Hosted AI gateway · Verified August 16, 2026

TrustedRouter vs Martian

Martian invented the commercial LLM router in 2023, then pivoted: today it is an interpretability research lab whose hosted gateway serves 289 models (as of August 2026) with explicit model selection and pass-through pricing. TrustedRouter is a routing gateway with hardware-attested execution on three clouds and no durable prompt logs on realtime inference.

Compare details

8decision dimensions

8official sources

1base URL to try TR

Choose Martian when

Its operating model is the feature.

Choose Martian for pass-through pricing at provider list rates (spot-verified across six models, August 2026), a public SOC 2 Type II and ISO 27001:2022 trust center, and OpenAI Chat Completions, Responses, and Anthropic Messages compatibility with coding-agent integration docs, when explicit model selection is all your application needs.

Choose TrustedRouter when

Privacy needs evidence.

Choose TrustedRouter when you need actual routing, including auto, cheap, fast, ZDR, E2E, and EU routes with provider failover, plus a written no-durable-prompt-logs commitment on realtime inference, enforced by a gateway whose running build you can verify against live attestations on three clouds.

DimensionMartianTrustedRouter

DeploymentHosted proprietary gateway run by Martian Learning, Inc.; nothing to self-hostHosted control plane, source-available (BUSL-1.1), with an attested API path

API surfaceOpenAI Chat Completions and Responses, plus Anthropic MessagesOpenAI Chat Completions and Responses plus Anthropic Messages

Model access289 models across 49 provider prefixes (2026-08-16); no embeddings or image endpoints documented500+ models across ~50 providers: prepaid, BYOK, and direct

RoutingExplicit model selection only; the 2023 automatic router is no longer a public productProvider fallback plus auto, cheap, fast, free, ZDR, E2E, and EU routes with composable privacy and jurisdiction preferences

ObservabilityDashboard usage and request history; martian_metadata tags; no log export documentedMetadata analytics and opt-in external broadcast

Prompt contentToS grants AI-training rights over Customer Materials; no published retention window, opt-out, or ZDRNo durable prompt or output logs on realtime inference; batch is opt-in encrypted retention

VerificationSOC 2 Type II and ISO 27001:2022 on a public Vanta trust center; pen test under NDALive gateway attestation on three clouds, bound to published source and release evidence

BillingPass-through at provider list prices (spot-verified); no published gateway feePrepaid at provider price + 5.5% ($0.01/M floor), or BYOK

What Martian is in 2026

Martian Learning, Inc. launched the first commercial LLM router out of stealth in November 2023, founded in 2022 by UPenn LLM researchers Shriyash Upadhyay and Etan Ginsberg. Its RouterBench paper (March 2024, over 405k inference outcomes) is still a standard reference for routing evaluation, and the code is MIT-licensed. The company's public identity has since split in two. The homepage now presents an interpretability research lab: active MIT-licensed benchmarks (ARES and code-review-benchmark both saw pushes in August 2026) and a $1M interpretability prize (submissions currently closed). The commercial product is the Martian Gateway, a hosted multi-provider API at api.withmartian.com/v1 whose catalog dates from about October 2025.

The fact most evaluators miss: the automatic router is no longer part of the public product. The gateway docs describe explicit model selection only, the live catalog contains no auto or router ids, and the old router site no longer resolves. Routing-style cost optimization now appears at a sister lab, Thesean AI. Third-party reviews of "the Martian router" quoting $20/month describe a product whose pages are dead.

Where Martian is strong

Three things hold up under checking. First, pricing: gateway model prices are pass-through at provider list rates. We spot-checked six models against provider price sheets on August 16, 2026 — anthropic/claude-sonnet-4-5 at $3.00/$15.00 per million tokens, openai/gpt-5 at $1.25/$10.00, google/gemini-2.5-flash at $0.30/$2.50 — all exact matches, with prices refreshed every five minutes.

Second, compliance. Martian runs a public Vanta trust center listing SOC 2 Type II and ISO 27001:2022, penetration test reports and a DPA available on request, and a published subprocessor list with change notifications. That is a serious posture for a company that raised a $9M seed in 2023 and took an Accenture Ventures investment in 2024.

Third, API breadth: OpenAI Chat Completions, the OpenAI Responses API, and Anthropic Messages, with streaming and tool calling, plus integration docs for Claude Code, Cursor, Codex, and Cline. The live catalog held 289 models across 49 provider prefixes on August 16, 2026, each with per-model pricing, cache pricing, and a reliability tier.

Three differences that decide it

Routing exists here. TrustedRouter ships trustedrouter/auto, cheap, fast, and free routes plus privacy-tier routes: trustedrouter/zdr restricts to providers with contractual or policy zero data retention, trustedrouter/e2e to confidential-compute providers (tinfoil, phala), and trustedrouter/eu to EU-focused providers, all with provider failover. The Martian gateway routes nothing automatically; you name the model.

Prompt handling is written down and checkable. Martian's terms of service (updated January 15, 2026) grant it a license to use Customer Materials to train and improve its AI models — deidentification is required only under a separate aggregate-data clause — with no public opt-out, no stated retention window, no ZDR option, and no data-handling page in the gateway docs. Our privacy policy commits realtime inference to no durable prompt or output logs; operational metadata (ids, model, tokens, latency, cost, region, key hash) is retained; batch is a separate opt-in mode with enclave-encrypted retention up to 30 days. The gateway that enforces this runs inside TEEs on GCP, AWS, and Azure, with live attestation endpoints at trust.trustedrouter.com binding the running build to published source and release digests.

