MorseSign inOpen Morse

OpenRouter alternatives: seven LLM gateways compared on fees, logging, and routing

The Morse teamPublished Facts checked 13 min read

The best OpenRouter alternatives depend on why you're leaving it. As of November 20, 2026, Morse Private Inference keeps no prompt or reply on any tier, Vercel AI Gateway charges no markup or platform fee on tokens, LiteLLM is an open-source gateway you run on your own servers, and Requesty adds 5% to provider prices.

If you'd rather drop the gateway altogether, DeepInfra, Fireworks AI, and Together AI sell open-weight models by the token from their own hardware, each with its own default for what it keeps.

We read each company's own pricing, privacy, and routing pages on November 20, 2026, along with OpenRouter's FAQ, pricing page, terms (last updated August 31, 2026), and its data collection, zero data retention, provider logging, and provider routing docs. We didn't send traffic through any of them, so every claim below is the company's own. OpenRouter deserves credit before anything else: it bills inference "at the provider's list price on every plan," labels each endpoint's data policy, and lets a single request demand zero data retention.

Morse, which publishes this page, makes a private AI chat app with GPT, Claude, Gemini, DeepSeek, and uncensored models, encrypted so even Morse can't read your chats. For developers it runs Morse Private Inference, a pay-as-you-go, OpenAI-compatible API for 80+ open-weight models with no subscription, and Morse keeps no prompt and no reply on any tier. Anyone can run npx @askmorse/verify inference.trymorse.ai to check that the Morse Enclave runs on genuine AWS Nitro hardware and runs the open-source code at github.com/trymorse/morse-enclave.

The short list

As of November 20, 2026, Morse Private Inference first, then the others in alphabetical order. Each cell is the company's own description, and the pages are listed under Sources at the end.

GatewayWhat it isHow you payPrompts kept by defaultHow you steer routing
Morse Private InferenceA hosted gateway for open-weight modelsEach model's listed price, paid from credits, with no subscriptionNo prompt and no reply, on any tierA privacy floor per request or per key, refused rather than sent lower
DeepInfraA model host you call directlyPer-token prices for many models, others by run time, with a card or prepayment firstHeld in memory during inference and not stored on disk, except Google and Anthropic models, which follow those companies' policiesOne host, so there's nothing to route
Fireworks AIA model host you call directlyPer-token prices drawn from prepaid credits, with no commitment"Zero data retention by default" for open models, outside the Batch and Responses APIsOne host, so there's nothing to route
LiteLLMAn open-source gateway you run yourselfEach provider directly. The open-source proxy is free, and enterprise features take a commercial licenseWhatever you set up, on your own serversYour own retry and fallback rules across deployments
RequestyA hosted gatewayThe provider's price plus 5%Set per organization, with request metadata kept either wayData-policy filters on the models you approve
Together AIA model host you call directlyPer-token prices for each modelPrompts and responses stored until you turn off Prompt storageOne host, so there's nothing to route
Vercel AI GatewayA hosted gatewayThe provider's list price with "no markup and no platform fee on tokens," plus paid add-onsVercel keeps no prompts or responses. Providers follow their own policies unless you ask for zero data retentionProvider order, an only list, model fallbacks, and zero data retention per request

Why do developers look for an OpenRouter alternative?

OpenRouter's own pages point to three reasons: the fee on credits, default settings that let prompts reach providers who may keep them, and a router that chooses the provider for you. OpenRouter documents each of them openly.

The fee on every top-up

OpenRouter charges "a 5.5% ($0.80 minimum) fee when you purchase credits" on its Standard plan, 8% on Business, and 5% on crypto payments, according to its FAQ and pricing page. Inference itself carries no markup, and bringing your own provider keys is free for the first $25,000 of list-price inference each month, then 5%. On a $1,000 top-up, the Standard fee comes to $55. Its terms reserve the right to expire unused credits 365 days after purchase, and its FAQ allows refunds of unused credits only within 24 hours, with platform fees never refunded.

What OpenRouter logs, and what its ZDR leaves out

OpenRouter "does not store your prompts or responses, unless you opt in," its data collection page says. It does store metadata such as token counts and latency for every request, and it "samples a small number of prompts for categorization to power our reporting and model ranking." Opting in to let OpenRouter use your prompts earns "a 1% discount on all model usage," and section 6.2 of its terms then grants it a "worldwide, perpetual, irrevocable" license to that content.

