Baseten
US inference platform selling autoscaled dedicated model deployments and training jobs billed per GPU-minute, plus per-token Model APIs, on a multi-cloud GPU pool.
Trust Score v0.1- Company & track record73
- Infrastructure control100
- Pricing & billing transparency65
- Reliability signals100
- Security & compliance100
- Customer signal—
Scored Sep 29, 2026 from the provider's own pages. Dimensions not yet measured are excluded from the denominator. Criteria ↓
Trust Score · 5 of 6 dimensions measuredThe score, criterion by criterion
Each point comes from something the provider publishes, with the page it was read on and the date. Anything not disclosed earns zero.
| Founding year published | 2019baseten.co · Sep 29, 2026 | 4 / 4 |
| Headquarters published | Baseten Labs, Inc.baseten.co · Sep 29, 2026 | 3 / 3 |
| Legal entity named | Baseten Labs, Inc.baseten.co · Sep 29, 2026 | 4 / 4 |
| Funding or ownership disclosed | not disclosedbaseten.co · Sep 29, 2026 | 0 / 4 |
| Capacity model published | operates or rents dedicated capacitybaseten.co · Sep 29, 2026 | 10 / 10 |
| Regions published | 2 regions or countriesbaseten.co · Sep 29, 2026 | 5 / 5 |
| Public prices without an account | yes (read on the pricing page)baseten.co · Sep 29, 2026 | 5 / 5 |
| Billing granularity published | per minutebaseten.co · Sep 29, 2026 | 4 / 4 |
| Egress policy published | not disclosedbaseten.co · Sep 29, 2026 | 0 / 4 |
| Stopped-instance billing published | Scaled to zero (min_replica 0): no charge. Idle warm replicas (min_replica >= 1) are billed; default scale_down_delay is 900 s, so replicas remain billed for 15 min after load dropsbaseten.co · Sep 29, 2026 | 4 / 4 |
| Public IPv4 policy published | not disclosedbaseten.co · Sep 29, 2026 | 0 / 3 |
| Status page | https://status.baseten.cobaseten.co · Sep 29, 2026 | 6 / 6 |
| SLA published | 99.9% monthly System Availability for Dedicated Inference and Model APIs on Baseten-managed infrastructure; credits 10% / 15% / 30% of the affected monthly bill, capped at 40%, to be claimed within 24 hbaseten.co · Sep 29, 2026 | 6 / 6 |
| SLA with a numeric uptime commitment | 99.9% monthly System Availability for Dedicated Inference and Model APIs on Baseten-managed infrastructure; credits 10% / 15% / 30% of the affected monthly bill, capped at 40%, to be claimed within 24 hbaseten.co · Sep 29, 2026 | 3 / 3 |
| Uptime history published | 90-day uptime on the status page on 2026-09-29: Dedicated Inference 99.99%, Model APIs 99.93%, Training 100%baseten.co · Sep 29, 2026 | 5 / 5 |
| Observed availability (30 days) | not measured yet | — |
| SOC 2 | Type IIbaseten.co · Sep 29, 2026 | 8 / 8 |
| ISO 27001 | ISO 27001:2022 certificate listed on Trust Centerbaseten.co · Sep 29, 2026 | 4 / 4 |
| HIPAA or GDPR DPA | yes (HIPAA report listed on Trust Center)baseten.co · Sep 29, 2026 | 3 / 3 |
| Aggregated public reviews | not collected yet | — |
Market report · Oct 25, 2026 · Specialised GPU clouds, hourly rental, pods and inference · #75 of the 200 to audit · prices read on the provider's pageWhat the market report notes about Baseten
Our internal report lists 248 actors with their zone, segment, status and the prices it could read, dated Oct 25, 2026. Its notes are reproduced as written, in French, with their sources.
- Listed as
- Baseten
Prices by GPU · 6 GPU types with a current offerWhat Baseten charges, against the market
Cheapest current on-demand true cost here, next to the market median for the same GPU. Negative means cheaper than the market.
| GPU | Cheapest here | Market median | vs market | Current offers | Modes | |
|---|---|---|---|---|---|---|
| H100 | $3.75 | $3.17 | +18% | 2 | on-demand | H100 prices |
| A100 80GB | $4.00 | $1.99 | +101% | 1 | on-demand | A100 80GB prices |
| B200 | $9.98 | $7.50 | +33% | 1 | on-demand | B200 prices |
| L4 | $0.85 | $0.80 | +6% | 1 | on-demand | L4 prices |
| T4 | $0.63 | $0.51 | +24% | 1 | on-demand | T4 prices |
| A10G | $1.21 | $1.01 | +20% | 1 | on-demand | A10G prices |
All offers · 56 readings on fileEvery Baseten offer, with its receipt
Current offers (last 45 days) ranked by true cost, all rental modes, each with its read date. Older readings stay listed below, in grey.
