What Tinfoil actually is
Tinfoil Inc. is a San Francisco company, founded in 2024, YC Spring 2025 batch, about five people per its YC profile. It runs open-source models inside hardware secure enclaves it operates: AMD SEV-SNP confidential VMs paired with NVIDIA Hopper and Blackwell GPUs in confidential-compute mode. Since July 2025 every inference request passes through a unified router at inference.tinfoil.sh, itself an attested enclave whose source is public under AGPL-3.0.
Three products. Private Chat is a consumer app at $20/month with a 3M-tokens/hour fair-use cap. The Private Inference API is OpenAI-compatible, billed per token. Tinfoil Containers (GA March 2026) runs any Docker image inside an enclave for $20/month plus usage, with H200 GPUs at $2,000/month and B200 at $5,000/month. As of August 2026 the API catalog is 6 chat models (Kimi K3, GLM-5.2, DeepSeek V4 Flash, Gemma 4 31B, GPT-OSS 120B, Llama 3.3 70B) plus vision, audio, embedding, and safety models, roughly a dozen models total. Open-source models only: Tinfoil's homepage says plainly that closed models like GPT-5 and Claude cannot run there. Containers is the escape hatch: bring your own weights, including proprietary fine-tunes.
Where Tinfoil is genuinely strong
Tinfoil verifies further down the stack than any gateway we compare against, ourselves included. Client SDKs check the router enclave's attestation against measurements published through GitHub and Sigstore; the router then verifies each downstream model enclave, including NVIDIA GPU attestation and model weights on dm-verity read-only volumes checked against Sigstore bundles. The chain proves which code and which weights served a request. Zero retention is architectural and the default on every request; per their security FAQ, prompts and completions live only inside enclaves and are never written to disk or logged. The enclave code is open source (AGPL-3.0), the SDKs Apache-2.0, and the company holds a SOC 2 Type II report covering security, availability, and confidentiality. Pricing is competitive despite confidential-compute overhead: GPT-OSS 120B costs $0.15 input / $0.60 output per million tokens, with prompt-caching discounts since July 2026.
A precision note our earlier page got wrong: the client directly verifies the router enclave only. Model-enclave, GPU, and weight checks happen transitively inside that router; Tinfoil's docs say client SDKs "only need to verify the first enclave in this chain." End-to-end trust therefore rests on the attested router doing the downstream checks. We had described it as client-side verification of everything. Corrected.
How TrustedRouter differs
Depth versus breadth. Tinfoil attests the model runtime. TrustedRouter attests the gateway: api.trustedrouter.com runs a published workload inside a TEE, verifiable live on three clouds (GCP Confidential Space, AWS Nitro Enclaves, Azure Confidential Containers) at trust.trustedrouter.com, with a runnable verifier script. Behind that gateway sit roughly 550 model routes across about 49 providers as of August 2026, including the closed models from Anthropic, OpenAI, and Google that Tinfoil cannot host. Downstream providers are covered by cited policy and contract tiers, not by our attestation. Tinfoil proves more about a narrow catalog; we prove the prompt path into a broad one.
Tinfoil is in our catalog. Tinfoil is one of our upstream providers. The trustedrouter/e2e route (alias confidential) restricts routing to providers running confidential compute with provider-side end-to-end encryption; today that pool is Tinfoil and Phala. The trustedrouter/zdr route enforces a contractual zero-data-retention floor across a wider pool that also includes Anthropic, OpenAI, and Google. One policy can therefore prefer enclave-hosted routes and fall back to policy-ZDR closed models when the open catalog cannot handle a task. Two caveats: on those routes the enclave posture is Tinfoil's or Phala's mechanism, outside our attestation boundary, and through our gateway you do not run Tinfoil's client-side verification.
Logging and pricing bases. Realtime inference keeps no durable prompt or output logs; we retain metadata (request ids, model, token counts, latency, cost, region, API-key hash). Batch is separate and opt-in, with enclave-encrypted retention up to 30 days (privacy). On price, the bases differ, so no single fee comparison is honest: TrustedRouter charges the upstream provider's token price plus 5.5% with a $0.01/M floor (pricing), while Tinfoil sets flat per-token prices for models it hosts itself. Compare final per-million prices on /models against tinfoil.sh/pricing.
What we do not claim
- Our attestation covers the gateway build, not model providers. On non-E2E routes, downstream zero retention is a contractual or policy claim with cited catalog flags, defaulting conservatively to assume-stored.
- Our gateway and control plane are source-available under BUSL-1.1, each release converting to Apache-2.0 four years after publication; the SDKs are Apache-2.0 or MIT. Tinfoil's enclave stack is AGPL-3.0 open source, a real difference in their favor. In both products, the security-critical code that touches your prompt is public.
- We publish no SOC 2 or HIPAA certification today. Tinfoil has SOC 2 Type II; HIPAA in progress per their security FAQ.
- We are young: public repos since late April 2026, public benchmark reports since June 2026, short on-page uptime history (about 72 hours detailed, plus monthly rollups). Tinfoil was founded in 2024 with about five people. Vendor-risk reviewers should weigh early-stage risk on both sides of this page.
Migration reality
Both APIs are OpenAI-compatible, so moving either direction is a base URL and key change. From Tinfoil to TrustedRouter, three things do not carry over: Tinfoil's client-side enclave verification (our SDKs verify our gateway attestation instead, a narrower check), the EHBP encrypted-body protocol, and Tinfoil model ids such as kimi-k3, which you map to catalog ids or to trustedrouter/e2e to keep enclave-hosted execution. Our only step-by-step migration doc today covers OpenRouter, so plan the id mapping by hand. In the other direction, Tinfoil has no closed models, so Claude or GPT traffic has nowhere to land there. Relevant to either plan: Tinfoil has shipped two breaking API changes since July 2025 (router unification, July 2025; mandatory EHBP, October 2025, SDKs v0.10.0+), and its catalog churns monthly; Kimi K2.6 and DeepSeek V4 Pro were removed in July 2026, with Kimi K3 and DeepSeek V4 Flash added in August. Budget for model-id maintenance either way, and see our security page for what our side lets you verify.