Its zero data retention setting has edges worth knowing before you rely on it. The ZDR guide says it "does not apply to plugins and tools you choose to enable," that self-serve accounts allow Safety Retention until you choose otherwise, and that for private endpoints OpenRouter "does not independently verify the upstream's contractual behavior." Its screening for child sexual abuse material runs on ZDR traffic too.

A router that picks the provider

By default, OpenRouter gives each provider a share of a model's traffic, and "each request picks one at random, weighted by those shares." A failed request moves on to another provider, a switch its FAQ says "happens transparently to the user." The data_collection field starts at allow, which its routing docs define as allowing "providers which store user data non-transiently and may train on it," and OpenRouter calls its data-policy tags "our best knowledge," not "a definitive source." You can narrow all of this with order, only, ignore, zdr, and account settings. Morse vs OpenRouter counts how many companies a default request can land at.

How we compared them

We read each company's pricing page and its privacy or data handling docs on November 20, 2026, and asked the same three questions of each: what it adds to the model's price, what it keeps of a prompt by default, and who decides which company serves a request. We left out Portkey, whose pricing page redirected to Palo Alto Networks' site that day, and any gateway whose pricing we couldn't read.

Morse Private Inference comes first, then the others in alphabetical order. Which gateway fits depends on whether you want a hosted gateway, one you run yourself, or a single host with no gateway at all, and the section after the list sorts them by need.

Morse Private Inference

Morse Private Inference fits developers who want OpenRouter's request format with privacy set as a rule rather than a preference. It's an OpenAI-compatible API for 80+ open-weight models, from DeepSeek, Kimi, and GLM to Qwen, MiniMax, and gpt-oss, and Morse calls every provider from an attested enclave that keeps no prompt or reply and never tells a provider who you are. Moving over from OpenRouter means changing three values: the base URL, the key, and the model. The model ids are mostly the vendor/model names OpenRouter already uses.

Every endpoint runs at one of four tiers: Anonymous, Private, TEE, and E2EE. Set provider.min_privacy on a request, or a floor on the key, and Morse serves the request at that tier or a more private one, or refuses it with 400 privacy_floor_unmet, checked before it calls a provider or charges you. It never sends a request lower. Morse also reads OpenRouter's zdr, data_collection, max_price, and allow_fallbacks fields, and picks the provider within a tier itself, with every provider in a tier charging the same, so you decide how private a request must be rather than which company serves it.

For billing, Morse records the model, the token counts, the cost, the time, and the key. It never records the prompt, the reply, or the user, safety_identifier, and prompt_cache_key fields, which it drops so a provider can't link one person's requests. Run npx @askmorse/verify inference.trymorse.ai before you deploy to check the enclave's AWS Nitro attestation, its place on Morse's signed list of images, and its source on GitHub. A plain request sent with a key is read by Morse's relay on its way to the enclave, so for code that should send nothing the relay can read, use attested() from the same package: it runs those checks itself and seals each request to the keys that passed. Credits pay each model's listed price, one credit is $0.01, and your first top-up in the console gets 50% extra. The docs cover streaming, tools, and errors, and every price is on the models page.

DeepInfra

DeepInfra fits developers who mostly call open models and want one host's per-token prices with no gateway in between. Its pricing page, read November 20, 2026, listed DeepSeek V4.1 Flash at $0.14 per million input tokens and $0.42 per million output tokens (marked 30% off), GLM-5.3 at $0.5625 and $2.50 (marked 38% off), and Kimi-K3 at $2.85 and $14.25. Some models are billed by execution time instead, and the page says "You have to add a card or pre-pay or you won't be able to use our services."

Its data privacy page says "we do not store the data you submit to our APIs on disk" and that it holds it "in memory during the inference process." DeepInfra says it doesn't train on what you send "except when using Google or Anthropic models," which follow those companies' own retention, and that it logs a request ID, the cost, and sampling parameters rather than content. Bulk inference jobs and image outputs are kept for a short period. Calling a single host means no automatic move to another company when it's busy, so build your own retries.

Fireworks AI

Fireworks AI fits teams that want zero data retention without having to ask for it. Its pricing page says "Zero data retention by default," and its data handling docs say Fireworks "does not log or store prompt or generation data for open models, without explicit user opt-in." Two services fall outside that: the Batch API, and the Responses API, which stores conversations by default (store=True) for up to 30 days unless you set store=False. Fireworks logs metadata such as the number of tokens in a request.

