Current as of March 2026. DeepSeek V3.1 upgrades V3 with a much bigger context window — 164K tokens for both input and output — while staying firmly in the budget tier at $0.20/M input. That output window in particular is unusual at this price.
Specs
| Provider | DeepSeek |
| Input cost | $0.20 / M tokens |
| Output cost | $0.80 / M tokens |
| Context window | 164K tokens |
| Max output | 164K tokens |
| Parameters | N/A |
| Features | function_calling, reasoning |
What it’s good at
Price at this context size
$0.20/M input is roughly 10x cheaper than GPT-4o for similar workloads. The 164K output limit is the headline — most models cap output at 8K or 16K, which means you’re constantly splitting long generations across multiple calls. V3.1 sidesteps that problem.
Reasoning on a budget
Built-in reasoning features are meaningful here. This isn’t a glorified summarization model — it handles multi-step logic and complex coding tasks well enough that I’d reach for it over V3 whenever the task involves actual reasoning rather than just text transformation.
Output volume
164K max output is a genuine differentiator. Generate a fully refactored module, a long spec, an entire test suite — in one pass.
Where it falls short
Latency
DeepSeek’s servers run slow compared to US-based providers, especially at peak hours. It’s unpredictable enough that you need retry logic on any production agent.
Content filtering
The safety filters over-trigger on anything touching sensitive geopolitical topics. Mostly a non-issue for developer tooling, but worth knowing if your use case brushes up against that territory.
Best use cases with OpenClaw
- Large-scale code refactoring — Feed it a large file, get the whole thing back refactored. The output window makes single-pass generation practical.
- High-cycle agents on a budget — Running thousands of agent loops on $0.20/M input is financially sustainable in a way that GPT-4o simply isn’t.
Not ideal for
- Real-time UIs — The latency variance makes it unsuitable for anything where a human is waiting for a response.
- Creative writing with sensitive content — The filters are stricter than Claude or GPT, and they fire on things that probably shouldn’t trigger them.
Run it through Haimaker
Skip juggling API keys. One Haimaker key gives you access to every model on the platform. Tell OpenClaw:
Add Haimaker as a custom provider to my OpenClaw config. Use these details:
- Provider name: haimaker
- Base URL: https://api.haimaker.ai/v1
- API key: [PASTE YOUR HAIMAKER API KEY HERE]
- API type: openai-completions
Add the auto-router model:
- haimaker/auto (reasoning: false, context: 128000, max tokens: 32000)
Create an alias "auto" for easy switching. Apply the config when done.
Or skip model selection entirely — Haimaker’s auto-router picks the best model for each task so you don’t have to.
OpenClaw setup
Configure this as a custom provider using the OpenAI-compatible schema. Set the base URL to https://api.deepseek.com and confirm your API key is active in the DeepSeek developer console before deploying.
{
"models": {
"mode": "merge",
"providers": {
"deepseek": {
"baseUrl": "https://api.deepseek.com/v1",
"apiKey": "YOUR-DEEPSEEK-API-KEY",
"api": "openai-completions",
"models": [
{
"id": "deepseek-chat-v3.1",
"name": "DeepSeek V3.1",
"cost": {
"input": 0.2,
"output": 0.7999999999999999
},
"contextWindow": 163840,
"maxTokens": 163840
}
]
}
}
}
}
How it compares
- vs GPT-4o mini — GPT-4o mini is faster and marginally cheaper on input, but V3.1 handles complex reasoning better and its output window is dramatically larger.
- vs Claude 3.5 Sonnet — Sonnet is more reliable and follows complex instructions better. It also costs $3.00/M input versus V3.1’s $0.20. That’s the tradeoff in a sentence.
Bottom line
If you’re generating large volumes of code or long-form output and cost is a real constraint, V3.1’s combination of 164K output and $0.20/M input is hard to argue with.
For setup instructions, see our API key guide. For all available models, see the complete models guide.