The prompt path is public. The Martian Gateway is a proprietary hosted service with no source repository. Every line of TrustedRouter code that touches your prompt is public: source-available under BUSL-1.1, converting to Apache-2.0 four years after each release, with Apache-2.0 or MIT SDKs.

On price, the bases differ, so state them plainly. Martian passes through provider list prices and publishes no gateway fee at all; how the gateway is monetized could not be established from public materials. TrustedRouter bills provider cost plus 5.5% with a $0.01 per million token floor, itemized per model on /models. Per token, Martian is cheaper today. Whether an undisclosed business model stays that way is a question to put to them.

What we do not claim

Our attestation covers the gateway build, not downstream model providers; provider privacy postures are policy and contract tiers, except E2E routes where the provider's own confidential-compute mechanisms apply. Our public repos date from late April 2026 and public benchmark history from June 2026, so our operating track record is short; Martian has been a company since 2022. We publish a security page, DPA, BAA, and subprocessor list, and hold no published SOC 2 or HIPAA certification today. On certifications, Martian is ahead.

Migration reality

Both sides speak all three API shapes, so every Martian surface has a direct target: change the base URL and key, then map creator/model-name ids to TrustedRouter ids on /models (about 550 model routes across 49 providers as of August 16, 2026). Two things do not carry over: the proprietary martian_metadata parameter, and Martian's per-model reliability tiers, for which the nearest equivalents are the provider health table on /status and our monthly frozen benchmark reports. There is no automatic-routing behavior to replicate; adopting trustedrouter/auto or a privacy route afterward is optional and one line.

Migration shape

Start with one real request.

Both sides speak OpenAI Chat Completions, Responses, and Anthropic Messages, so every Martian surface has a direct TrustedRouter target: swap the base URL and key, then map creator/model-name ids to TrustedRouter ids. The martian_metadata parameter and per-model reliability tiers do not carry over, and there is no automatic-routing behavior to replicate.

Keep the first test small, stream the response, and compare output, latency, provider selection, and billed usage before moving production traffic.

Agent setup Run a small eval

TrustedRouter sideOpenAI SDK

from openai import OpenAI

client = OpenAI(
    base_url="https://api.trustedrouter.com/v1",
    api_key="sk-tr-v1-...",
)

with client.chat.completions.create(
    model="trustedrouter/zdr",
    messages=[{"role": "user", "content": "Reply PONG"}],
    stream=True,
) as response:
    for chunk in response:
        print(chunk.choices[0].delta.content or "", end="")

Official evidence

Sources checked August 16, 2026

Report a change

Martian Gateway quickstart (base URL, explicit model choice)https://gateway-docs.withmartian.com/quickstart
Martian Gateway endpoints (Chat Completions, Responses, Messages)https://gateway-docs.withmartian.com/api-reference/endpoints
Live Martian model catalog APIhttps://api.withmartian.com/v1/models
Martian Terms of Service (Jan 15, 2026)https://withmartian.com/terms-of-service
Martian trust center (SOC 2 Type II, ISO 27001:2022)https://trust.withmartian.com
RouterBench paper (arXiv 2403.12031)https://arxiv.org/abs/2403.12031
TechCrunch launch coverage (Nov 15, 2023)https://techcrunch.com/2023/11/15/martians-tool-automatically-switches-between-llms-to-reduce-costs/
Accenture investment press release (Sep 17, 2024)https://newsroom.accenture.com/news/2024/accenture-invests-in-martian-to-bring-dynamic-routing-of-large-language-queries-and-more-effective-ai-systems-to-clients
TrustedRouter monthly benchmark reportsStable first-party route evidence
TrustedRouter live trust recordRelease and attestation evidence

All gateways

TrustedRouter vs Vercel AI GatewayHosted AI gateway
TrustedRouter vs RequestyHosted AI gateway
TrustedRouter vs OpenRouterHosted model marketplace
TrustedRouter vs LiteLLMSelf-hosted AI gateway

Questions

Does Martian still route prompts to the best model automatically?

No. As of August 2026 the gateway docs describe explicit model selection only, the live catalog contains no auto or router model ids, and the old router site no longer resolves. The routing research is real (RouterBench, over 405k inference outcomes), and routing-style cost optimization now appears at sister lab Thesean AI. On TrustedRouter, trustedrouter/auto, cheap, fast, zdr, e2e, and eu are live routes with provider failover.

Who has the stronger compliance certifications?

Martian. It lists SOC 2 Type II and ISO 27001:2022 on a public Vanta trust center, with penetration test reports and a DPA available on request, and a published subprocessor list. TrustedRouter publishes a DPA, BAA, and subprocessor list, and has no published SOC 2 or HIPAA certification today; our verification story is live hardware attestation of the gateway build, which is a different kind of evidence, not a substitute for a certification requirement.

What happens to prompt content on each service?

Martian's terms of service (updated January 15, 2026) grant it a license to use Customer Materials to train and improve its AI models — deidentification is required only under a separate aggregate-data clause — with no public opt-out, stated retention window, or ZDR option, and the gateway docs have no data-handling page. TrustedRouter's realtime inference keeps no durable prompt or output logs; operational metadata (ids, model, tokens, latency, cost, region, key hash) is retained; batch is a separate opt-in mode with enclave-encrypted retention up to 30 days. The gateway enforcing this is attested at trust.trustedrouter.com.