OpenRouter Alternatives and Competitors (2026): 11 Providers and 6 AI Gateways Compared on Price, Rate Limits, and Compliance

What OpenRouter Actually Charges

OpenRouter does not mark up inference. You pay the same token rate you would pay the underlying provider. The money OpenRouter makes is the 5.5% fee on credit purchases (and a 5% fee on bring-your-own-key usage above a free allowance of $25,000 in list-price inference per month on pay-as-you-go, $200,000 on Enterprise). At scale on the homepage: 400+ active models, 60+ providers, 100T monthly tokens, 8M+ users, 250k+ apps.

The free tier is tight. Without credits you get 50 requests per day. After buying at least $10 in credits, 1,000 requests per day. Free model variants (the :free suffix) are capped at 20 requests per minute. Unused credits can expire one year after purchase. None of this is a markup, but it is the cost of the convenience layer.

So the decision is not "is OpenRouter overpriced?" It is: do you need unified billing and automatic fallback across 60+ providers, or do you want the lowest direct price and full control? If the latter, you have two moves: go direct to a single cheaper provider, or self-host LiteLLM to keep the unified API with no per-request fee.

OpenRouter Limitations in Production

Most teams start looking at OpenRouter alternatives and competitors after one of four limits shows up in production.

• Fees at volume. Card credit purchases carry 5.5% ($0.80 minimum) and crypto purchases 5%. Bring-your-own-key traffic is fee-free up to $25,000 a month of list-price inference on pay-as-you-go, then 5%.
• Free-model rate limits. Free models are capped at 50 requests a day, or 1,000 a day once you have bought at least $10 in credits. OpenRouter's own FAQ says they are usually not suitable for production use.
• Hosted only. There is no self-hosted OpenRouter. Requests and keys pass through its infrastructure, so teams with VPC or data-residency rules move to a gateway they run (LiteLLM, Bifrost) or an EU-hosted one (Opper).
• It does not run the GPUs. Latency, quantization, and uptime depend on the provider behind each route. Going direct puts you on that provider's own rate limits and support.

Source: openrouter.ai/docs/faq, fetched October 12, 2026.

Same Model, Five Prices

OpenRouter routes the same open model to multiple providers, each with its own direct rate. Going direct to the cheapest one is the largest lever, larger than the 5.5% credit fee. Prices below are per 1M tokens, input / output, from each provider's pricing page on June 9, 2026.

Provider-by-Provider

The serverless providers OpenRouter routes to fall into three groups: per-token open-model shops (DeepInfra, Fireworks, Together, Novita), speed specialists (Groq, Cerebras), and run-your-own-code platforms that bill by GPU-second (Modal, Replicate, Baseten). A fourth option is a different gateway in front of the same providers: Opper does that from the EU.

DeepInfra: cheapest per-token, no free tier

DeepInfra is the price leader on open models: DeepSeek V4 Pro $1.30/$2.60, V4 Flash $0.10/$0.20, Kimi K2.6 $0.75/$3.50, GLM-5.1 $1.05/$3.50, Qwen3-235B-A22B-Instruct $0.09/$0.10, Llama 3.1 8B $0.02/$0.05. It also resells Anthropic (Sonnet 4.6 at $3/$15, Opus 4.8 at $5/$25). Limit is 200 concurrent requests per account, postpaid with no free tier and no minimum. SOC 2 and ISO 27001 certified, zero retention on inference (only metadata logged), with HIPAA/GDPR measures. See the Fireworks vs DeepInfra breakdown.

Fireworks: predictable rate limits, fine-tuning

Fireworks publishes clear ceilings: 6,000 RPM with a payment method (10 RPM without), gated by monthly spend tiers ($50, $500, $5,000, $50,000). Serverless covers Kimi K2.6 $0.95/$4.00, DeepSeek V4 Flash $0.14/$0.28, GLM 5.1 $1.40/$4.40, GPT-OSS 120B $0.15/$0.60, MiniMax 2.7 $0.30/$1.20. Cached input gets 50% off and batch runs at 50% of serverless. Fine-tuning is priced per 1M training tokens, from $0.50 (LoRA SFT, up to 16B) to $20 (full SFT, >300B). SOC 2 Type II, HIPAA, zero data retention on open models. Fireworks alternatives.

