The best OpenRouter alternatives, compared honestly
OpenRouter is still the easiest way to reach 500-plus models behind one API key, and it does something none of its rivals do: it publishes a machine-readable, per-provider table of who retains your prompts and who might train on them. Read that table and a different problem appears. On 31 August 2026 it listed 83 providers, of which 46 are zero retention, 11 retain for 30 days, one (Google AI Studio) for 55 days, and 25 retain prompts for a period nobody states. The docs then say plainly that OpenRouter does not have routing rules based on retention, while default routing prioritises price. This page compares six alternatives on the only axis that matters once you have noticed that: what the serving provider actually does with your prompt, and what it costs to make that answer enforceable.
The best OpenRouter alternative depends on which part of the data question is yours:
- The routing layer must never hold your prompt → LiteLLM: MIT-licensed self-hosted gateway, 57.6k stars, 0 dollars forever, and no message content persisted by default.
- You want a managed provider with the tightest published default → Fireworks AI: no prompt or generation logging for open models without explicit opt-in, prompts in volatile memory for the request only.
- You want price parity plus a zero-retention switch → Together AI: gpt-oss-120B at 0.15 dollars per 1M input tokens, the same rate OpenRouter shows, but retention is on until you flip the switch and the switch is not retroactive.
- You need the contract in your own name → the model vendor's own API: real terms and named zero-retention programmes, though first-party endpoints are often the retaining ones (Anthropic 30 days, Google AI Studio 55 days).
Why teams look elsewhere
What pushes teams off OpenRouter
Nothing here is a claim that OpenRouter got worse at its job: model coverage, failover and unified billing all still work, and every criticism below is only possible because OpenRouter publishes the underlying data. What changed is that legal and security reviews started reading the provider table, the plan table and the routing defaults together, and those three documents do not point the same way.
Retention is documented, not routed on
The Provider Logging docs state that if you opt out of training, OpenRouter will not route to providers that train, and in the next section that OpenRouter does not have routing rules that change based on data retention policies. So the training flag is enforced for you and the retention column is, by default, information only. On 31 August 2026 that column covered 37 of 83 providers retaining something, 25 of them for an unstated period.
The control that fixes it is a paid feature
On the pricing page's three-column plan table, Data Policy-Based Routing is marked unavailable on Free and available on Pay-as-you-go. Org-wide Managed Policy Enforcement is Enterprise only, alongside SSO/SAML and contractual SLAs. So the free tier that most evaluations start on is exactly the tier where the retention table cannot be turned into a routing rule.
Default routing is price-first, and cheap is often not zero-retention
The routing docs say the default is to load balance across providers, prioritising price. Cross-referencing OpenRouter's endpoint API against its own zero-retention endpoint list on 31 August 2026: deepseek/deepseek-v4-pro-0813 had 16 endpoints, 11 on the zero-retention list, and the three cheapest input rates were all off that list (1.1154, 1.1220 and 1.1220 dollars per 1M). The honest counterexample: openai/gpt-oss-120b had 20 of 20 endpoints on it.
Zero retention stops at the model call
Straight from the ZDR docs: enforcement only applies to provider routing for inference requests and does not apply to plugins and tools you enable, such as web search. OpenRouter has also taken the position that in-memory caching of prompts is not retaining data, so 45 of the 821 zero-retention endpoints report implicit caching support. Both are defensible calls; both are things a security reviewer wants to know before signing.
The 1 percent discount is paid for in licence terms
Opting in to OpenRouter's use of your inputs and outputs buys a 1 percent discount on all model usage. The Terms of Service, last updated 29 July 2026, spell out the price in section 6.2: a worldwide, perpetual, irrevocable, non-exclusive, royalty-free, fully paid, sublicensable licence to store and commercially use those anonymized inputs and outputs. A separate perpetual licence covers anonymized input categorization regardless of your logging preference.
Small top-ups pay far more than 5.5 percent
The credit purchase fee is 5.5 percent with a 0.80 dollar minimum by card, 5 percent by crypto. The minimum is regressive: a 5 dollar top-up pays 16 percent, a 10 dollar top-up pays 8 percent, and 5.5 percent only becomes the binding rate above 14.55 dollars. Credits are refundable within 24 hours only, and the platform fee itself is never refundable.
The shortlist
6 OpenRouter alternatives worth evaluating
A router does two separable jobs: it aggregates providers (one key, automatic failover, one invoice) and it chooses which provider serves each token. Only the second job creates the data question, so the replacements split the same way: run the aggregation yourself and pick providers deliberately, or drop aggregation and buy from one provider whose published policy you accept. This ranking optimises for how enforceable and how legible each provider's prompt handling is, not for the lowest rate per million tokens. Every pick lists one honest strength and one real weakness.
