API that auto-routes to the cheapest AI provider (OpenAI/Anthropic/Gemini) | Hacker News Reader

Gemini)

tokensaver.org

24 pointsh2o_wine38 comments

input tokens: $0.5 per 1,000

output tokens: $1.5 per 1,000

that's either one hell of a typo or my god I'll be broke in an hour if I accidentally use this service

Yeah, agreed. This sounds like a proper scam to me...

It totally isn't. In a reply above I explained how I accidentally routed to Anthropic first & ended up losing $13.62 before I realized my mistake.

The moral of the story,if something gets screwed up, I end up paying

The website looks like AI, so do we call it a typo or a hallucination?

Not only is it AI its outdated AI.

https://tokensaver.org/api/pricing

Is offering GPT 3.5 Turbo and Gemini 1.5 Pro.

So right now it seems like it actually just provides Claude 3.5 sonnet, all you need to do is curl it.

I wonder how many dollars OP loaded on their API key. So far according to stats you've spent $13.65 for a few hundred thousand tokens people sent up.

{ "success": true, "totals": { "requests": 491, "revenue": 0, "cost": 13.650696, "profit": -13.650696, "inputTokens": 411557, "outputTokens": 258956, "customers": 32 }, "byProvider": [ { "provider": "anthropic", "requests": 491, "cost": 13.650696, "revenue": 0, "profit": -13.650696 } ] }

Actually I updated it to Gemini 2.0, GPT-4o, GPT-4o-mini, and Claude 3.5 Sonnet, Haiku. I did it when I realized I accidentally routed everything to Anthropic first & got hit with a $13.62 bill on rollout.

Moral of the story is, the only one who loses $ is me

That was exactly my impression. I thought the price was for 1 million tokens.

Actually, I accidentally routed to Anthropic first and ended up losing $13.62 but it does work. I've used it for two of my other projects.

I created it for people using AI in development. The 1st 50 tokens are free.

Mom, I want openrouter.

We have openrouter at home!

In all seriousness, the value proposition is weird to me. The most expensive queries are the ones with huge contexts, and therefore the ones I'd less likely to use cheap models.

The authentication for the API seems poorly designed. The auth token is your email address rather than a real auth token. If I know someone uses this service I can send a massive number of requests to cause a large credit card charge with just their email address. I thought this was just a mistake in the obviously LLM-written home page, but the API really does work this way after testing.

On top of that logging in does not require a password, just an email address.

This has a nice made up "case study": https://tokensaver.org/blog/how-i-saved-500-dollars-on-ai-co...

> Six months ago, I was running a customer support chatbot for a SaaS product. Nothing fancy - ...

I'm sure this toooootally happened

> curl -X POST https://tokensaver.org/api/chat \ -H "Content-Type: application/json" \ -d '{ "email": "your@email.com", "messages": [ {"role": "user", "content": "Hello!"} ] }'

Am I getting this right that there's no auth? Just provide an email and get free requests?

Edit:

This seems to always use sonnet 3.5 no matter the request. I asked it a USAMO problem and it still used sonnet (and hallucinated the wrong answer of course).

Vibe coded most likely. The creator might figure out the problems with that approach the hard way.

I screwed up the routing & it sent everything to Anthropic when I updated from Beta & ended up losing money. The case study was done 6 months ago.

It's a simple tool. Use it or don't. The only person who would lose money in error is me.

I am not sure who is the intended customer for this service.

The prompt and the model go hand in hand. If you randomly select the model the likelihood of getting something consistent is basically zero.

Also model pricing don't very that much. I have never heard of spot-instance equivalent for inference although that will be cool. The demand for GPU is so high right now that I think most datacenters are at 100% utilisation.

Btw landing page does not bring much confidence this is serious. Might want to change it to communicate better and also to be attractive to "developers" I guess.

> Also model pricing don't very that much.

I'm curious when AI pricing will couple with energy markets. Then the location of the datacentre will matter considerably

Depends on what you're doing. Something like "read this text and extract all the phone numbers" or "write a 3-point summary of this email" will perform about the same on all good models.