Together: serverless plus GPU clusters

Together spans serverless (DeepSeek V4 Pro $2.10/$4.40, GLM-5.1 $1.40/$4.40, Qwen3.5-397B-A17B $0.60/$3.60, GPT-OSS-120B $0.15/$0.60, Llama 3.3 70B $1.04/$1.04) and rentable GPU clusters (HGX H100 $5.49/hr on-demand, reserved from $3.99/hr). Rate limits are dynamic per model and scale with sustained traffic; for a fixed guarantee Together points you to a dedicated endpoint. Batch API gives up to 50% off on selected models. SOC 2 Type 2. Together AI alternatives.

Groq: throughput specialist

Groq optimizes for tokens per second on a fixed catalog: GPT-OSS 120B at 500 tok/s ($0.15/$0.60), GPT-OSS 20B at 1,000 tok/s ($0.075/$0.30), Llama 3.3 70B at 394 tok/s ($0.59/$0.79), Qwen3 32B at 662 tok/s ($0.29/$0.59), Kimi K2 Instruct $1.00/$3.00 (cached input $0.50). Free plan exists (e.g. 30 RPM, 14.4K requests/day on llama-3.1-8b-instant). Batch is 50% off with a 24h-7day window. SOC 2 Type II; HIPAA via BAA with preview/beta and compound systems excluded. Compare Fireworks vs Groq.

Opper: EU-hosted gateway, 3% platform fee

Opper is a hosted router like OpenRouter, run from the EU. One API in front of 700+ models across 30+ providers, drop-in compatible with the OpenAI, Anthropic, and Google SDKs (base URL https://api.opper.ai/v3/compat). The platform fee is 3% on the Gateway tier against OpenRouter's 5.5% credit fee, and bring-your-own-key is free there; the Control Plane tier is 5.5% and adds scoring, routing rules, PII masking, and policy enforcement. The platform runs only in AWS Stockholm: traces, routing tables, and metadata stay in the EU, and prompts and responses are not written to disk unless you add a retention rule (30 days maximum). Comply rules reject a call before it leaves the platform if it would route outside your allowed providers, regions, or zero-retention set, and one sub-processor covers every model, so turning on a new provider needs no new DPA amendment. Morph serves DeepSeek, MiniMax, and Qwen through it.

Modal, Replicate, Baseten: run your own model

These bill for compute, not tokens, which fits custom or fine-tuned models OpenRouter does not host. Modal bills per second (H100 ~$3.95/hr, B200 ~$6.25/hr), boots containers in ~1s, scales to zero, and gives $30/month free credits on Starter. Replicate scales to zero by default and only charges running prediction time (H100 $5.49/hr, A100 80GB $5.04/hr). Baseten bills per minute (H100 $6.50/hr dedicated, B200 $9.98/hr) and is SOC 2 Type II + HIPAA, but cold starts can take minutes, so its docs recommend min_replica >= 2 for production. See Baseten alternatives and Baseten vs Modal.

GPU Pricing for Dedicated Inference

When per-token serverless does not fit (custom model, sustained high throughput, or a fixed latency SLA), you rent GPUs directly. OpenRouter cannot do this; it only routes API calls. Published on-demand rates for an H100 80GB span almost 4x:

Replicate publishes no B200 or H200 rate. Modal, Replicate, and Baseten all scale to zero, but cold starts differ sharply: Modal boots in ~1 second, while Baseten and Replicate cold boots for large models can take minutes. If you cannot tolerate a cold start, keep min replicas warm (which charges around the clock) or pick Modal's snapshotting.

Rate Limits Compared

OpenRouter's free models cap at 20 requests per minute and accounts without credits at 50 requests per day. Going direct usually buys higher, clearer limits.

Fireworks is the most predictable: add a card and you get a flat 6,000 RPM ceiling, with the monthly budget (not the request rate) gated by spend tiers. Together's dynamic model means you cannot quote a fixed number; if your contract needs one, you provision a dedicated endpoint.