The only option here where nobody else's server sees the prompt before the model does, and you still keep the two things people actually use a router for: one OpenAI-shaped API across providers, and automatic fallbacks. The proxy does not persist message content by default; turn_off_message_logging, a per-request no-log flag and beta header-level redaction let you keep spend tracking without keeping text. Licence, precisely: the repository LICENSE is MIT (copyright 2023 Berri AI) with everything under enterprise/ licensed separately, which is why GitHub's API reports NOASSERTION and not MIT. Self-hosted tier is 0 dollars, free forever; 57,642 stars and a push on 31 August 2026. Weakness: you inherit each provider's default policy with no negotiated agreements, you carry the ops and the uptime, and the enterprise tier has no published price at all (annual, sized to gateway request capacity, quote only).
The clearest published default of any managed provider on this page: Fireworks does not log or store prompt or generation data for any open models without explicit user opt-in, and states that prompt and generation data exist only in volatile memory for the duration of the request, with cached prompts and KV caches held in volatile memory for several minutes. Metadata such as token counts is logged to run the service. Start with 1 dollar of free credits and no monthly minimum; dedicated capacity is per GPU-hour, at 8.00 dollars per hour for an H100 80GB and 15.00 dollars per hour for a B300 288GB from 1 September onwards. Weakness: a narrower catalog than a router, and the serverless per-token rate card lives in the docs rather than on the pricing page, so headline comparisons are awkward.
Price parity is real and checkable: Together lists gpt-oss-120B at 0.15 dollars input and 0.60 dollars output per 1M tokens, and that is exactly what OpenRouter shows for the Together endpoint, which is the cleanest confirmation available that the router does not mark up inference. The privacy policy, last updated 17 December 2025, says Together does not use data collected from you to train its models without explicit opt-in, and a zero-retention toggle under Privacy and Security removes storage and retention entirely. Other pay-as-you-go rates per 1M tokens: gpt-oss-20B 0.05 and 0.20; DeepSeek V4 Flash 0731 0.14 and 0.28; Llama 3.3 70B 1.04 both ways. Weakness: retention is on by default until you flip that switch, and Together states the setting applies only from the moment you enable it and does not affect data processed before.
The strongest wording on this page comes from Groq's Services Agreement: Groq is not permitted to use inputs or outputs for training or fine-tuning any model unless explicitly granted permission, and does not access, use, store or retain inputs or outputs except as necessary to provide the cloud services, with customer data deleted within 30 days of termination. Note that the general privacy policy, effective 12 November 2025, deliberately does not cover GroqCloud customer data and points at that agreement instead, so the agreement is the document to read. Weakness: no published rate card. Checked on 31 August 2026, groq.com/pricing redirects to the homepage and both console pricing paths return 404; the plan names Free, Developer and Enterprise appear in the docs with no prices attached. The only public per-token figure for Groq is the one OpenRouter prints for its endpoints.
Sometimes the answer is not another gateway. Going straight to OpenAI, Anthropic or Google gets you a data processing agreement you signed, a named zero-retention programme you can be approved for, and a support path that does not run through a third party. OpenAI does not train on API data unless you opt in, keeps abuse-monitoring logs up to 30 days, and offers zero data retention subject to prior approval. Anthropic states it will not use inputs or outputs from commercial products to train models by default. Weakness: one integration and one contract per vendor, no failover between them, no consolidated invoice, and going direct does not by itself buy zero retention. OpenRouter's own table puts Anthropic at 30 days and OpenAI at an unstated period, and OpenAI's zero-retention programme excludes conversations, assistants, threads and vector stores, which hold state until deleted.
Replicate publishes the sharpest number in this whole comparison: input and output, including any files, are automatically deleted after an hour for any prediction created through the API. No other vendor here commits to a default deletion window in a single sentence. Billing is by run time rather than tokens for most public models, at 0.000225 dollars per second for an Nvidia T4 (0.81 dollars per hour), 0.001400 for an A100 80GB (5.04 per hour) and 0.001525 for an H100 (5.49 per hour), with some models billed per input and output instead: DeepSeek-R1 at 3.75 dollars per 1M input tokens, FLUX-1.1-Pro at 0.04 dollars per output image. No monthly fee or minimum is published. Weakness: it is not an LLM router. Per-second hardware billing punishes idle time and cold starts, chat-completions parity is partial, and there is no cross-provider failover to inherit.
Also considered and left off deliberately: Vercel AI Gateway, Portkey and similar hosted gateways are credible OpenRouter substitutes on features, but they reproduce the same structure this page is examining, a third party choosing your provider, so adding them lengthens the list without moving the axis. None was dropped for a factual defect.
