We publish a picker at llmapi.ai/openrouter-alternatives that narrows 20 routers, gateways and inference clouds down to the three that fit your workload. This post is the working behind it: what we looked at, how we scored, and where the comparison stops being reliable.
Why people leave OpenRouter in the first place
OpenRouter is a good product and a fair baseline. In practice teams move for four reasons, and they are rarely about the model list.
- Modality. The catalogue is deep in text and thin in speech, video and OCR. Once a product grows a voice feature, a second vendor appears in the stack.
- Billing. Prepaid credits and a per-request fee are fine for a side project and awkward for a finance team that wants an invoice, a PO and a cost centre.
- Governance. Data residency, audit logs, per-team budgets and an SLA with a name on it tend to arrive with the first enterprise customer.
- Control. Some security teams will not accept a third party in the request path at all, which points at a self-hosted gateway rather than a hosted one.
The ten criteria
Every entry was read against the same ten questions, using the vendor's own documentation, pricing page and status page, plus hands-on use where we had an account.
- Model breadth — how many models, and how quickly new releases land.
- Modality coverage — text, vision, speech, video, OCR, embeddings.
- Routing and failover — what happens when an upstream provider degrades.
- Pricing shape — markup, credits, invoices, commitments, hidden minimums.
- Spend controls — budgets, per-key limits, per-team reporting.
- Data residency — regional hosting, EU options, retention policy.
- Self-hosting — whether you can run it yourself, and under which licence.
- White label — whether you can resell it under your own brand.
- Enterprise readiness — SLAs, SOC 2, DPAs, procurement paperwork.
- Migration effort — how close it is to a drop-in base URL swap.
How the picker scores
We did not publish a single overall ranking, because there isn't one. A gateway that is perfect for a bank is the wrong answer for a two-person team shipping a demo. Instead each alternative carries tags on ten axes — workload size, modality, model ecosystem, hosting, privacy regime, keys and billing, routing behaviour, stated priority, migration driver and migration effort — and the picker adds a weight each time one of your answers matches a tag. We started at six questions and added four more — ecosystem, privacy, billing and routing behaviour — because those were the distinctions readers kept writing in about.
What the picker weighs
Hosting carries the heaviest weight, because it is the only answer that is usually non-negotiable: if your security team said self-hosted, a managed router is not a compromise, it is a rejection. Modality and your stated priority come next, workload size last, since most platforms stretch across sizes. The result is three names with the matched reasons spelled out, plus the two worst fits for your answers, so you can see what the scoring rejected and why.
What the gateway layer actually costs
Token prices are set by whoever runs the GPU. What a router adds on top is a separate, much smaller number — and it is the number people argue about. Here is every fee structure in the comparison, as published on the vendors' own pricing pages, checked 21 September 2026.
Where a flat fee beats a percentage
| Platform | What it charges on top of tokens | Free entry point |
|---|---|---|
| OpenRouter | 5.5% on card credit purchases ($0.80 minimum), 5.0% flat on crypto. No per-token markup. BYOK free below $25,000/month of list-price inference, then 5%. Business plan 8% | 25+ free models, 50 requests/day |
| Requesty | ~5% markup on token spend | Trial credits |
| Cloudflare AI Gateway | Gateway features free; 5% fee on credits bought through Unified Billing, no markup on inference | Free on all plans, 100,000 logs on the free tier |
| Portkey | $49/month flat for the production control plane, 100,000 logged requests included, $9 per extra 100,000. No token markup | Dev plan, 10,000 requests/month |
| LiteLLM | $0 for the open-source gateway. Enterprise is priced on annual request capacity, never per token | Self-host, free forever |
| Eden AI | ~5.5% at checkout, 0% markup on tokens | Starter credits |
| Hugging Face Inference | 0% routing markup, pass-through provider pricing | Monthly included credits |
| Together, Fireworks, Groq, DeepInfra, Novita, Replicate | Nothing — they are the inference provider, the price is the price | Groq: free tier on selected models |
| AWS Bedrock, Azure AI Foundry, Google Vertex AI | Nothing separate — per-token or provisioned-capacity rates land on the cloud bill you already have | Cloud free credits |
| LLM API | Pay per token with volume discounts, invoiced; no separate credit-purchase fee | Trial credits |
Put in money, the percentage models and the flat models cross over quickly. Fees only, one year, tokens billed separately:
| Monthly inference spend | OpenRouter, 5.5% | Requesty, 5% | Portkey, $49 base | LiteLLM, self-hosted |
|---|---|---|---|---|
| $100 | $66/yr | $60/yr | $588/yr | $0 + your ops time |
| $1,000 | $660/yr | $600/yr | $588/yr | $0 + your ops time |
| $10,000 | $6,600/yr | $6,000/yr | $588/yr | $0 + your ops time |
| $40,000 | $26,400/yr | $24,000/yr | $588/yr | $0 + your ops time |
The crossover against Portkey's flat fee sits near $900 a month of inference. Below it, a percentage is cheaper; above it, a flat control-plane fee is. That single line explains most of the migration traffic between these two shapes of product, and it is also why the small-top-up trap matters: OpenRouter's $0.80 minimum turns a $5 credit purchase into an effective 16% fee, so top up in blocks rather than in dribs.