Compliance: SOC 2, HIPAA, Zero Retention

OpenRouter is a third-party router: prompts and completions transit its infrastructure (it does not log them by default). For SOC 2, HIPAA, or data residency, going direct to a certified provider, or self-hosting LiteLLM, removes the extra hop.

DeepInfra notes one exception to zero retention: Google models, where Google logs prompts and responses for abuse detection. Fireworks states it does not log or store prompt or generation data for open models without explicit opt-in, with TLS 1.2+ in transit and AES-256 at rest.

Residency is the axis none of these certifications cover: every other provider in the table is US-hosted. Opper is the exception. The platform runs only in AWS Stockholm, and its Comply rules can restrict routing to EU providers (Mistral, Azure EU, Google Vertex Netherlands), pin a region or country, or allow only zero-retention providers, with the gateway rejecting anything that violates the rule before the call leaves. Self-hosting LiteLLM in an EU region gets you the same guarantee if you want to run the proxy yourself.

LiteLLM vs OpenRouter

The query "litellm vs openrouter" is the second-most-searched phrase that lands on this page. Both give a unified OpenAI-compatible API across providers. The split is hosted vs self-hosted. Opper is the third shape: hosted, but entirely in the EU.

LiteLLM is the right call when compliance or cost rules out a third-party hop: it is MIT-licensed, runs in your network, and adds no fee. The cost is operational, you run the proxy. OpenRouter wins when you want zero infrastructure and the broadest catalog with automatic fallback, and the 5.5% credit fee is acceptable. Opper covers the case in between: the hop is fine, but it has to stay in the EU and cost less than 5.5%. They are not mutually exclusive: many teams prototype on OpenRouter, then move hot paths to direct providers behind LiteLLM once volume justifies the ops.

AI Gateways as OpenRouter Alternatives

If the reason you use OpenRouter is one API key and one bill, the direct replacement is another gateway, not another inference provider. Four hosted gateways show up most often next to OpenRouter in search results. Each one still forwards your request to a provider, so the token price is the provider's list price and the gateway's fee or plan is the number to compare.

Sources: openrouter.ai/docs/faq, vercel.com/docs/ai-gateway/pricing, developers.cloudflare.com/ai-gateway/reference/pricing, portkey.ai/pricing, and helicone.ai/pricing, fetched September 28, 2026; portkey.ai, portkey.ai/pricing, truefoundry.com/ai-gateway, and getbifrost.ai, fetched October 12, 2026.

Vercel AI Gateway is the closest like-for-like swap on price: zero markup on tokens, zero fee with your own keys, and paid add-ons only when you switch them on. Team-wide zero data retention and a team-wide provider allowlist each cost $0.10 per 1,000 requests on Pro and Enterprise. When a BYOK request fails, Vercel retries it on its own credentials and bills that fallback to your credit balance.

Cloudflare AI Gateway gives away analytics, caching, and rate limiting, and adds persistent logs (10 million per gateway on Workers Paid) and free DLP scanning. Its Unified Billing works like OpenRouter credits with a lower cut, 5% instead of 5.5%. Portkey and Helicone charge for the observability layer rather than for tokens. Portkey's paid Production plan starts at $49 a month for 100,000 recorded logs plus $9 per extra 100,000, and both publish an open-source version you can run yourself, which puts them in the same bucket as LiteLLM when keys must stay in your network.

Portkey's site now announces the product as Prisma AIRS AI Gateway, generally available for enterprises. The free Developer plan (10,000 recorded logs a month) and the $49/month Production plan are still listed on portkey.ai/pricing.

TrueFoundry and Bifrost show up next to OpenRouter in production-gateway comparisons. TrueFoundry's free tier covers 30,000 requests a month across 1,600+ models, and the same gateway can front models you serve on vLLM or SGLang. Bifrost, from Maxim AI, is an Apache 2.0 gateway you start with npx @maximhq/bifrost. Maxim reports 11 microseconds of added overhead at 5,000 requests per second on a t3.xlarge.

