US AI Models & US-Operated Providers
Which AI models come from US labs, which providers are operated by US companies, and how to require US-operated routes on a request. Both facts, kept apart.
"US AI models" is two questions. This page answers both, separately.
Who built it and where your prompt goes are different facts. 160 models here come from 7 US labs: OpenAI, Google DeepMind, Anthropic, Meta, NVIDIA, xAI, Thinking Machines Lab. 34 of 51 providers are US-headquartered.
Need US-only routing? Set provider.jurisdiction to us. It fails closed when nothing qualifies.
from openai import OpenAI
client = OpenAI(
base_url="https://api.trustedrouter.com/v1",
api_key="sk-tr-...", # your TrustedRouter key
)
reply = client.chat.completions.create(
model="trustedrouter/liberty-3.0",
messages=[{"role": "user",
"content": "Draft the incident summary."}],
extra_body={"provider": {"jurisdiction": "us"}},
)Liberty: US labs, open weights, one route
open weightOur own family. Each Liberty id calls several component models and returns one answer — and every component is an open-weight model from a US lab. Both facts are computed from the catalog when this page renders, so if a component were ever swapped the counts would move and the claim would retire itself.
| Route | Context | Component models | From US labs | US-operated route available |
|---|---|---|---|---|
trustedrouter/liberty-3.0TrustedRouter Liberty 3.0 |
1,048,576 |
nvidia/nemotron-3-ultra-550b-a55b
google/gemma-4-31b-it
openai/gpt-oss-120b
trustedrouter/liberty-1.0-1m
thinkingmachines/inkling
|
6 of 6 | yes |
trustedrouter/liberty-2.0TrustedRouter Liberty 2.0 |
262,144 | 6 models once the nested Liberty presets are expanded | 6 of 6 | yes |
parasail/liberty-2.0Parasail Liberty 2.0 |
262,144 | 6 models once the nested Liberty presets are expanded | 6 of 6 | yes |
trustedrouter/liberty-1.0-1mTrustedRouter Liberty 1.0 1M |
1,048,576 | 2 of 2 | yes | |
trustedrouter/liberty-1.0TrustedRouter Liberty 1.0 |
262,144 | 3 of 3 | yes |
Two limits worth stating plainly. First, the component models are open-weights models that many providers serve, and Liberty's default candidate pool includes operators outside the US; set provider.jurisdiction to us when the serving provider has to be US-operated. Second, operator jurisdiction is not datacentre location — a US-registered company can run inference on hardware elsewhere, and the catalog records the company, not the machine.
Measurement for these routes is published rather than summarised here: our own DRACO run for Liberty 2.0 is written up in the Liberty 2.0 post, and live route behaviour is on the leaderboard.
Open weights from US labs
57 routesThe labs themselves. Weights you can download, run anywhere, and inspect.
- Meta20 open-weight models · meta-llama/Llama-3.3-70B-Instruct
- Google DeepMind15 open-weight models · google/gemma-2-2b-it
- NVIDIA15 open-weight models · nvidia/Nemotron-3-Nano-Omni
- OpenAI4 open-weight models · openai/gpt-oss-120b
- Thinking Machines Lab3 open-weight models · thinkingmachines/inkling
Two separate facts: who operates the endpoint, and who built the model.
Counts below are read from the routes configured in TrustedRouter right now. Operator country is the legal home of the company running the endpoint a request reaches, taken from that company's own terms, privacy policy, or regulatory filing. It is not a claim about which datacentre serves the request.