The same job, nine bills
Percentages are abstract. Take one concrete workload — 40M input and 10M output tokens a month on DeepSeek V4 Pro, roughly a busy RAG product — and price it at each platform's published list rate on 21 September 2026.
One workload, nine bills
| Platform | Rate, $/1M in · out | Token cost | Layer fee | Monthly total |
|---|---|---|---|---|
| DeepInfra | $1.30 · $2.60 | $78.00 | — | $78.00 |
| Requesty (routed to cheapest endpoint) | $1.30 · $2.60 | $78.00 | $3.90 | $81.90 |
| Cloudflare AI Gateway (unified billing) | pass-through | $78.00 | $3.90 | $81.90 |
| OpenRouter (routed to cheapest endpoint) | $1.30 · $2.60 | $78.00 | $4.29 | $82.29 |
| Novita AI | $1.60 · $3.20 | $96.00 | — | $96.00 |
| Together AI | $1.74 · $3.48 | $104.40 | — | $104.40 |
| LLM API | $1.74 · $3.48 | $104.40 | — | $104.40 before volume discount |
| Portkey + DeepInfra | $1.30 · $2.60 | $78.00 | $49.00 | $127.00 |
| LiteLLM + DeepInfra, self-hosted | $1.30 · $2.60 | $78.00 | hosting + ops | $78.00 plus your own time |
Two things worth saying out loud. First, at this size the entire argument is worth about $49 a month — less than the engineering hours you would spend switching, which is why we score migration effort as a criterion rather than a footnote. Second, we are not the cheapest row in our own table, and pretending otherwise would make every other number here suspect. On a single open model routed to the cheapest endpoint, a raw inference provider wins on price. The case for a router is what happens when that endpoint is down, when the second modality arrives, or when finance asks for one invoice.
Token prices for the same model, side by side
List prices per 1M tokens, input · output, as published on 21 September 2026. A dash means the platform does not serve that model.
Output price spread on the same model
| Model | Together AI | DeepInfra | Novita AI | Groq | LLM API |
|---|---|---|---|---|---|
| DeepSeek V4 Pro | $1.74 · $3.48 | $1.30 · $2.60 | $1.60 · $3.20 | — | $1.74 · $3.48 |
| DeepSeek V4 Flash | $0.14 · $0.28 | $0.09 · $0.18 | $0.14 · $0.28 | — | $0.19 · $0.51 |
| GLM-5.2 | $1.40 · $4.40 | $0.563 · $1.80 | $0.65 · $2.04 | — | $1.40 · $4.40 |
| Kimi K2.6 | $1.20 · $4.50 | $0.75 · $3.50 | $0.80 · $3.40 | — | $0.95 · $4.00 |
| gpt-oss-120b | $0.15 · $0.60 | $0.037 · $0.17 | $0.05 · $0.25 | free tier, then paid | $0.15 · $0.60 |
| Llama 3.3 70B | $1.04 · $1.04 | — | — | $0.59 · $0.79 | $0.13 · $0.40 |
The spread on a single model reaches 2.5x — GLM-5.2 output runs $1.80 on DeepInfra and $4.40 on Together — which is larger than any platform fee in the comparison. If price is genuinely your first criterion, the endpoint matters far more than the router, and the right question is not "which gateway is cheapest" but "which gateway lets me pin the cheap endpoint and fall back when it breaks."
Free tiers, measured
The hobby end of the market is where published limits move fastest, so these are quoted with dates attached.
- Groq — 30 requests/minute, 1,000 requests/day, 8,000 tokens/minute and 200,000 tokens/day on gpt-oss-120b, gpt-oss-20b and the Qwen 3 27B models, no card required. Llama 3.1 8B and Llama 3.3 70B left the free and Developer tiers on 16 August 2026 and are enterprise-only now — a good example of why a free tier is a bad thing to architect around.
- OpenRouter — 25+ free models at 50 requests/day, rising to 1,000/day after a lifetime $10 credit purchase.
- Portkey — Dev plan, 10,000 requests/month.
- Cloudflare AI Gateway — core gateway free on all plans; 100,000 logs total on Workers Free, 10M per gateway on Workers Paid.
- LiteLLM — the whole gateway, free forever, on hardware you pay for.
How wide the catalogues actually are
Model counts are the most quoted and least useful number in this market, because they count text models that nobody calls. The figure that changes an architecture is modality coverage. Our own catalogue, read from the live feed on 21 September 2026, carries 398 models: 567 text endpoints, 217 image, 150 video, 76 audio and 11 embedding entries across input and output. OpenRouter publishes 500+ models across 80+ providers, overwhelmingly text with some vision. Eden AI returned 221 expert-model entries across 36 sub-features from 42 specialist providers in a September 2026 count, weighted towards OCR, document parsing and speech rather than chat.
Modality coverage, three catalogues compared
Read those three lines together and the split is clear: OpenRouter is the widest text catalogue, Eden AI is the deepest document and speech catalogue, and platforms like ours sit in between because we carry video and speech alongside text on one key. None of these numbers is a quality claim. They tell you what you will not have to integrate separately.