| GPU | Configuration | Mode | Region | Stock | List | True cost | Read | |
|---|---|---|---|---|---|---|---|---|
| T4 | 1× T4 | On-demand | — | Stock unknown | $0.63 | $0.63 | 21:35 UTCfresh | Receipt |
| L4 | 1× L4 | On-demand | — | Stock unknown | $0.85 | $0.85 | 21:35 UTCfresh | Receipt |
| A10G | 1× A10G | On-demand | — | Stock unknown | $1.21 | $1.21 | 21:35 UTCfresh | Receipt |
| H100SXM | 1× H100 SXM | On-demand | — | Stock unknown | $3.75 | $3.75 | 21:35 UTCfresh | Receipt |
| A100 80GB | 1× A100 80GB | On-demand | — | Stock unknown | $4.00 | $4.00 | 21:35 UTCfresh | Receipt |
| H100SXM | 1× H100 SXM | On-demand | — | Stock unknown | $6.50 | $6.50 | 21:35 UTCfresh | Receipt |
| B200 | 1× B200 | On-demand | — | Stock unknown | $9.98 | $9.98 | 21:35 UTCfresh | Receipt |
True cost & billing · read Sep 29, 2026What the hourly price does not say
Billing granularity, minimums, what a stopped machine costs, traffic, addresses, deposits and identity checks, as published by the provider.
Billing
- Billing granularity
- per minute
- Minimum charge
- Partial minutes round up; image builds and cold starts (model load) are billed; failed boots and image pulls are not
- Stopped instance
- Scaled to zero (min_replica 0): no charge. Idle warm replicas (min_replica >= 1) are billed; default scale_down_delay is 900 s, so replicas remain billed for 15 min after load drops
- Egress
- not published
- Public IPv4
- not published
- Billing model
- postpaid invoice
- Minimum deposit
- not published
- Free credits
- New workspaces receive credits (amount not published); models are deactivated when credits run out and no payment method is on file
Account
- Payment
- card · invoice · crypto (not published)
- Sign-up
- email (or Google / GitHub sign-in), payment method (card or US bank account) once starting credits are used, organization verification via support for regional selection and higher Model API limits
- Identity check
- not published
- Support
- email, in-app chat, slack (Pro), zoom (Pro)
- Enterprise support
- not published
- SLA
- 99.9% monthly System Availability for Dedicated Inference and Model APIs on Baseten-managed infrastructure; credits 10% / 15% / 30% of the affected monthly bill, capped at 40%, to be claimed within 24 h
- Status page
- status.baseten.co ↗
- Published uptime
- 90-day uptime on the status page on 2026-09-29: Dedicated Inference 99.99%, Model APIs 99.93%, Training 100%
Infrastructure
- Regions
2 named regions in 2 countries
us (United States) · eu (European Union (country not specified))- Countries
- United States, European Union (country not specified)
- Platform
- api, cli, serverless, ssh, custom_docker_image · SDK : python (Truss, Chains, Training, Loops SDKs) · Performance client for Python, JavaScript, Rust · Terraform : — · Kubernetes : no · interconnexion : InfiniBand between nodes for multi-node deployments (node_count) · clusters : Multi-node deployments (node_count), multi-node Training Jobs, one-command Slurm workstations · single-tenant and self-hosted (customer VPC) options on Enterprise
- Storage
- object · not published · Weight caching (BDN, cached weights) and a training cache exist; no storage price published
- Migration out
- Models are packaged in the open-source Truss format or as custom Docker images and pull weights from Hugging Face, S3 or GCS, so the same container can be redeployed elsewhere. Training artifacts and checkpoints belong to the customer. There are no inference volumes to export; training cache data is deleted with the project.
Security & compliance
- SOC 2
- Type II
- ISO 27001
- ISO 27001:2022 certificate listed on Trust Center
- HIPAA
- yes (HIPAA report listed on Trust Center)
- GDPR / DPA
- Data Processing Agreement listed on Trust Center
- Other
- SOC 3, PCI DSS SAQ D AOC, CCPA, FedRAMP 20x package overview listed on Trust Center; annual third-party penetration tests
- Legal entity
- Baseten Labs, Inc.
- Funding
- not published
From the study · assessmentBaseten, in short
What stood out when we read the documents: strengths, weaknesses, who it suits, and what we could not find.