Providers operated from the United States
34 of 51Send a request to one of these and the operator is a company headquartered in the United States — head office, not place of registration. Privacy tier is a separate column: it is the highest bar we have recorded from that provider's published policy.
| Provider | Operator country | Privacy tier | Recorded claims | Models served | Policy source |
|---|---|---|---|---|---|
| Novita AInovita | United States | Standard | no verified privacy claim | 111 | policy |
| DeepInfradeepinfra | United States | No-store | no verified privacy claim | 95 | policy |
| GMI Cloudgmi | United States | Standard | no verified privacy claim | 76 | policy |
| Atlas Cloudatlas-cloud | United States | Standard | no verified privacy claim | 75 | policy |
| Venicevenice | United States | Standard | no verified privacy claim | 58 | policy |
| OpenAIopenai | United States | Standard | ZDR on prepaid | 44 | policy |
| Lightning AIlightning | United States | Standard | no verified privacy claim | 36 | policy |
| Parasailparasail | United States | Zero retention | ZDR | 27 | policy |
| Cloudflare Workers AIcloudflare-workers-ai | United States | Standard | no verified privacy claim | 25 | policy |
| Togethertogether | United States | Zero retention | ZDR | 21 | policy |
| DigitalOcean Gradient AIdigitalocean | United States | Standard | no verified privacy claim | 18 | policy |
| Crusoecrusoe | United States | Standard | no verified privacy claim | 15 | policy |
| Telnyxtelnyx | United States | Standard | no verified privacy claim | 14 | policy |
| Basetenbaseten | United States | Zero retention | ZDR | 13 | policy |
| Fireworks AIfireworks | United States | Standard | no verified privacy claim | 12 | policy |
| Databricksdatabricks | United States | Standard | no verified privacy claim | 11 | policy |
| Google AI Studiogoogle-ai-studio | United States | Standard | no verified privacy claim | 11 | policy |
| Anthropicanthropic | United States | Standard | no verified privacy claim | 10 | policy |
| Chuteschutes | United States | Confidential + E2EE | provider E2EE confidential ZDR | 10 | policy |
| Google Vertex AIgoogle-vertex | United States | Standard | ZDR on prepaid | 10 | policy |
| Morphmorph | United States | Standard | no verified privacy claim | 8 | policy |
| xAI Grokgrok | United States | Standard | no verified privacy claim | 7 | policy |
| FriendliAIfriendli | United States | Standard | no verified privacy claim | 6 | policy |
| Neurometric AIneurometric | United States | No-store | no verified privacy claim | 6 | policy |
| Tinfoiltinfoil | United States | Confidential + E2EE | provider E2EE confidential ZDR | 6 | policy |
| Makoramakora | United States | Standard | no verified privacy claim | 5 | policy |
| Cerebrascerebras | United States | Zero retention | ZDR | 4 | policy |
| Waferwafer | United States | Standard | no verified privacy claim | 4 | policy |
| 0G Private Computerzero-g | United States | No-store | confidential | 3 | policy |
| Thinking Machines Labthinkingmachines | United States | Standard | no verified privacy claim | 2 | policy |
| Runwayrunway | United States | Standard | no verified privacy claim | 1 | policy |
| Voyage AIvoyage | United States | Zero retention | ZDR | 1 | policy |
| Meta via OpenRoutermeta | United States | Standard | no verified privacy claim | 0 | policy |
| TrustedRoutertrustedrouter | United States | Confidential + E2EE | provider E2EE confidential ZDR | 0 | policy |
Every provider's full policy note and route list is on its own page. 2 providers in the catalog publish no operator TrustedRouter could pin down, so they appear in no jurisdiction list on this site and no request preference treats them as matching one. Review every provider.
Models built by US labs
160 modelsGrouped by the lab that built the weights, with the source TrustedRouter read that from. The operator columns show who runs the endpoints for each route, which is a different question from the lab's country and frequently a different answer.
