One endpoint. Every model, routed for you.
Route by price, latency or quality across 20 models from 8 providers. Fall back automatically when one degrades, and see every dollar in one dashboard.
import OpenAI from "openai";
// Same SDK. One base URL. Every provider.
const client = new OpenAI({
apiKey: process.env.FINAL_ROUTER_KEY,
baseURL: "https://finalrouter.com/api/v1",
});
// "auto" lets Final Router pick the model.
const response = await client.chat.completions.create({
model: "auto",
messages: [{ role: "user", content: prompt }],
});Seven providers behind the endpoint - and the tools you already use in front of it
OpenAI
Anthropic
Google
Meta
Mistral
DeepSeek
Kimi
Grok
CursorClaude Code
VS Code
ChatGPTGemini CLI
Cline
Roo Code- 33x
- Price difference across the catalogue
- 20
- Models behind one endpoint
- 0%
- Markup on your tokens
From $0.15 to $5 per million input tokens. Send everything to one model and you pay the top rate for the easy work too.
Across 8 providers, all reachable with the OpenAI client you already have. One base URL. No new SDK to learn.
You pay what each provider charges. We take our cut when you buy credit, so we gain nothing by sending you to a pricier model.
You could build all of this yourself
None of it is difficult. There is just a lot of it, and it never really finishes. Every team that adds a second provider ends up writing the same routing, the same fallback and the same cost tracking, then maintaining all three while the models keep changing underneath them.
Wiring up providers yourself
A separate SDK for every provider, each with its own errors and its own streaming format
Changing model means changing code, deploying, and having a way back
When a provider goes down, you are down until someone notices and switches it over
Five invoices at the end of the month and no idea which feature spent what
A leaked key has to be found, rotated and redeployed everywhere it was used
No straight answer to which model handled a request, or what it cost
With Final Router
One endpoint that speaks the OpenAI format. Change the base URL, leave the rest alone
Swap between 20 models by changing a string, or send "auto" and let us pick
A slow provider gets skipped while the request is still running, so the answer still arrives
One balance and one ledger, with the cost of each request shown as it happens
Provider keys are encrypted once, never sit in your app, and can be revoked in a click
Every response tells you which model answered, why it was picked, and what it cost
The cheapest model in the catalogue costs $0.15 per million input tokens. The dearest costs $5. That is 33 times the price for work that often does not need it. Sending the easy requests to the cheap model is the whole idea.
Everything a multi-model app needs, minus the plumbing
You should be shipping product, not writing your fifth provider adapter or reconciling five invoices at the end of the month.
Smart routing
Every request is scored against the models you allow, then sent to the one that wins. Ask for cheapest, fastest or best, or let it balance the three.
Four strategies: cost, latency, quality, balanced
Automatic fallback
A provider that rate-limits or times out gets skipped while the request is still running. It carries on down your chain instead of failing.
429 and 5xx retry elsewhere, 401 surfaces
One key, not five
Store each provider credential once, encrypted. Your app only ever carries a single Final Router key, and you can revoke it in a click.
AES-256-GCM at rest, hashed gateway keys
Spend caps
Set a monthly limit. A request that would take you past it is refused before it reaches a provider, so it costs you nothing.
Checked before the call, not after
Usage you can read
Tokens, cost and latency broken down by model, provider and key. Written as each request lands, not totted up at the end of the month.
Written per request, not batched
Load balancing
Spread traffic across several keys so no single account hits its own rate limit while the others sit idle.
Per-key limits with RateLimit headers
Live in three steps, not three sprints
Connect your providers
Paste the API keys you already have. They are encrypted at rest and never returned to a browser.
Point your app at one URL
Change the base URL in the SDK you already use. Nothing else in your code has to move.
Watch it route
Every request picks a model, falls back if it has to, and lands in your dashboard with its cost and latency attached.
