Can you own it?
This page is a projection of the one entry record, the Use & modify and Transparency factors that Assess covers. The full verdict is set by all four factors together, floor-weighted so the weakest caps the whole.
- AssessUse & modify + Transparency
- ImplementData control + Reliability
- UseReliability
- SupportTransparency
Data governance - retention, ZDR & training
Zero Data Retention is contractual and controlling in the Terms of Service (last modified 2026-10-15). Sec 7(b): "Provider will not retain, store, or log any Customer Data submitted to or generated by the Services beyond the period strictly necessary to process and return the applicable request... ('Zero Data Retention')", and Sec 5(a) says this commitment "controls over the Privacy Policy and any other Provider policy, webpage, or documentation". Training is barred except as necessary to provide the Services (for example a fine-tune you ask for). The exceptions are bounded: support data kept at your written request and deleted within 30 days of resolution, non-content metadata, legal and abuse records, and data kept for a feature you elect (batch requests, asynchronous media outputs, third-party integrations or paid cache retention) for the documented period.
ZDR is scoped: it "applies to models hosted by Provider". The Terms state in the binding text that data you send to a third-party (partner) model is governed by that partner's terms. In the live listing, 94 of 232 public entries are flagged partner, including Anthropic Claude and Google Gemini, and the trust-centre FAQ says partner models "run on the partner's infrastructure, and the partner's data policy applies". Segregate open-model and partner-model traffic when governance matters. Nothing independent attests the ZDR itself.
Data & IP ownership
For DeepInfra-hosted models, Sec 7(b) leaves all IP in Customer Data with you, gives the provider only a licence to host and process it to provide the Services, and bars sale and training. Sec 17 is a mutual confidentiality duty with three-year survival. Two limits apply: the definition is general and does not name Customer Data as Confidential Information, and Sec 17(f) lets the provider use your name and logo in marketing. Ownership is not uniform across the catalogue: partner models carry the partner's terms. Sub-processors are disclosed (AWS, Google Cloud Platform, Stripe, all United States), but Sec 6(a) lets DeepInfra engage new ones "without Customer's consent". The trust-centre FAQ says the DPA gives 7 days notice, which is unverified because the DPA PDF is blocked.
Compliance & attestations
Terms Sec 5(a) incorporates a data processing agreement "by reference" that applies to personal data in Customer Data, with no separate signature. The trust centre lists it as "2026-10-05 - Deep Infra - DPA" and its FAQ says it "includes the EU Standard Contractual Clauses (Modules 1 to 3), the UK IDTA and Swiss terms", and that DeepInfra is not certified under the EU-US Data Privacy Framework. The DPA PDF returns HTTP 403 from CloudFront, so the SCC content is a publisher statement.
The trust centre shows SOC 2 and ISO 27001 as "Compliant", GDPR as "In progress" and HIPAA as "Compliant". It lists a SOC 2 Type 1 report with a bridge letter, an ISO 27001 certificate, an ISMS manual, a pen-test report and a BAA, all unverified. A BAA needs an enterprise agreement and covers hosted models only. Read compliance as: DPA in place, attestations claimed and unverified.
Data residency & jurisdiction
The trust-centre FAQ says "All DeepInfra inference runs in data centers in the United States", for both the serverless API and dedicated deployments, with "no EU or other non-US region" and no choice of US region. The Privacy Policy says data is processed "specifically in the United States". The SCCs in the DPA give a transfer mechanism, not residency. As a US-headquartered company, DeepInfra carries standing US CLOUD Act exposure. For workloads with EU residency requirements this is a hard constraint.
Security controls
The trust centre states SOC 2 and ISO 27001 as "Compliant", shows control families (MFA, network and endpoint controls, vulnerability remediation) and lists an ISMS manual and a pen-test report. The docs and FAQ say inference input is held in memory only, batch data is encrypted on disk then deleted, and the prompt cache is in memory per account and model. The assessment is held to partial because no certificate, SOC 2 report or pen-test report is verified, SOC 2 is Type 1, and tenant isolation on shared endpoints is not described in the sources.
Pricing & cost model
First-party list prices (USD per million tokens, input / output): Llama 3.1 8B Turbo 0.02 / 0.04, Llama 3.3 70B Turbo 0.10 / 0.32, gpt-oss-120B 0.037 / 0.17, DeepSeek V4 Flash 0.09 / 0.18, Kimi K3.1 Pro 2.00 / 10.00. Service tiers add Priority (1.5x) and Flex (0.8x). Dedicated GPUs cost A100 $1.49, H100 $2.90, H200 $3.49, B200 $4.89 and B300 $5.89 per GPU-hour, 59% to 75% above the July 2026 figures. A card or prepayment is required and a spending limit can be set. Serverless prices are very low, but "among the cheapest" is not supported for dedicated capacity, so cost scores 4.
Reliability posture
A public status page (status.deepinfra.com) shows measured 90-day uptime for the API and website of 100% and per-model figures of 99.98% to 100%, a maintenance page and scheduled deprecations with replacements. There is no uptime SLA or credits; Terms Sec 4(b) offers only a 24-hour repair or replacement of a failed provided server. The page has no incident-history view, so past incident communication cannot be reviewed. DeepInfra suits cost-led work, not workloads that need a contractual availability commitment.
How this scores
The ownership factors this domain covers, drawn from the one entry record.
Use and modify freelyCan you use it freely and leave without lock-in?
StrongOpenAI-compatible API over a public catalogue that is mostly open-weight (DeepSeek V4, Kimi K3.1, GLM-5.x, Qwen3.5, gpt-oss, Gemma 4, Llama); the same models run elsewhere or self-hosted, and custom weights can be hosted on dedicated GPUs.
TransparencyAre the binding terms published, legible and independently checkable?
ModerateThe Terms, Privacy Policy, status page, trust centre and sub-processor list are readable, and the DPA is incorporated by Terms Sec 5(a) and listed (2026-10-05). Held at moderate because the DPA PDF, SOC 2 report, ISO certificate and pen-test report are not verifiable, partner models run under other terms, and the Terms still read as an Order Form MSA.
Sources
The same evidence records as the entry sheet. Read means the text was verified; unverified means it is known to exist but not yet read.