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
Baseten's Security Practices (Last Updated 2026-11-12) say Baseten "adheres to a Zero Data Retention posture" for Model APIs and Dedicated Inference and will not "store, retain, or otherwise make a persistent copy of model inputs or outputs". The docs add that Dedicated Inference synchronous requests "run under ZDR by default". The Security Practices do not say "ZDR is the default configuration". A Customer-Controlled Factors clause says Baseten is not in breach of the ZDR commitments where retention follows from customer configuration, customer runtime code, a customer-enabled Third-Party Product or a written customer direction. Terms Sec 2.5 and DPA Sec 4.5 tie the commitment to that ZDR designation. Async Inference (an elective Dedicated Inference feature) queues inputs for up to 72 hours before deletion, and outputs are never stored. Terms Sec 6.3 bars Baseten from using Customer Content - including deployed model weights, outputs and inference logs - to train, fine-tune or otherwise develop models.
Data & IP ownership
Terms & Conditions Sec 6.1: "Customer reserves all rights, title, and interest in and to Customer Content," a definition that includes deployed Customer Models and Model Outputs. A mutual confidentiality clause (Sec 11.1-11.2) names Customer Content and Customer Models and requires both parties to protect confidential information "with the same degree of care it uses for its own (but no less than reasonable care)," surviving three years after termination (trade secrets indefinitely, Sec 11.4). Most peers do not match this. The DPA gives 15 days' notice of new sub-processors (Sec 7.1) and a ten-day objection right. The list at trust.baseten.co has 13 rows, and six cloud-compute additions since August sit behind a gated compute-providers PDF.
Residency & sovereignty
The Regional deployments docs publish an EU region (eu, next to us), selectable per GPU
model deployment through the CLI or Management API, with a regional endpoint. Model APIs use a
shared endpoint whose serving regions "can change", so staying in specific regions needs a
compliance policy enabled by support. Chains and training jobs cannot select a region. The docs
do not say where the Baseten-operated control plane runs. Baseten Self-Hosted (Enterprise) is
the stronger path, running the whole workload plane inside the customer's own VPC.
CLOUD Act disclosure: Baseten is US-headquartered (San Francisco). Standard US jurisdiction applies to the managed service, including a deployment pinned to the EU region.
Compliance & attestations
SOC 2 Type II is confirmed on the Security Practices (named audit firm Sensiba San Filippo LLP on the vendor's blog), as are HIPAA (BAAs with qualifying customers) and a GDPR DPA. The Vanta Trust Center renders only in a browser. It labels ISO 27001:2022, PCI DSS SAQ D, CCPA and GDPR next to SOC 2 and HIPAA, and its updates feed announced the ISO 27001 certification on 2026-08-24. Pages readable without JavaScript show no trace of ISO 27001. The certificate, the 5.31.26 SOC 2 and SOC 3 reports, the HIPAA report and the PCI attestation are all behind "Request access", so they are vendor claims and the compliance score is 4. ISO 27701 and CSA STAR are not listed. A FedRAMP 20x package is listed but is not an authorisation.
Security controls
Isolation is logical namespace separation for shared use (access controls, unique customer identifiers) plus Dedicated single-tenant Kubernetes namespaces with Calico and Cilium network policies ("Baseten never shares GPUs across users") and Self-Hosted deployment inside the customer's own VPC. Encryption is TLS 1.2+ in transit (mTLS between clusters) and AES-256 at rest. Access is least-privilege with enforced MFA for all personnel, with quarterly access reviews. Customers are notified of an unauthorised disclosure within 48 hours. The Security Practices say Baseten engages qualified third-party firms for periodic penetration testing, and the docs say annually. The trust centre lists an October 2026 test report that is gated. No public bug-bounty programme was found. By default a deployment can reach any internet host unless the customer configures egress restrictions.
Pricing & cost model
Per-token pricing for hosted Model APIs and per-GPU-minute pricing for dedicated deployments are both published on Baseten's pricing page, with pay-as-you-go and "Production scale" routes and published default rate limits. Example rates per 1M tokens (input / cached / output): GLM-5.3 $1.40 / $0.14 / $4.40; DeepSeek V4.1 Flash $0.30 / $0.007 / $1.20; Kimi K3.1 Pro $2.00 / $0.20 / $10.00. Dedicated GPU-minute rates run from T4 $0.01052 up to H100 80GB $0.10833 (about $6.50 an hour) and B200 $0.16633.
Reliability posture
The SLA at baseten.co/service-level-agreement (stamped 2026-09-15) is a "commercially reasonable measures" undertaking to reach 99.9% monthly availability for Dedicated Inference and Model APIs on Baseten-managed infrastructure. Service credits are 10/15/30%, capped at 40% of monthly fees, with a 24-hour claim window. It is not an unqualified guarantee. The public status page shows all systems operational and logs 7 major or critical incidents between 2026-09-20 and 2026-11-18, all resolved. They include an 11 h 45 min regional event on 2026-09-29 and a 21-minute outage of all Model APIs on 2026-11-11.
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/Anthropic-compatible APIs over a predominantly open-weight Model Library keep workloads portable. The Terms & Conditions confirm the customer retains all IP in Customer Content, including deployed model weights and outputs (Sec 6.1). The deployment ladder (shared, dedicated, single-tenant, self-hosted, hybrid) gives control over where and how the workload runs.
TransparencyAre the binding terms published, legible and independently checkable?
ModerateThe Terms & Conditions, DPA, Security Practices, SLA and docs are legible, and the Trust Center renders in a browser (13-row sub-processor list, dated updates feed). Held at moderate because every certificate and report behind it (SOC 2 and SOC 3, ISO 27001, HIPAA, pen test, PCI AOC, compute-provider roster) is access-gated. The Security Practices have no explicit 'ZDR is the default' sentence and carry a customer-controlled-factors carve-out, and the docs do not say where the control plane runs.
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.