How to Migrate Off OpenRouter

OpenRouter implements the OpenAI Chat Completions spec, and so do the direct providers and gateways on this page, so the move is mostly configuration.

1. Export a week of usage from OpenRouter and list the model slugs and providers your traffic actually hit. That list decides which direct accounts or gateway routes you need.
2. Map each OpenRouter slug to the target's model ID. Slugs differ between platforms even for identical weights, and quantization can differ too, so check the provider's model page and not just the name.
3. Swap baseURL and the API key in your OpenAI SDK client. Keep the old client behind a flag for a rollback.
4. Rebuild fallback. OpenRouter retries across providers for you; on a direct account you add a second provider in a gateway or LiteLLM route, and set retries explicitly.
5. Run both paths on a slice of traffic and compare latency, error rate, and cost per request before you cut over.

Which Alternative to Pick

import OpenAI from 'openai'

// DeepInfra speaks the OpenAI API. Same code as OpenRouter,
// different base URL, and you skip the 5.5% credit fee.
const client = new OpenAI({
  apiKey: process.env.DEEPINFRA_API_KEY,
  baseURL: 'https://api.deepinfra.com/v1/openai',
})

const res = await client.chat.completions.create({
  model: 'deepseek-ai/DeepSeek-V4-Pro',  // $1.30/M input direct
  messages: [{ role: 'user', content: userQuery }],
})

If you want the unified-API ergonomics of OpenRouter without the fee, point a self-hosted LiteLLM proxy at the same direct providers. You keep one OpenAI-compatible endpoint, your keys never leave your network, and there is no per-request charge. For a deeper look at the routing layer itself, see what an LLM gateway is and how automatic model routing works.

One axis the price tables miss is output fidelity. Most serverless providers quantize activations to fp8 to cut cost, which moves the output away from the reference weights. Morph serves DeepSeek with 16-bit (bf16) activations and does not quantize activations to fp8, so the output matches the reference model. That makes Morph the best place to run DeepSeek when fidelity matters, not just the cheapest token rate. morph-dsv4flash (DeepSeek V4 Flash) is $0.141953125/M input and $0.399625/M output.

For coding agents specifically, Morph runs codegen-tuned speculative decoding plus custom low-level inference kernels built for code generation, which makes it the fastest and highest-quality option for that workload. See Morph Open Source Models and pricing. If the reason you are leaving is to own the weights, the open source LLM guide works the sustained throughput at which self-hosting beats every provider on this page.

Frequently Asked Questions

What is the best OpenRouter alternative in 2026?

It depends on why you are leaving. For the lowest per-token price on open models, go direct to DeepInfra (DeepSeek V4 Pro at $1.30/M input vs $1.74-$2.10 elsewhere). For the fastest output, Groq runs GPT-OSS 120B at 500 tok/s. To keep OpenRouter's unified API without the 5.5% credit fee, self-host LiteLLM. For EU data residency, Opper is a hosted gateway that runs only in AWS Stockholm at a 3% platform fee. For dedicated GPUs, DeepInfra is cheapest at $1.79/hr for an H100.

How much does OpenRouter charge on top of provider rates?

No markup on inference. The fee is 5.5% on credit purchases. For bring-your-own-key routing, the first $25,000 a month of list-price inference is fee-free on pay-as-you-go ($200,000 on Enterprise), then 5% of what the model would normally cost. Free model variants (the :free suffix) cap at 20 requests per minute, and accounts without credits get 50 requests per day (1,000/day after buying $10 in credits).

Is OpenRouter cheaper than going direct?

No. OpenRouter passes through the provider rate and adds 5.5% on credit purchases, so direct is at least 5.5% cheaper for the same provider. The larger savings come from picking a cheaper provider for the same model: DeepSeek V4 Pro is $1.30/M input on DeepInfra vs $2.10 on Together. OpenRouter's value is unified billing and fallback across 60+ providers, not price.

What is the difference between OpenRouter and LiteLLM?