Every serverless request bills on input, cached input, and output tokens from prepaid credits, with no commitment. On November 20, 2026, its standard prices included DeepSeek V4.1 Flash at $0.30 and $1.20 per million input and output tokens, GLM 5.3 Flash at $0.15 and $0.50, GPT OSS 120B at $0.15 and $0.60, and Kimi K3 at $3.00 and $15.00. The same page lists SOC 2 Type II, ISO 27001, and HIPAA with a BAA.

LiteLLM

LiteLLM fits developers who want the gateway inside their own walls. Its docs describe the proxy as "a self-hosted OpenAI-compatible gateway" for "100+ LLMs," with virtual keys, cost tracking per key, team, and user, an admin UI, and "Built-in retry / fallback logic across multiple deployments." Every response follows OpenAI's Chat Completions format whatever the provider, so switching providers doesn't change how you parse replies.

No third company sits between your code and each provider, and the open-source version, whose docs list an MIT license, costs nothing to LiteLLM. Enterprise features fall under a separate "LiteLLM Commercial License." In exchange you hold an account and a key with every provider, set each one's retention and training terms yourself, and keep the proxy running. Logs go only where you send them, through callbacks to tools such as Langfuse and MLflow. Its GitHub repository showed 63.6k stars on November 20, 2026.

Requesty

Requesty fits teams that want a hosted gateway with governance built in. Its pricing page offers "600+ models behind one OpenAI-compatible endpoint, with routing, caching and governance included," and charges "what your application spends, plus 5%," so a model listed at $10 per million tokens costs $10.50. A free plan gives "200 requests per day" on free models with no card, and EU data residency comes on every plan.

Its data privacy docs open by saying prompts and responses are "never stored unless you explicitly enable logging," but new keys inherit the organization's default_logging setting, and the docs' default configuration example sets it to true, so check yours. Organization-wide zero data retention is "a one-way operation" that can't be switched off later, and it covers only Requesty's own storage: in the docs' words, "ZDR controls what Requesty stores," while providers are governed separately by data-policy filters on approved models. Token counts, costs, timestamps, and routing details are kept either way, never the content.

Together AI

Together AI fits developers who want a large open-model host and don't mind changing one setting on day one. Its pricing page bills serverless models per million tokens: on November 20, 2026, DeepSeek V4.1 Flash cost $0.30 for input and $1.20 for output, GLM-5.3 $1.40 and $4.40, and Kimi K3 $3.00 and $15.00.

Its zero data retention docs say "ZDR is not enabled by default," and that "Together stores the prompts you send and the responses models return, and may use them for product improvements." To stop that, turn off Prompt storage under Privacy in your organization settings. Training on your data is "a separate opt-in and is off by default," and usage metadata such as token counts and model IDs stays either way.

Vercel AI Gateway

Vercel AI Gateway is the closest like-for-like swap for OpenRouter's hosted model. Its pricing docs say "AI Gateway charges no markup and no platform fee on tokens," including when you bring your own key, though payment processing fees can apply. It's a hosted API at ai-gateway.vercel.sh that speaks OpenAI Chat Completions, OpenAI Responses, and Anthropic Messages, and it needs a Vercel account but no Vercel deployment. A free tier covers a subset of models with a monthly credit.

Vercel says the gateway "does not retain the prompts or responses in your requests," and it records metadata such as the model, provider, tokens, and cost. By default it "does not route based on the data retention policy of providers," so zero data retention is something you ask for: free per request with zeroDataRetention: true on a paid account, or $0.10 per 1,000 requests team-wide. Its ZDR docs say the setting doesn't cover tools a provider runs, such as a provider's own web search, and that Claude Fable models aren't ZDR on any provider. Requests made with your own keys fall back to Vercel's keys by default, and turning that off costs $0.20 per 1,000 requests.

Which OpenRouter alternative fits your project?

  • Choose Morse Private Inference if you want an OpenAI-compatible API that keeps no prompt or reply on any tier, a privacy floor that refuses rather than downgrades, and an enclave you can check before you send.
  • Choose Vercel AI Gateway if you want a hosted gateway at providers' list prices with no fee on tokens, and you'll switch on zero data retention where you need it.
  • Choose LiteLLM if you'd rather run the gateway yourself and hold your own terms with each provider.
  • Choose Requesty if you want routing, caching, and EU data residency on every plan, and 5% on top is fine.
  • Choose one of the three direct hosts above if you mostly call open models and want one host's prices, after checking what each keeps by default.

