DeepSeek API Privacy & Zero Data Retention
Run DeepSeek V4 Pro and V4 Flash through an attested OpenAI-compatible gateway. No prompt or output logging, with enforceable ZDR route filters.
DeepSeek V4 on attested infrastructure: your prompts never reach the model vendor.
The worry behind "is the DeepSeek API safe" is specific. api.deepseek.com runs on the vendor's own infrastructure, several US states have restricted it on government devices, and a January 2025 incident exposed plaintext chat logs. The V4 weights themselves are MIT-licensed. The risk lives in where the endpoint runs.
TrustedRouter serves deepseek-v4-pro and deepseek-v4-flash through non-Chinese hosting providers inside a hardware-attested confidential-compute gateway. TrustedRouter never logs prompt or output content. Ordinary synchronous and streaming inference does not retain it. The opt-in Batch API temporarily retains enclave-encrypted artifacts for up to 30 days. Zero-Data-Retention and TEE tiers are available per route. Migration is one line: change base_url, keep your OpenAI SDK.
from openai import OpenAI
client = OpenAI(
base_url="https://api.trustedrouter.com/v1",
api_key=TR_KEY,
)
r = client.chat.completions.create(
model="deepseek-v4-pro", # or deepseek-v4-flash
messages=[{"role": "user", "content": "Summarize this contract."}],
)Served outside China
DeepSeek V4 routes run on non-Chinese hosting providers on attested infrastructure. Prompts do not go to the model vendor. EU-focused routing is available through the EU gateway.
Attestation you can run
The gateway source is open and the image digest is published. One curl with a nonce returns a JWT signed by the CPU vendor's root key, bound to the live TLS session. Details at /security.
No prompt or output logs
TrustedRouter never logs prompt or output content. Ordinary synchronous and streaming inference does not retain it. The opt-in Batch API temporarily retains enclave-encrypted artifacts for up to 30 days. Zero-Data-Retention routes add contractual guarantees; TEE routes keep prompts sealed even from the hosting provider. See zero data retention.
Make downstream retention a hard routing rule.
TrustedRouter never logs prompt or output content. Ordinary synchronous and streaming inference does not retain it. The opt-in Batch API temporarily retains enclave-encrypted artifacts for up to 30 days. For ordinary real-time DeepSeek requests, set provider.min_privacy to zdr to consider only providers with a recorded zero-data-retention posture.
The filter fails closed. If no eligible DeepSeek endpoint is available, the router returns an error instead of silently sending the request to a weaker privacy tier.
{
"model": "deepseek/deepseek-v4-flash",
"provider": {"min_privacy": "zdr"},
"messages": [
{"role": "user", "content": "Summarize this contract."}
]
}Run the attestation check yourself.
Other US hosts also serve DeepSeek outside China with zero-retention policies, and that is a real improvement over first-party routing. They still ask you to take the policy on faith. TrustedRouter's gateway proves it: run curl "https://api.trustedrouter.com/attestation?nonce=$NONCE" and check the returned JWT against the published image digest. Your compliance team gets evidence, plus the DPA, subprocessor list, and SOC 2 readiness documentation to assess it with.
Hosting is the control point for Chinese open-weights generally. In our testing, the same GLM weights answered 60 of 60 sensitive questions through TrustedRouter's enclave versus 30 of 60 via the vendor-hosted API; the filtering is host-applied, not in the weights. One anonymized proof point: a document-processing company scaled from zero to over a billion tokens per week through the attested gateway within three weeks of onboarding. Use the live model catalog for current DeepSeek model IDs, try one in the playground, compare current latency on the live leaderboard, and inspect stable history in the monthly benchmark reports.
Current routes, prices, privacy, and measured performance.
Catalog facts come from the routes currently configured in TrustedRouter. Performance uses the same cached metadata snapshot as the public leaderboard. Prompts and outputs are not part of these measurements.
| Model | Providers | Context | Input | Output | Privacy | Measured route |
|---|---|---|---|---|---|---|
DeepSeek: DeepSeek V4 Flash 0731deepseek/deepseek-v4-flash-0731 |
+19
|
1,048,576 | $0.047475/1M to $2.11/1M | $0.09495/1M to $10.55/1M | ZDR | 1930 ms TTFT deepseek · 152 tok/s · 95.71% available · n=939 |
DeepSeek V4 Pro 0813deepseek/deepseek-v4-pro-0813 |
|
1,024,000 | $0.6963/1M to $1.3926/1M | $2.0889/1M to $4.1778/1M | ZDR | 1714 ms TTFT baseten · 75 tok/s · 100.00% available · n=77 |
DeepSeek: DeepSeek V4 Flash 0423deepseek/deepseek-v4-flash |
+15
|
1,024,000 | $0.07174/1M to $0.4642/1M | $0.17724/1M to $1.3926/1M | E2EE 5 cited scores | 1929 ms TTFT tinfoil · 78 tok/s · 95.24% available · n=52 |
DeepSeek: DeepSeek V4 Pro 0423deepseek/deepseek-v4-pro |
+12
|
1,024,000 | $0.91785/1M to $2.532/1M | $1.8357/1M to $5.064/1M | ZDR 5 cited scores | 3433 ms TTFT arcee · 100.00% available · n=33 |
DeepSeek: DeepSeek V3.2deepseek/deepseek-v3.2 |
+7
|
163,840 | $0.26375/1M to $1.508756/1M | $0.4009/1M to $2.538647/1M | E2EE 3 cited scores | 8083 ms TTFT chutes · 100.00% available · n=17 |
DeepSeek: DeepSeek V4 Flash Vision Expdeepseek/deepseek-v4-flash-vision-exp |
+2
|
1,048,576 | $0.13715/1M to $0.4642/1M | $0.2954/1M to $1.3926/1M | varies | 1268 ms TTFT deepseek · 100.00% available · n=6 |
Browse every modelReview provider policiesOpen the full leaderboardSnapshot 2026-09-07T09:16:50.445Z
Questions
Is the DeepSeek API safe to use?
It depends on which endpoint you call. api.deepseek.com is the vendor's own service, and several US states have restricted it on government devices. TrustedRouter serves the same MIT-licensed V4 weights through non-Chinese hosting providers inside a hardware-attested gateway. Prompts do not go to the model vendor. TrustedRouter never logs prompt or output content. Ordinary synchronous and streaming inference does not retain it. The opt-in Batch API temporarily retains enclave-encrypted artifacts for up to 30 days. You can verify the attestation live at any time instead of relying on a policy.
Does using DeepSeek through TrustedRouter send data to China?
No. DeepSeek V4 routes are served by non-Chinese hosting providers on attested infrastructure, so prompts never reach the model vendor. Zero-Data-Retention routes add a contractual guarantee that providers keep nothing, and TEE routes keep the prompt sealed even from the hosting provider. Each route's privacy tier is listed on the models page, and the attestation backing the claim is checkable live.
How do I require DeepSeek zero data retention?
Set provider.min_privacy to zdr on a DeepSeek request. The router then considers only endpoints with a recorded zero-data-retention posture and fails closed if none are eligible. TrustedRouter never logs prompt or output content. Ordinary synchronous and streaming inference does not retain it. The opt-in Batch API temporarily retains enclave-encrypted artifacts for up to 30 days. The ZDR filter adds the downstream provider requirement.
Which DeepSeek model ID should I use?
Read the live TrustedRouter model catalog and choose the current DeepSeek route for your workload. The API is OpenAI-compatible, so changing generations is a model-id change rather than an SDK migration. The live leaderboard shows current route behavior, while monthly benchmark reports preserve historical measurements.