Side by side
OpenRouter alternatives compared
Prices as of August 2026, from each vendor's own pricing page, with the billing cycle stated in every cell. The column that decides most migrations is Default prompt handling: not the rate, but whether the party serving your tokens keeps them, for how long, and whether you can make the answer binding. An empty rate is itself a finding, so it is written out rather than guessed. OpenRouter's own row is included as the baseline.
| Tool | Type | Default prompt handling | Enforceable zero retention | Fee on top of inference | Self-host / licence | Best for |
|---|---|---|---|---|---|---|
| OpenRouter | Hosted router, 80+ providers | Per provider, published: 46 of 83 zero retention, 11 at 30 days, 1 at 55 days, 25 unstated | ✓ global, per model group, per request: Pay-as-you-go and up, not Free | 5.5% per credit purchase, min 0.80 dollars; BYOK 5% above 25,000 dollars/month of list-price inference; no markup on inference | ✗ proprietary SaaS | Widest catalog and failover, if the retention column is acceptable |
| LiteLLM | Self-hosted gateway | Gateway persists no message content by default; each upstream provider's own policy still applies | ✓ by choosing upstreams yourself, per key and per request | 0 dollars self-hosted; enterprise price not published, sold annually by gateway request capacity | ✓ MIT with an enterprise/ carve-out (GitHub reports NOASSERTION) |
Teams that must keep the routing layer in their own perimeter |
| Fireworks AI | Managed inference provider | No prompt or generation logging for open models without opt-in; volatile memory only, caches for minutes | ✓ it is the documented default, not a toggle | None stated; 1 dollar free credits, no monthly minimum, GPU-hour rates for dedicated capacity | ✗ proprietary SaaS | The strongest managed default with no configuration to forget |
| Together AI | Managed inference provider | Retained by default; no training without explicit opt-in | ✓ account toggle, but forward-looking only, not retroactive | None stated; no plan fee, no stated minimum, per-1M-token rates | ✗ proprietary SaaS | Matching OpenRouter's rate while owning the provider relationship |
| Groq | Managed inference provider | Contractually no training without permission; no retention beyond delivering the service | ✓ in the Services Agreement, 30-day deletion after termination | No published rate (checked 31 Aug 2026: pricing page 308 to homepage, console paths 404) | ✗ proprietary SaaS | Latency-sensitive work where you can negotiate a quote |
| Vendor's own API | First-party model API | No training on API data by default; abuse logs up to 30 days at OpenAI, 30 days for Anthropic per OpenRouter's table | ✓ named programmes, but by prior approval and with endpoint exclusions | None; list price per 1M tokens, per vendor, monthly invoice | ✗ proprietary SaaS | Regulated buyers who need the DPA in their own name |
| Replicate | Model hosting and prediction API | Input and output deleted after 1 hour for API predictions | ✗ no separate programme; the 1-hour default is the commitment | None; billed per second of hardware, some models per input and output, no monthly minimum | ✗ proprietary SaaS | Image, video and custom models rather than chat routing |
Numbers that do not fit in cells. OpenRouter's zero-retention endpoint list held 821 endpoints across 292 models and 50 providers on 31 August 2026, of which 45 support implicit caching; its full catalog was 396 models. Only 4 of 83 providers were flagged as possibly training (DeepSeek, Liquid, NVIDIA, Thinking Machines) and 11 of 83 require your end-user identifiers, including Anthropic, OpenAI, Google Vertex, Amazon Bedrock and Meta. Free-model limits are 20 requests per minute and 50 requests per day below 10 dollars of lifetime credits purchased, rising to 1,000 per day at 10 dollars or more, with failed attempts counting toward the quota. Together's dedicated endpoints run 5.49 dollars per hour for an HGX H100 and 8.99 dollars per hour for an HGX B200, both on demand. Pricing and data policies change often; check each vendor for current terms. Compiled August 2026.
Official pages: OpenRouter provider logging · OpenRouter ZDR · OpenRouter pricing · OpenRouter terms · LiteLLM pricing · Fireworks data handling · Fireworks pricing · Together pricing · Together privacy · Groq services agreement · Replicate pricing · OpenAI data usage
A fair call
When OpenRouter is still the right choice
Migrating off a router is cheap in code and expensive in coverage. The API shape is OpenAI-compatible on every option here, so the client change is small; what you lose is the failover across 80-plus providers, the single invoice, and a catalog of 396 models you did not have to negotiate for one at a time. Worth saying plainly: every criticism on this page was assembled out of documents OpenRouter itself publishes, and there is a real case that a router with a published retention table is safer than a provider with no table at all.
OpenRouter is still right if…
- Your models happen to be fully covered:
openai/gpt-oss-120bhad all 20 endpoints on the zero-retention list on 31 August 2026, which makes the entire retention debate moot for that workload. - You are on Pay-as-you-go and will actually switch on data-policy routing, per request or account-wide. The mechanism exists, is documented, and works at model-group granularity.
- You value not being billed for failures. Failed and fallback attempts are not charged, and zero completion insurance is enabled automatically on every account with no configuration, refunding prompt, completion and reasoning tokens even when the underlying provider charges for prompt processing.
