DeepSeek
DeepSeek provides high-performance AI models with a focus on reasoning and code generation at highly competitive prices. DeepSeek R1 is one of the leading open-weight reasoning models, and DeepSeek V3.2 is a strong general-purpose model.
Getting an API Key
- Visit platform.deepseek.com/api_keys
- Sign in or create an account
- Generate a new API key (starts with
sk-...) - Paste the key into AI Supreme Council under Settings > AI Model > DeepSeek
API keys are stored locally in your browser (localStorage) and are never included in shared bot URLs.
Supported Models
| Model | Context Window | Max Output | Input Price | Output Price | Capabilities |
|---|---|---|---|---|---|
| DeepSeek V3.2 | 128K | 8K | $0.28/MTok | $0.42/MTok | Tools, code, streaming |
| DeepSeek V3.2 Reasoner | 128K | 64K | $0.28/MTok | $0.42/MTok | Reasoning, code, streaming |
| DeepSeek R1 0528 | 65K | 8K | $0.55/MTok | $2.19/MTok | Reasoning, code, streaming |
Prices are per million tokens (MTok). Cached input available at reduced rates.
- DeepSeek V3.2 is the general-purpose default -- fast, affordable, and good at code. Best for everyday tasks where you do not need step-by-step reasoning.
- DeepSeek V3.2 Reasoner adds reasoning to the V3.2 base with a larger 64K output window. Great for complex analysis that benefits from chain-of-thought.
- DeepSeek R1 0528 is the dedicated reasoning model -- best for math, logic, theorem proving, and multi-step problem solving.
Reasoning Support
DeepSeek's reasoning models (V3.2 Reasoner and R1 0528) show their step-by-step thinking process before delivering a final answer. The reasoning content appears in a collapsible block above the response, similar to Claude's extended thinking and Grok's reasoning output.
Since DeepSeek uses the OpenAI-compatible API format, reasoning is controlled via the Reasoning Effort setting in the bot configuration panel:
| Setting | Behavior |
|---|---|
low | Minimal reasoning -- faster responses |
medium | Balanced reasoning depth |
high | Deep reasoning -- best quality for hard problems |
The reasoning content streams in real time via SSE, so you can watch the model's thought process as it works through a problem.
DeepSeek R1 is also available for free through OpenRouter using the model ID deepseek/deepseek-r1-0528:free. If you want to try DeepSeek's reasoning without setting up a DeepSeek account or spending money, use OpenRouter with a free key.
Competitive Pricing
DeepSeek is one of the most affordable providers available:
| Comparison | DeepSeek V3.2 | GPT-4o Mini | Claude Haiku 4.5 |
|---|---|---|---|
| Input | $0.28/MTok | $0.15/MTok | $0.80/MTok |
| Output | $0.42/MTok | $0.60/MTok | $4.00/MTok |
DeepSeek V3.2 offers output pricing lower than GPT-4o Mini while providing strong performance on coding and reasoning tasks. The V3.2 Reasoner offers 64K max output tokens at the same price as the base V3.2 model.
Tool/Function Calling
DeepSeek V3.2 supports tool calling via the OpenAI-compatible format. Define tools in your bot configuration, and DeepSeek will invoke them as structured function calls. The R1 reasoning model does not support tool calling -- use V3.2 or V3.2 Reasoner for tool-use scenarios.
OpenAI-Compatible API
DeepSeek uses a fully OpenAI-compatible API, which means:
- Standard
POST /v1/chat/completionsendpoint atapi.deepseek.com/v1/chat/completions - Bearer token authentication
- SSE streaming
- Tool/function calling support (V3.2)
- Same request/response format as OpenAI
The platform registers DeepSeek with the shared openaiCompatible() streaming factory, so no special configuration is needed. All standard features (streaming, reasoning content parsing, tool call normalization) work automatically.
Configuration
When creating a bot profile, select DeepSeek as the provider and choose your preferred model. You can set a per-bot API key in the bot configuration panel to override the global key.
Registry Entry
The model registry defines DeepSeek with the following metadata:
| Field | Value |
|---|---|
| Key placeholder | sk-... |
| Key URL | platform.deepseek.com/api_keys |
| API format | OpenAI-compatible |
| Auth type | Bearer token |
| Features | Reasoning, streaming |
Use in Councils
DeepSeek models are strong council members, especially for reasoning-heavy deliberation:
| Council Role | Recommended Model | Why |
|---|---|---|
| Reasoning specialist | DeepSeek R1 0528 | Deep chain-of-thought for math and logic |
| Code reviewer | DeepSeek V3.2 | Fast, affordable code analysis with tool support |
| Budget member | DeepSeek V3.2 Reasoner | 64K output at V3.2 pricing for detailed analysis |
Pair DeepSeek with vision-capable models from other providers (Claude, Gemini, GPT-4o) for councils that need both reasoning depth and multimodal input.
Tips for Best Results
- Use V3.2 for speed and cost. It is the most responsive model for general tasks at a fraction of the cost of larger models.
- Use R1 for math and logic. DeepSeek R1 excels at mathematical reasoning, theorem proving, and complex analytical tasks.
- Try it free via OpenRouter first. DeepSeek R1 is available for free on OpenRouter, so you can evaluate it before committing to a DeepSeek API key.
- Combine with councils. DeepSeek R1 is a strong reasoning member in multi-model councils, complementing vision-capable models from other providers.
- Use V3.2 Reasoner for long output. With 64K max output tokens, V3.2 Reasoner is ideal for tasks that need detailed, lengthy responses with reasoning chains.
- Watch the context limit. R1 0528 has a 65K context window, which is smaller than V3.2's 128K. For long documents, prefer V3.2 or V3.2 Reasoner.