto for the DeepSeek API?

Pay 20% less than DeepSeek’s published peak rates for Flash. GPT Proto’s listed rates are $0.24 per million uncached input tokens, $0.96 per million output tokens, and $0.0048 per million cached input tokens. For a job you can schedule, check the applicable off-peak rates before deciding when and where to run it. The current rates are on the DeepSeek Flash model page.

Keep one key and balance as you test more models. Start with Flash for ticket triage. If the trickier reports need more reasoning, test V4 Pro against the same inputs with the same GPT Proto credentials. The GPT Proto DeepSeek model catalog lists the available routes and their prices.

Give your app a fallback you can test. You can configure a second supported model in your app for timeouts or upstream errors. Run the same input through both models and check that both responses meet your app’s requirements before relying on the fallback. If you hit a setup or billing issue, use the console support chat listed in the GPT Proto install guide.

How to Connect DeepSeek to Codex with a GPT Proto API Key

Use the ChatGPT (Codex) installer when you want to configure Codex with your GPT Proto key. Use Copy Markdown for AI when you want Codex to add a DeepSeek API call to your app’s code.

1. Open Quick Start on the DeepSeek Flash model page.

2. In Let AI finish the integration, choose ChatGPT (Codex).

3. Select or create a GPT Proto API key, choose your operating system, and follow the on-screen steps in the GPT Proto installation guide.

4. Open Codex and confirm the active provider and model. Start with a read-only task such as “Explain what this test covers and suggest one missing case.” If your workflow depends on tool calls, try a second task that reads a file and uses its contents in a follow-up.

A single reply confirms basic connectivity. If you plan to use Codex for longer sessions, test a multi-step task too. One developer reported an HTTP 400 during tool-call continuations in a specific DeepSeek Responses setup. That report covers a direct DeepSeek configuration and is documented in the Codex issue tracker.

Key DeepSeek API Models Available on GPT Proto

Start with DeepSeek V4.1 Flash for a new integration. Evaluate V4 Pro if difficult text tasks regularly need a second pass, or V3.2 if you run a large volume of short text tasks with predictable outputs. The other routes are useful when you have an existing integration or want a baseline for testing. All seven are listed in the GPT Proto DeepSeek model catalog.

These are GPT Proto catalog routes. On DeepSeek’s direct API, the older V4 Flash and Vision Exp IDs now route to V4.1 Flash. Check the ID in the model page’s Quick Start whenever you change providers or migrate an existing integration. See the official model listing and the DeepSeek Flash model ID guide.

How to Choose the Right DeepSeek Model

If your input includes screenshots or charts, start with V4.1 Flash. In the ticket example, send the error screenshot with the support message and check that the response captures the visible error. If you already use Vision Exp, run both on the same images before switching.

If your input is mostly short, repetitive text, compare V3.2 with Flash on a batch of real tickets. Check how many results pass your JSON schema, use the right labels, and avoid inventing details. Compare the cost per usable ticket, including retries.

If a bad answer is expensive to fix, try Flash on the routine cases and test V4 Pro on cases Flash misses. For a code fix, use passing tests and the number of retries as your criteria. Route the difficult cases to Pro if it improves those results enough to justify the extra cost.

To switch models in the cURL example, replace the value of model with the GPT Proto model ID above. The GPT Proto key stays the same. If you switch between text-only and image-capable models, make sure the request input matches what the new model accepts.

If speed matters, measure how long the user waits for the first chunk and how long the full task takes. Fast token generation can still leave a user waiting on a long reasoning step. Both metrics appear in the independent benchmarks for DeepSeek’s direct API.

For a small coding workflow, see the DeepSeek Flash and Claude Haiku 5.5 comparison. For a wider evaluation, compare DeepSeek Flash with Kimi K3 for coding agents or DeepSeek Flash with GLM 5.3 Flash for code and cost.

DeepSeek API Pricing per Million Tokens

Rates below are in USD per million tokens and were checked on November 4, 2026. GPT Proto rows show the listed rates. For DeepSeek’s direct API, peak hours are 01:00–04:00 and 06:00–10:00 UTC on weekdays, excluding Chinese public holidays.

For current prices, see the GPT Proto DeepSeek catalog and the official DeepSeek API rate card.

Suppose your ticket queue uses 1 million uncached input tokens and 250,000 output tokens. At the listed GPT Proto Flash rate, that comes to $0.48 in model usage. The same token mix costs $0.60 at DeepSeek’s official peak rates or $0.30 at its off-peak rates. These are estimates for that token mix. When budgeting for an agent, include tokens from extra turns and retries, plus any separately billed tools your setup uses.