Models
Every model, priced side by side
Prices are per million tokens. Quality is a composite benchmark score; latency is the median across the last 24 hours of real traffic.
| Model | Provider | Context | Input | Output | Quality | Latency |
|---|---|---|---|---|---|---|
Claude Opus 5US - processed in US. We call api.anthropic.com, served from the United States. Their models are also sold through EU regions of Bedrock and Vertex, which we do not integrate. anthropic/claude-opus-5 | Anthropic | 1M | $5.00 | $25.00 | 98 | 2.1s |
Claude Sonnet 5US - processed in US. We call api.anthropic.com, served from the United States. Their models are also sold through EU regions of Bedrock and Vertex, which we do not integrate. anthropic/claude-sonnet-5 | Anthropic | 1M | $2.00 | $10.00 | 93 | 980ms |
Claude Haiku 4.5US - processed in US. We call api.anthropic.com, served from the United States. Their models are also sold through EU regions of Bedrock and Vertex, which we do not integrate. anthropic/claude-haiku-4-5 | Anthropic | 200K | $1.00 | $5.00 | 84 | 420ms |
GPT-5.1US - processed in US. We call api.openai.com, which is served from the United States. They sell EU data residency on other endpoints; we are not on one, so this says what happens to your request rather than what the company offers. openai/gpt-5.1 | OpenAI | 400K | $1.25 | $10.00 | 95 | 1.4s |
GPT-5 miniUS - processed in US. We call api.openai.com, which is served from the United States. They sell EU data residency on other endpoints; we are not on one, so this says what happens to your request rather than what the company offers. openai/gpt-5-mini | OpenAI | 400K | $0.25 | $2.00 | 82 | 380ms |
Gemini 2.5 ProUS - processed in US. The Gemini API endpoint we call is global, which in practice means processing outside the EU. Vertex AI offers European regions; that is a different integration and not this one. google/gemini-2.5-pro Prompts over 200k tokens cost $2.50 in and $15.00 out, per 1M. | 1.048576M | $1.25 | $10.00 | 92 | 1.1s | |
Gemini 2.5 FlashUS - processed in US. The Gemini API endpoint we call is global, which in practice means processing outside the EU. Vertex AI offers European regions; that is a different integration and not this one. google/gemini-2.5-flash | 1.048576M | $0.30 | $2.50 | 78 | 310ms | |
Muse Spark 1.2US - processed in US. Served from the United States on the endpoint we call. meta/muse-spark-1.2 | Meta | 1.048576M | $1.25 | $4.25 | 88 | 1.8s |
Muse Spark 1.1US - processed in US. Served from the United States on the endpoint we call. meta/muse-spark-1.1 | Meta | 1.048576M | $1.25 | $4.25 | 83 | 1.7s |
Mistral LargeEU - processed in EU. Mistral is a Paris company directly subject to the GDPR, and La Plateforme processes in the EU by default with no configuration. Their own documentation notes temporary transfers for some features and subprocessors, so this is EU by default rather than EU without exception. mistral/mistral-large-latest | Mistral | 262K | $0.50 | $1.50 | 85 | 760ms |
Mistral SmallEU - processed in EU. Mistral is a Paris company directly subject to the GDPR, and La Plateforme processes in the EU by default with no configuration. Their own documentation notes temporary transfers for some features and subprocessors, so this is EU by default rather than EU without exception. mistral/mistral-small-latest | Mistral | 262K | $0.15 | $0.60 | 70 | 260ms |
Kimi K3SG - processed in SG. The international API (api.moonshot.ai) is run by their Singapore entity, which states data is stored on servers in Singapore. The parent company is Beijing-based; the .cn API is a separate service we do not call. moonshot/kimi-k3 | Moonshot AI | 1.048576M | $3.00 | $15.00 | 94 | 2.2s |
Kimi K2.6SG - processed in SG. The international API (api.moonshot.ai) is run by their Singapore entity, which states data is stored on servers in Singapore. The parent company is Beijing-based; the .cn API is a separate service we do not call. moonshot/kimi-k2.6 | Moonshot AI | 262K | $0.95 | $4.00 | 87 | 1.3s |
Kimi K2.7 CodeSG - processed in SG. The international API (api.moonshot.ai) is run by their Singapore entity, which states data is stored on servers in Singapore. The parent company is Beijing-based; the .cn API is a separate service we do not call. moonshot/kimi-k2.7-code | Moonshot AI | 262K | $0.95 | $4.00 | 89 | 1.1s |
Kimi K2.7 Code HighspeedSG - processed in SG. The international API (api.moonshot.ai) is run by their Singapore entity, which states data is stored on servers in Singapore. The parent company is Beijing-based; the .cn API is a separate service we do not call. moonshot/kimi-k2.7-code-highspeed | Moonshot AI | 262K | $1.90 | $8.00 | 89 | 700ms |