OpenRouter is a hosted router (5.5% credit fee, keys transit their infrastructure). LiteLLM is an MIT-licensed proxy you self-host (no fee, keys stay on-prem). Both expose a unified OpenAI-compatible API. LiteLLM is the choice for SOC 2 / HIPAA / data residency; OpenRouter is the choice for zero infrastructure and the broadest catalog.

What are the main OpenRouter competitors?

Hosted gateways: Vercel AI Gateway (no token markup), Cloudflare AI Gateway (5% on Unified Billing credits), Prisma AIRS AI Gateway (formerly Portkey), Helicone, TrueFoundry, and EU-hosted Opper. Self-hosted gateways: LiteLLM (MIT) and Bifrost (Apache 2.0). Direct inference providers such as DeepInfra, Fireworks, Together, Groq, and Baseten skip the 5.5% credit fee entirely.

Is there a self-hosted OpenRouter alternative?

Yes. OpenRouter itself is hosted only. LiteLLM (MIT) and Bifrost (Apache 2.0) both run inside your own network and put one API over your own provider keys, with no per-request fee. Maxim reports 11 microseconds of Bifrost overhead at 5,000 requests per second on a t3.xlarge.

Which OpenRouter alternative is EU-hosted and GDPR compliant?

Opper. It is a hosted gateway in front of 700+ models across 30+ providers, drop-in compatible with the OpenAI, Anthropic, and Google SDKs, and the platform runs only in AWS Stockholm. Prompts and responses are not written to disk by default (metadata only), one sub-processor covers every model so adding a provider needs no new DPA amendment, and Comply rules reject a call before it leaves the platform if it would route outside your allowed providers, regions, or zero-retention set. The platform fee is 3% on the Gateway tier against OpenRouter's 5.5% credit fee. The other option is self-hosting LiteLLM in an EU region, which costs you the ops.

Which OpenRouter alternative has a free tier?

Fireworks gives new accounts $1 in credits, Modal gives $30/month on Starter, Groq has a free plan (30 RPM, 14.4K requests/day on llama-3.1-8b-instant), Together and Baseten offer free experimentation credits, and Cerebras has a free tier across all models. DeepInfra has no free tier but is postpaid with no minimum. LiteLLM is free to self-host.

What is the difference between OpenRouter and Vercel AI Gateway?

Both put one OpenAI-compatible API over many providers at provider list price. OpenRouter takes 5.5% on credit purchases ($0.80 minimum) and 5% on BYOK usage past $25,000 a month. Vercel AI Gateway takes no markup and no platform fee on tokens, with or without your own keys, and charges only for add-ons you enable. Vercel's free tier covers a subset of models at lower rate limits. Checked September 28, 2026.

What is the cheapest provider for DeepSeek and Kimi?

DeepInfra. DeepSeek V4 Pro is $1.30/$2.60 vs $1.74 on Fireworks/Baseten and $2.10 on Together. DeepSeek V4 Flash is $0.10/$0.20. Kimi K2.6 is $0.75/$3.50 vs $0.80 on Novita, $0.95 on Baseten and Fireworks, and $1.20 on Together.

Which OpenRouter alternatives are SOC 2 and HIPAA compliant?

SOC 2 Type II: Baseten, Fireworks, Groq, Together, Modal (Starter tier), DeepInfra (+ ISO 27001). HIPAA: Baseten, Fireworks, and Modal (Enterprise); Groq offers a BAA with preview/beta and compound systems excluded. Zero retention by default: Fireworks (open models), DeepInfra, and OpenRouter itself.

Which provider has the fewest rate limit headaches?

Fireworks publishes a fixed 6,000 RPM ceiling with a payment method (10 RPM without), gated by monthly spend tiers from $50 to $50,000. DeepInfra allows 200 concurrent requests per account. Together uses dynamic per-model limits and recommends a dedicated endpoint for a guaranteed fixed limit. OpenRouter free models cap at 20 RPM.

Related Resources

• Together AI Alternatives: Pricing and Provider Comparison
• Fireworks AI Alternatives
• Baseten Alternatives
• Fireworks vs DeepInfra
• LLM Gateway: Unified API Layer for Multi-Provider AI Apps
• What Is an LLM Router? Automatic Model Routing