What we could not measure
Three things are missing on purpose.
- Latency and throughput. These vary by model, region, hour and upstream provider. Any single number would be stale before you read it, and vendor-published benchmarks are marketing. Where speed is the deciding factor, run your own prompt on your own region for a day.
- Real-world uptime. We monitor our own per-model availability and publish it, but we have no comparable instrumentation on other platforms, so we report what routing features exist rather than how often they save you.
- Support quality. Reported from public plan documents, not tested. Community Discord and a named account manager are different products.
Where a vendor does not publish a fact, the comparison says so rather than estimating. Model counts move weekly; treat them as orders of magnitude.
The shortlist at a glance
| Platform | Modalities | Hosting | Pricing shape |
|---|---|---|---|
| OpenRouter | Text, some vision | Managed cloud | Provider price + platform fee, prepaid credits |
| LLM API | Text, vision, speech, video, OCR | Managed cloud, EU region | Pay per token with volume discounts, invoiced |
| Together AI | Text, image, embeddings | Managed cloud | Per token, plus hourly dedicated GPUs |
| Fireworks AI | Text, vision, audio, image | Managed cloud | Per token, plus on-demand GPU hours |
| Groq | Text, speech-to-text | Managed cloud | Per token, free developer tier |
| DeepInfra | Text, image, speech, embeddings | Managed cloud | Per token, pay as you go |
| Novita AI | Text, image, video, speech | Managed cloud | Per token or per generation |
| Replicate | Image, video, audio, text | Managed cloud | Per second of compute |
| AWS Bedrock | Text, image, embeddings | Managed cloud, many regions incl. EU | Per token, on your AWS bill; provisioned throughput available |
| Azure AI Foundry | Text, image, speech, vision | Managed cloud, EU data boundary | Per token or provisioned units, on the Azure bill |
| Google Vertex AI | Text, image, video, speech | Managed cloud, EU regions | Per token or per media unit, on the GCP bill |
| Portkey | Whatever your providers support | Managed cloud or self-hosted | Per request tiers, free developer plan, enterprise self-host |
| LiteLLM | Whatever your providers support | Self-hosted (managed tier available) | Free and open source, paid enterprise tier |
| Kong AI Gateway | Whatever your providers support | Self-hosted or hybrid | Open source core, enterprise licence |
| Cloudflare AI Gateway | Text, image, speech via Workers AI | Managed edge, EU options | Free tier, usage-based extras |
| Anyscale | Whatever you deploy | Your cloud account / VPC | Platform fee on top of your own compute |
| Hugging Face Inference | Text, image, audio, embeddings | Managed cloud | Free credits, then pass-through provider pricing |
| Requesty | Text, vision | Managed cloud | Usage fee on top of provider prices |
| Eden AI | Text, OCR, speech, image, translation | Managed cloud, EU company | Per call, with a monthly plan |
| Unify | Text | Managed cloud | Provider price plus routing fee |
Four situations, four answers
Prototyping with maximum choice. Stay where you are, or add Hugging Face Inference. At small volume the per-request fee is noise and the catalogue breadth is the whole point.
Production across text, voice and video. You want one key that covers all of it, with uptime monitoring and a single invoice, which is the case LLM API was built for. Novita AI is the cheaper alternative if lighter guarantees are acceptable.
Nothing third-party in the data path. Self-hosted LiteLLM, or Kong AI Gateway if your traffic already runs through Kong. You keep provider prices with no markup and you own the uptime.
Procurement has to sign. AWS Bedrock, Azure AI Foundry or Google Vertex AI put the spend on a contract that already exists, with residency and compliance attached. You trade catalogue breadth for paperwork that is already done.
Sources, and when we checked them
Every number above comes from a vendor's own published page, read on 21 September 2026. Prices in this market move monthly; if you are budgeting, re-check the two or three rows that decide your answer rather than trusting a blog post, including this one.
- OpenRouter pricing and BYOK terms — openrouter.ai/pricing and the OpenRouter FAQ.
- Together AI serverless rates — together.ai/pricing.
- DeepInfra rates — deepinfra.com/pricing.
- Novita AI rates — novita.ai/pricing.
- Groq rates and free-tier limits — Groq's pricing and rate-limit documentation.
- Portkey plans and log overage — Portkey feature comparison.
- LiteLLM licensing — litellm.ai/pricing.
- Cloudflare AI Gateway fees and log limits — Cloudflare AI Gateway pricing and Unified Billing docs.
- LLM API rates and modality counts — our own catalogue feed, refreshed nightly and shown on each model page.
Where we are biased, stated plainly
We build one of the twenty. LLM API appears in the comparison on the same criteria as everyone else, it wins the recommendations where the answers point at mixed modalities, unified billing, failover or white-label reselling, and it loses them where you asked for self-hosting, a free hobby tier or a single-vendor cloud contract. If the honest answer to your situation is LiteLLM on your own cluster, the picker will say LiteLLM.
Take the ten-question picker and see which three come back for your workload.