Grok 4.6US - processed in US. xAI is a United States company and api.x.ai is a single global endpoint with no regional variant to choose. Their published documentation does not name a storage region for API traffic, so this is recorded from where the company and its infrastructure sit rather than from a residency commitment they have made. xai/grok-4.6 Prompts over 200k tokens cost $4.00 in and $12.00 out, per 1M. | xAI | 500K | $2.00 | $6.00 | 96 | 4.3s |
Grok 4.3US - processed in US. xAI is a United States company and api.x.ai is a single global endpoint with no regional variant to choose. Their published documentation does not name a storage region for API traffic, so this is recorded from where the company and its infrastructure sit rather than from a residency commitment they have made. xai/grok-4.3 Prompts over 200k tokens cost $2.50 in and $5.00 out, per 1M. | xAI | 1M | $1.25 | $2.50 | 91 | 2.8s |
Grok Build 0.1US - processed in US. xAI is a United States company and api.x.ai is a single global endpoint with no regional variant to choose. Their published documentation does not name a storage region for API traffic, so this is recorded from where the company and its infrastructure sit rather than from a residency commitment they have made. xai/grok-build-0.1 Prompts over 200k tokens cost $2.00 in and $4.00 out, per 1M. | xAI | 256K | $1.00 | $2.00 | 88 | 3.6s |
DeepSeek V4 ProCN - processed in CN. Their privacy policy states data is collected, processed and stored in the People's Republic of China. The one provider in the catalogue where the answer is neither the EU nor the US. deepseek/deepseek-v4-pro Between 01:00-04:00 and 06:00-10:00 UTC this model costs $1.32 in and $3.96 out, per 1M. | DeepSeek | 1M | $0.66 | $1.98 | 90 | 1.6s |
DeepSeek V4 FlashCN - processed in CN. Their privacy policy states data is collected, processed and stored in the People's Republic of China. The one provider in the catalogue where the answer is neither the EU nor the US. deepseek/deepseek-v4-flash Between 01:00-04:00 and 06:00-10:00 UTC this model costs $0.44 in and $1.32 out, per 1M. | DeepSeek | 1M | $0.22 | $0.66 | 79 | 700ms |
Pricing
No markup on tokens. Ever.
You pay exactly what the provider charges. We add 5.5% when you top up your credits, and nothing after that. No subscription, no seat fees, no minimum spend.
Free
Enough to try the API properly before you pay for anything.
- GPT-5 mini and DeepSeek V4 Flash, 25 requests a day
- 20 requests a minute
- Replies up to 8,000 tokens
- Fallback, streaming and tool calling
- 2 API keys, 1 guardrail
Pay as you go
Every model, at the price the provider charges.
- No markup on tokens, ever
- All 20 models across 8 providers, no daily cap
- Rate limit rises with your balance
- Routing policies: yours, plus 4 built in
- Unlimited keys and guardrails
- Bring your own provider keys at 5% of list price
Volume
Spending enough that a percentage is a real number? There is usually something we can do.
- Terms on the credit fee
- Higher bring-your-own-key allowance
- A person who knows your account
The full comparison, and the questions people ask before paying
Questions, answered straight
What is Final Router?
An AI gateway: one OpenAI-compatible endpoint over eight providers - OpenAI, Anthropic, Google, Meta, Mistral, DeepSeek, Moonshot and xAI. Point your existing SDK at our base URL and routing, automatic fallback, spend caps, guardrails and cost tracking apply to every request.
What does it cost?
Tokens are billed at each provider's published list price with no markup - every rate is on the pricing page with the date it was checked. A 5.5% fee (minimum $0.80) applies when you buy credits, and bring-your-own-key usage carries a 5% fee after a $25,000 monthly free allowance.
Do I have to change my code?
One line: the base URL. If your code speaks OpenAI's API, it already speaks ours - tools, streaming, image input and json_schema structured output included. Unknown fields are refused with named 400s instead of being silently dropped.
What happens when a provider goes down?
Your request falls back to the next healthy model automatically, with tools and conversation history intact, and the response names what it fell back from. Scheduled probes feed the same health signal to routing and to the public status page.
Do you store or train on our prompts?
We never train on your data, and prompt text is not stored unless you opt in - then it is kept 30 days, visible only to you, and turning storage off deletes everything. Each provider's own policy is recorded per model in the catalogue.
Can I try it without a card?
Yes - the free tier gives 25 requests a day on two starter models, up to 8,000 output tokens a reply, no card required. Enough to wire up an integration and watch real responses explain their routing.
Stop maintaining five integrations
Swap your base URL and be routing across every major provider in under five minutes.