OpenAI 57 models
OpenAI's terms of use name OpenAI OpCo, LLC, a Delaware company at 1455 3rd Street, San Francisco, CA 94158. Source.
| Model | Context | Routes | Operator jurisdictions | US-operated routes |
|---|---|---|---|---|
openai/gpt-oss-120bOpenAI: gpt-oss-120b |
131,072 | 26 |
United States ×13
Netherlands ×1
Singapore ×1
|
+7 more on the model page |
openai/gpt-oss-20bOpenAI: gpt-oss-20b |
131,072 | 15 |
United States ×8
Singapore ×1
|
+2 more on the model page |
openai/gpt-5.6-terraOpenAI: GPT-5.6 Terra |
1,050,000 | 6 |
United States ×4
|
|
openai/gpt-4oOpenAI: GPT-4o |
128,000 | 5 |
United States ×3
|
|
openai/gpt-3.5-turboOpenAI: GPT-3.5 Turbo |
16,385 | 4 |
United States ×2
|
|
openai/gpt-4OpenAI: GPT-4 |
8,191 | 4 |
United States ×2
|
51 further OpenAI models are in the catalog, ranked below these by route count. Browse every model.
Google DeepMind 34 models
Gemini and Gemma come from Google DeepMind, part of Alphabet Inc., whose 10-K gives principal executive offices at 1600 Amphitheatre Parkway, Mountain View, California. The code records that parent's home, not the location of every team that worked on a model. Source.
| Model | Context | Routes | Operator jurisdictions | US-operated routes |
|---|---|---|---|---|
google/gemma-4-31b-itGoogle: Gemma 4 31B |
262,144 | 24 |
United States ×11
Israel ×1
Singapore ×1
|
+5 more on the model page |
google/gemma-4-26b-a4b-itGoogle: Gemma 4 26B A4B |
262,144 | 12 |
United States ×6
Singapore ×1
|
|
google/gemini-2.5-flashGoogle: Gemini 2.5 Flash |
1,048,576 | 8 |
United States ×5
|
|
google/gemma-3-27b-itGoogle: Gemma 3 27B |
262,144 | 8 |
United States ×3
Netherlands ×1
|
|
google/gemini-3.1-pro-previewGoogle: Gemini 3.1 Pro Preview |
1,048,576 | 7 |
United States ×5
|
|
google/gemini-3.5-flashGoogle: Gemini 3.5 Flash |
1,048,576 | 7 |
United States ×5
|
28 further Google DeepMind models are in the catalog, ranked below these by route count. Browse every model.
Anthropic 20 models
Anthropic's commercial terms contract through Anthropic, PBC under California law outside the EEA, Switzerland, and the UK, where Anthropic Ireland, Limited applies instead. Source.
| Model | Context | Routes | Operator jurisdictions | US-operated routes |
|---|---|---|---|---|
anthropic/claude-opus-5Claude Opus 5 |
1,000,000 | 5 |
United States ×4
|
|
anthropic/claude-fable-5Anthropic: Claude Fable 5 |
1,000,000 | 3 |
United States ×2
|
|
anthropic/claude-haiku-4.5Anthropic: Claude Haiku 4.5 |
200,000 | 3 |
United States ×2
|
|
anthropic/claude-opus-4.5Anthropic: Claude Opus 4.5 |
200,000 | 3 |
United States ×2
|
|
anthropic/claude-opus-4.6Anthropic: Claude Opus 4.6 |
1,000,000 | 3 |
United States ×2
|
|
anthropic/claude-opus-4.7Anthropic: Claude Opus 4.7 |
1,000,000 | 3 |
United States ×2
|
14 further Anthropic models are in the catalog, ranked below these by route count. Browse every model.
Meta 20 models
The Llama licence names Meta Platforms, Inc. for licensees outside the EEA and Switzerland, and Meta Platforms Ireland Limited within them. Source.