The equivalent of "Spot Instance" is basically the OpenAI Batch API

It's a tool not an SEO branded, shiny website. It's a utility. & model pricing varies considerably for now. This tool will be useless in another year or two.

Out of frustration, I built an AI API proxy that automatically routes each request to the cheapest available provider in real-time.

The problem: AI API pricing is a mess. OpenAI, Anthropic, and Google all have different pricing models, rate limits, and availability. Switching providers means rewriting code. Most devs just pick one and overpay.

The solution: One endpoint. Drop-in replacement for OpenAI's API. Behind the scenes, it checks current pricing and routes to whichever provider (GPT-4o, Claude, Gemini) costs least for that specific request. If one fails, it falls back to the next cheapest.

How it works: - Estimates token count before routing - Queries real-time provider costs from database - Routes to cheapest available option - Automatic fallback on provider errors - Unified response format regardless of provider

Typical savings: 60-90% on most requests, since Gemini Flash is often free/cheapest, but you still get Claude or GPT-4 when needed.

30 free requests, no card required: https://tokensaver.org

Technical deep-dive on provider pricing: https://tokensaver.org/blog/openai-vs-anthropic-vs-gemini-pr...

I wrote up how to reduce AI costs without switching providers entirely: https://tokensaver.org/blog/reduce-ai-api-costs-without-swit...

Happy to answer questions about the routing logic, pricing model, or architecture.

> Typical savings: 60-90% on most requests, since Gemini Flash is often free/cheapest, but you still get Claude or GPT-4 when needed.

This claim seems overstated. Accurately routing arbitrary prompts to the cheapest viable model is a hard problem. If it were reliably solvable, it would fundamentally disrupt the pricing models of OpenAI and Anthropic. In practice, you'd either sacrifice quality on edge cases or end up re-running failed requests on pricier models anyway, eating into those "savings".

I genuinely wonder the use cases are where the required accuracy is so low (or I guess the prompts are so strong) that you don't need to vigorously use evals to prevent regressions with the model that works best--let alone actually just change models on the fly based on what's cheaper.

Yes and in addition for some reason that use case is also not a fit for some cheap OS model like qwen or kimi, but must be run on the cheapest model of the big three.

Hi, curious, did you know about OpenRouter before building this?

> OpenRouter provides a unified API that gives you access to hundreds of AI models through a single endpoint, while automatically handling fallbacks and selecting the most cost-effective options. Get started with just a few lines of code using your preferred SDK or framework.

It isn't OpenAI API compatible as far as I know, but they have been providing this service for a while...

OpenRouter can also prioritize providers by price: https://openrouter.ai/docs/guides/routing/provider-selection...

You should probably take the service down before the HN crowd maxes out your credit card with the already discovered security and auth issues. Then find a technical co founder of you still want to pursue this idea and build it from scratch.

My credit card isn't in there. The app was written 6 months ago where it stayed in Beta. I rolled it out as a way to reduce the cost of development. Use it or don't. It will be obsolete in another year or two when AI calls level in price.

Inferred tokens are a commodity. It's crude oil, not an oil painting.

https://news.ycombinator.com/item?id=45837691

You don't have to take my word for it.

https://www.aiville.com/c/anthropic/api-that-auto-routes-to-...

For input,

- GPT-5.1 is $1.25 / 1M tokens

- You are $0.50 / 1,000 tokens

Output:

- GPT-5.1 is $10.00 / 1M tokens

- You are $1.50 / 1,000 tokens

Am I reading that wrong? Is that a typo?

Really interesting approach.

As these routing engines evolve, I wonder how you see them handling drift or divergence when different models produce structurally incompatible outputs.

Any thoughts on lightweight harmonization layers?

A free router is https://huggingface.co/models

This landing page is vibe coded and littered with mistakes/typos (per 1,000 tokens), outdated models (Gemini 1.5?), the security link at the bottom of the page is an href=#, and I can see the "dashboard" without logging in or signing up.

> Message Privacy: Your API requests are processed and immediately forwarded. We never store or log conversation content.

> Minimal Data: We only store your email and usage records. Nothing else. Your data stays yours.

Source: trust me bro.

Reply on news.ycombinator.com