- You want provider terms better than the provider's own default. Together retains by default on its own platform yet appears as zero retention on OpenRouter, which is exactly the negotiated-agreement effect the ZDR docs describe.
- You buy credits in real amounts. The 5.5 percent fee is the whole markup, inference is passed through at list price (we checked one endpoint against the provider's own page and it matched to the cent), and BYOK is free up to 25,000 dollars per month of list-price inference.
- You would rather have a conservative unknown: where OpenRouter cannot establish a provider's policy, it assumes the endpoint both retains and trains.
Look elsewhere if…
- The routing layer itself is the problem, not the providers: LiteLLM, self-hosted, keeps the decision and the traffic inside your perimeter.
- You are staying on the free tier: data-policy routing is not available there, so the retention table stays informational.
- You want one provider whose default needs no configuration: Fireworks AI publishes no logging of prompts or generations for open models without opt-in.
- Someone will ask which sub-processor served a given request, in writing: go to Together AI, Groq or the vendor's own API and hold one contract.
- Your compliance scope covers web search and other plugins: zero-retention enforcement is documented as not extending to them.
- You need the data processing agreement and any zero-retention approval in your own name: that is the first-party API, whatever it costs in failover.
- You top up in 5 dollar increments: at the 0.80 dollar minimum that is a 16 percent fee, so either buy larger credits or price a provider with no purchase fee.
Common questions
Common questions about OpenRouter alternatives
What is the best OpenRouter alternative in 2026?
If the reason you are looking is data handling, LiteLLM is the strongest answer: it is a self-hosted gateway under an MIT licence with an enterprise/ carve-out, so the routing layer itself runs on your infrastructure and never holds your prompt on someone else's server. You keep one key and automatic fallbacks. If you want a managed provider instead, Fireworks AI publishes the tightest default: no prompt or generation logging for open models without explicit opt-in.
Does OpenRouter train on my prompts?
OpenRouter itself says it does not: its privacy policy, last updated 6 July 2026, states that OpenRouter does not use your inputs or outputs for model training, and prompts are not retained unless you opt in to logging. The question is the provider behind the model. On 31 August 2026 OpenRouter's own provider table flagged 4 of 83 providers as possibly training on prompts: DeepSeek, Liquid, NVIDIA and Thinking Machines. If you opt out of training in account settings, OpenRouter will not route to them.
Which providers on OpenRouter retain my prompts?
Counting OpenRouter's own provider data on 31 August 2026: 46 of 83 providers are zero retention, 11 retain for 30 days, Google AI Studio retains for 55 days, and 25 retain prompts for a period that is not stated anywhere. So 37 of 83 retain something. Anthropic, Cohere, Meta and Mistral sit in the 30-day group; OpenAI sits in the unknown-period group. Eleven providers also require your end-user identifiers.
Can I force zero data retention on OpenRouter?
Yes, but not on the free plan. Data Policy-Based Routing is marked unavailable on Free and available on Pay-as-you-go, and org-wide Managed Policy Enforcement is Enterprise only. Once enabled you can enforce zero retention globally, per model group, per guardrail, or per request. Two limits are worth reading: enforcement covers provider routing for inference only, not plugins and tools such as web search, and OpenRouter treats in-memory prompt caching as not retention.
How much does OpenRouter cost in 2026?
Inference is passed through at provider list price with no markup, which we spot-checked: Together's gpt-oss-120B is 0.15 dollars per million input tokens on both together.ai and OpenRouter. The fee is on credit purchases: 5.5 percent with a 0.80 dollar minimum by card, 5 percent by crypto. The minimum is the catch, so a 5 dollar top-up pays 16 percent and 5.5 percent only becomes the real rate above 14.55 dollars. Bring-your-own-key costs 5 percent above 25,000 dollars per month of list-price inference.
Is there a truly open-source OpenRouter alternative?
LiteLLM is the one to look at, and the licence needs stating precisely. The repository LICENSE is MIT, copyright 2023 Berri AI, with a carve-out: everything under the enterprise/ directory is licensed separately. Because of that carve-out GitHub's own API reports the licence as NOASSERTION rather than MIT. The self-hosted proxy is free forever at 0 dollars; the enterprise tier has no published price and is sold annually, sized to gateway request capacity rather than per token.
Is going direct to OpenAI or Anthropic more private than a router?
Not automatically. Going direct gets you a contract in your own name, which matters, but the first-party endpoints are often the retaining ones: OpenRouter's table lists Anthropic at 30 days, Google AI Studio at 55 days and OpenAI at an unstated period. OpenAI's zero data retention needs prior approval and does not extend to conversations, assistants, threads or vector stores, which keep state until deleted. Anthropic says it does not train on commercial inputs by default, but feedback you submit stores the conversation for up to 5 years.