| Model | Context | Routes | Operator jurisdictions | US-operated routes |
|---|---|---|---|---|
meta-llama/llama-3.3-70b-instructMeta: Llama 3.3 70B Instruct |
131,072 | 13 |
United States ×7
|
+1 more on the model page |
meta-llama/llama-3.1-8b-instructMeta: Llama 3.1 8B Instruct |
131,072 | 5 |
United States ×3
|
|
meta-llama/llama-4-maverickMeta: Llama 4 Maverick |
1,048,576 | 5 |
United States ×3
|
|
meta-llama/llama-4-scout-17b-16e-instructLlama 4 Scout Instruct |
131,072 | 5 |
United States ×3
|
|
meta-llama/llama-3.1-70b-instructMeta: Llama 3.1 70B Instruct |
131,072 | 4 |
United States ×2
|
|
meta-llama/llama-4-maverick-17b-128e-instruct-fp8Llama 4 Maverick Instruct |
1,048,576 | 4 |
United States ×2
|
14 further Meta models are in the catalog, ranked below these by route count. Browse every model.
NVIDIA 18 models
The Nemotron family comes from NVIDIA Corporation, whose corporate address for stockholder communications is 2788 San Tomas Expressway, Santa Clara, California 95051. Source.
| Model | Context | Routes | Operator jurisdictions | US-operated routes |
|---|---|---|---|---|
nvidia/nemotron-3-ultra-550b-a55bNVIDIA: Nemotron 3 Ultra |
512,288 | 9 |
United States ×4
Netherlands ×1
|
|
nvidia/nemotron-3-nano-30b-a3bNVIDIA: Nemotron 3 Nano 30B A3B |
262,144 | 6 |
United States ×3
|
|
nvidia/nemotron-3-super-120b-a12bNVIDIA: Nemotron 3 Super |
1,000,000 | 6 |
United States ×2
Netherlands ×1
|
|
nvidia/nemotron-3.5-lightningNVIDIA: Nemotron 3.5 Lightning |
1,000,000 | 4 |
United States ×2
|
|
nvidia/Cosmos3-Super-ReasonerCosmos3-Super-Reasoner |
8,000 | 2 |
Netherlands ×1
|
none configured |
nvidia/Llama-3_1-Nemotron-Ultra-253B-v1Llama 3_1 Nemotron Ultra 253B v1 |
128,000 | 2 |
Netherlands ×1
|
none configured |
12 further NVIDIA models are in the catalog, ranked below these by route count. Browse every model.
xAI 8 models
xAI's consumer terms take legal notices at 1450 Page Mill Rd., Palo Alto, CA 94304 and are governed by the laws of the State of Texas. Source.
| Model | Context | Routes | Operator jurisdictions | US-operated routes |
|---|---|---|---|---|
x-ai/grok-4.3SpaceXAI: Grok 4.3 |
1,000,000 | 3 |
United States ×2
|
|
x-ai/grok-4.5SpaceXAI: Grok 4.5 |
500,000 | 3 |
United States ×2
|
|
x-ai/grok-4.6SpaceXAI: Grok 4.6 |
500,000 | 3 |
United States ×2
|
|
x-ai/grok-build-0.1SpaceXAI: Grok Build 0.1 |
256,000 | 3 |
United States ×2
|
|
x-ai/grok-4.20SpaceXAI: Grok 4.20 |
2,000,000 | 2 |
United States ×1
|
|
x-ai/grok-4.20-reasoninggrok-4.20-0309-reasoning |
0 | 2 |
United States ×1
|
2 further xAI models are in the catalog, ranked below these by route count. Browse every model.
Thinking Machines Lab 3 models
Thinking Machines Lab Inc. is a public benefit corporation based in San Francisco, California. Its own site names the company without stating a location. Source.
| Model | Context | Routes | Operator jurisdictions | US-operated routes |
|---|---|---|---|---|
thinkingmachines/inklingThinking Machines: Inkling |
1,048,576 | 7 |
United States ×4
|
|
thinkingmachines/inkling-smallThinking Machines: Inkling Small |
524,288 | 6 |
United States ×3
|
|
thinkingmachines/inkling-1mInkling |
1,048,576 | 2 |
United States ×1
|
Model origin is keyed by the vendor prefix of a model id, and a prefix earns a row once at least three of its models are in the catalog. 40 callable catalog models sit under a prefix below that line and carry no recorded origin, so they are on no page here rather than assigned a country by guesswork.