How Morse Private Inference keeps prompts off the record

On most gateways, what happens to a prompt after it leaves your server depends on settings you have to remember for each account, key, and request. Morse Private Inference starts from the strict end, so there's no setting to forget.

  • No prompt and no reply, on any tier. There's nothing to switch on. The usage ledger holds the model, the tokens, the cost, the time, and the key, never the content, and Morse keeps each request's billing record for 7 days, then only daily totals.
  • A floor that holds. Set min_privacy to private, tee, or e2ee on a request or on a key, and a request Morse can't serve at that floor is refused, never sent lower.
  • No name on your requests. A provider sees Morse, never your account, your card, or your address.
  • A check you can run yourself. The verifier checks the enclave's AWS Nitro attestation and its open-source code, and attested() seals each request so only an attested enclave can open it. Verify it yourself walks through each step, and what is a trusted execution environment explains the hardware.

The same enclave carries every message in the Morse chat app, where your history stays on your device, encrypted, and the zero data retention questions in this article are covered in what is zero data retention.

Read the quickstart in the Morse docs, or make a free account and create your first key.

Morse publishes this comparison. We've tried to keep it fair; if something's wrong or out of date, email hello@trymorse.ai.

OpenRouter is a trademark of OpenRouter, Inc. Morse is not affiliated with or endorsed by OpenRouter, Inc.

Sources

All read November 20, 2026.

Common questions

What is the best alternative to OpenRouter?

It depends on why you're leaving. Vercel AI Gateway is the closest hosted swap and adds no fee to token prices, LiteLLM suits teams that want to run the gateway themselves, and a direct model host is usually cheapest if you only call open models.

Is there a free OpenRouter alternative?

LiteLLM's open-source proxy costs nothing beyond your own servers and each provider's prices. Vercel AI Gateway gives every team a monthly free credit on a subset of models, and Requesty's free plan allows 200 requests a day on free models with no card.

Does OpenRouter charge a fee?

Yes, when you buy credits: 5.5% with a $0.80 minimum on Standard, 8% on Business, and 5% on crypto payments. Inference itself is billed at the provider's list price, and bringing your own keys is free up to $25,000 of list-price inference a month, then 5%.

Does OpenRouter log your prompts?

Not unless you opt in, its data collection page says. It does store metadata such as token counts and latency for every request, and it samples a small number of prompts for categorization to power its reporting and model rankings.

Is Vercel AI Gateway cheaper than OpenRouter?

On tokens, yes. Both pass through the provider's list price, but Vercel adds no fee while OpenRouter adds 5.5% when you top up on Standard. Vercel charges for some add-ons, such as team-wide zero data retention, a team-wide provider allowlist, and regional endpoints.

What is the difference between an LLM gateway and an inference provider?

An inference provider runs the models on its own hardware and sells them by the token. A gateway gives you one API and one bill in front of many providers, and decides or lets you decide which one serves each request.

Is LiteLLM more private than OpenRouter?

It removes one company from the path, since the proxy runs on your own servers. Each provider still receives your prompts under the account and terms you hold with it, so its retention and training settings are yours to check.

Does zero data retention mean nothing is kept?

Not on every gateway. OpenRouter's and Vercel's ZDR both leave out tools a provider runs and don't count short-lived prompt caches, OpenRouter's self-serve accounts allow Safety Retention unless you choose otherwise, and Requesty's ZDR covers only what Requesty itself stores.

How do I switch from OpenRouter to Morse Private Inference?

Change the base URL to https://inference.trymorse.ai/v1, use a Morse key, and keep the model id, since Morse's ids are mostly the vendor/model names OpenRouter uses. Add provider.min_privacy to set how private each request must be.

Does Morse keep the prompts I send through its API?

No. Morse's usage ledger records the model, the tokens, the cost, the time, and which of your keys made the request, never the content, and Morse keeps each request's billing record for 7 days, then only daily totals.

Ask what you actually want to ask.

80+ models, characters, and images, with every chat encrypted and kept on your device. Free to start.

Start a private chat

Read next