Strengths
- Scale-to-zero autoscaling of dedicated GPU replicas billed per minute; no charge for failed boots or image pulls
- 39 published GPU configurations across 8 GPU models (T4 to B200), 1 to 8 GPUs per replica, plus multi-node InfiniBand deployments
- Contractual 99.9% SLA with service credits and a status page with per-component 90-day uptime
- SOC 2 Type II, SOC 3, ISO 27001:2022, HIPAA and PCI DSS SAQ D documents listed on the Trust Center
- Self-hosted (customer VPC), hybrid and single-tenant options; per-token Model APIs for hosted LLMs
- Training Jobs and the Loops RL SDK at the same per-minute instance prices, with SSH into training containers
Weaknesses
- H100 at 0.10833 USD/min = 6.4998 USD per GPU-hour, H200 7.50, B200 9.9798, A100 and RTX PRO 6000 4.0002: between 2.4x and 4.2x GPUBank list prices
- Idle warm replicas are billed, the default 900 s scale-down delay keeps replicas billed after traffic stops, and cold starts and image builds are billed
- Only two selectable regions (us, eu); region selection is restricted to verified organizations
- No published storage, egress or reserved-capacity prices; volume discounts through sales only
- SSH into inference deployments must be enabled by support; no general-purpose VM or pod product
- Payment by card or US bank account; no crypto published
Ideal for
- Teams serving custom or open-source models in production that want autoscaling, scale-to-zero and managed inference optimisation rather than raw GPU machines
- Regulated buyers needing SOC 2, ISO 27001, HIPAA and a contractual SLA
- Enterprises that want self-hosted or single-tenant inference
Not published
- Funding and investors: not stated on the pages read
- Amount of starting credits for new workspaces: not published
- Storage, egress and IP pricing: not published
- Reserved or committed capacity and volume-discount prices: sales only
- Spot or interruptible tier: none published
- Terraform provider: none found in the docs index
- What organization verification involves (documents, card, company check): not documented
- Exact GPU form factors (H100 SXM vs PCIe, A100 SXM vs PCIe) not named
- inferred from published memory bandwidth
- Physical data-center locations: only region slugs us and eu, plus subprocessor locations on the Trust Center (AWS, Google Cloud global
- Crusoe US, Ireland
- Oracle US, UK, UAE)
Observations
Pricing page shows GPU prices per minute by default with a per-hour toggle, visible without an account
Pricing page lists 7 GPU instances (T4, L4, A10G, A100, H100 MIG, H100, B200), each only in its smallest configuration; H200, RTX PRO 6000 and all multi-GPU sizes appear only in the docs instance reference
Three plans: Basic ($0 per month, pay as you go), Pro and Enterprise (quote, volume discounts)
Model API token prices for 11 LLMs shown on the same page
Site-wide banner on 2026-09-29: 'Blaxel is joining Baseten'
Pricing FAQ says customers do not pay for idle time, while the billing docs state idle warm replicas are billed
'Talk to sales about compute in other countries and regions' under each GPU table
Sign-up asks only for an email, or Google / GitHub sign-in
Pages read
- baseten.co/pricing/ — Plans, per-minute GPU prices on the page, Model API prices, FAQ (credits, idle time, support, discounts
- docs.baseten.co/deployment/resources — Full GPU instance table with per-minute prices, vCPU, RAM, VRAM, GPU specs
- docs.baseten.co/organization/billing — Per-minute metering, lifecycle billing, payment methods, credits, invoice cadence, x60 conversion
- docs.baseten.co/deployment/autoscaling/overview — Scale-to-zero, default scale-down delay 900 s
- docs.baseten.co/deployment/regional-deployments — Regions us and eu, verified-organization requirement
- docs.baseten.co/hosting-options/cloud — Multi-cloud pool, GPUs never shared, self-service
- docs.baseten.co/hosting-options/single-tenant — Enterprise custom pricing, private networking
- docs.baseten.co/observability/security — Data retention, isolation, pentesting
- docs.baseten.co/reference/overview — APIs, CLIs, SDKs
- docs.baseten.co/inference/ssh — SSH into deployments, enabled on request
- docs.baseten.co/training/concepts/cache — Training storage behaviour
- docs.baseten.co/inference/model-apis/pricing-and-limits — Account tiers Basic unverified/verified, rate limits
- baseten.co/service-level-agreement/ — 99.9% SLA and service credits
- status.baseten.co/ — 90-day uptimes
- trust.baseten.co — SOC 2, ISO 27001, HIPAA, PCI DSS, FedRAMP 20x, subprocessors
- baseten.co/terms-and-conditions/ — Legal name, San Francisco address, invoice terms
- baseten.co/about-us/ — Founded 2019, founders
- baseten.co/deployments/baseten-cloud/ — Deployment options, compliance claims
- baseten.co/products/training/ — Training Jobs and Loops
- login.baseten.co/sign-up — Email, Google or GitHub sign-up
- baseten.co/ — Positioning
Questions about Baseten
What GPUs does Baseten offer?
Baseten has 8 GPU types with a fresh offer today, for example H100 from $3.75, A100 80GB from $4.00, B200 from $9.98, L4 from $0.85, T4 from $0.63, A10G from $1.21 per GPU-hour on-demand. The full list with prices and market medians is on this page.
How does Baseten bill?
Billing granularity: per minute. Minimum charge: Partial minutes round up; image builds and cold starts (model load) are billed; failed boots and image pulls are not Stopped instance: Scaled to zero (min_replica 0): no charge. Idle warm replicas (min_replica >= 1) are billed; default scale_down_delay is 900 s, so replicas remain billed for 15 min after load drops
How trustworthy is Baseten?
Its Trust Score is 87/100 (v0.1), read from its own documents. Strong on infrastructure control, reliability signals, security & compliance. Weaker on pricing & billing transparency. The score measures transparency and published guarantees, not customer satisfaction; every criterion and its source is listed above.
Where are Baseten's data centres?
2 named regions in 2 countries: United States, European Union (country not specified).
How do you pay Baseten, and are there credits?
Payment by card, invoice, crypto. Free credits: New workspaces receive credits (amount not published); models are deactivated when credits run out and no payment method is on file.