Browse every modelReview provider policiesAI models from the European UnionAI models from China
A lab's country is fixed.
Where a model was built does not change when you route it differently. It is a procurement fact, and it is the one this page's lab sections answer, each with the primary source it was read from.
Your route is yours to pick.
Where prompts go is decided per request. provider.jurisdiction accepts us today; provider.only pins an exact allowlist for any other requirement, and both fail closed rather than falling back.
What this page will not say.
Not that a US operator keeps data in the US, not that a US lab's model is safer, and not that any of it is certified. It reports two recorded fields and the source behind each. Residency is a separate control.
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 |
|---|---|---|---|---|---|---|
Meta: Llama 3.3 70B Instructmeta-llama/llama-3.3-70b-instruct |
+3
|
131,072 | $0.142425/1M to $1.84625/1M | $0.422/1M to $2.90125/1M | E2EE 6 cited scores | 3370 ms TTFT tinfoil · 100.00% available · n=23 |
OpenAI: GPT-5.4 Miniopenai/gpt-5.4-mini |
|
400,000 | $0.79125/1M | $4.7475/1M | ZDR | 1164 ms TTFT openai · 100.00% available · n=20 |
Google: Gemini 3.5 Flashgoogle/gemini-3.5-flash |
+1
|
1,048,576 | $1.5825/1M | $9.495/1M | ZDR | 2852 ms TTFT google-vertex · 112 tok/s · 100.00% available · n=15 |
Google: Gemini 3.1 Pro Previewgoogle/gemini-3.1-pro-preview |
+1
|
1,048,576 | $2.11/1M | $12.66/1M | ZDR 3 cited scores | 3652 ms TTFT google-ai-studio · 72 tok/s · 100.00% available · n=12 |
Google: Gemini 3.1 Flash Litegoogle/gemini-3.1-flash-lite |
|
1,048,576 | $0.26375/1M | $1.5825/1M | ZDR | 1914 ms TTFT google-ai-studio · 100.00% available · n=12 |
Google: Gemini 3.5 Flash Litegoogle/gemini-3.5-flash-lite |
|
1,048,576 | $0.3165/1M | $2.6375/1M | ZDR | 2938 ms TTFT google-ai-studio · 100.00% available · n=11 |
Browse every modelReview provider policiesOpen the full leaderboardSnapshot 2026-08-20T14:22:52.106Z
Questions
What counts as a US AI model?
Two different things, so TrustedRouter records them separately. A model's origin is the country of the lab that built the weights, read from that lab's own licence, terms, or regulatory filing. A route's jurisdiction is the country of the company operating the endpoint the request reaches, read from that company's own terms or filing. A model from a US lab is regularly served by providers registered outside the US, and open-weights models from other countries are regularly served by US providers, so neither fact implies the other.
How do I make sure my prompts only reach US providers?
Set provider.jurisdiction to "us" on the request. The router then considers only providers whose recorded operator country is the United States and fails closed when none qualify, rather than falling back to a provider you did not approve. For an exact list rather than a country test, use provider.only with the provider slugs you approved.
Does a US provider mean my data stays in the United States?
No, and this site does not claim it. The recorded country is the legal home of the company operating the endpoint, taken from its published terms, privacy policy, or regulatory filing. It says nothing about which datacentre answers a given request. Where processing location is a contractual requirement, it comes from an agreed provider allowlist and a signed agreement, not from a country code in a catalog.
What is the Liberty model family?
Liberty is TrustedRouter's own set of panel routes: each id calls several component models and returns one answer. Every component model under every Liberty route resolves to a lab recorded as US-based in the catalog's model-origin table, and the US-AI-models page shows that as a count computed when the page renders rather than as a fixed claim. US-operated provider routes are available for each Liberty id; the default candidate pool also includes operators outside the US, so add provider.jurisdiction="us" when US-operated serving is a requirement.