Posted by rvz 18 hours ago
The good thing about the developer experience is that it is dead simple to create keys that are time-bound, model restricted, and have budgets against them in an easy, programmatic way (also in bulk). Costs are guaranteed to be on par or lower than any model provider, so it's an easy sell to the finance department.
To look at it from the other side, Gemini is the absolute worse when it comes to key management. The labyrinth of actions that one has to navigate to generate budgeted keys for a team is not worth it and I still haven't figured out how to manage the keys at a granularity I need.
OpenAI and Anthropic both have what I would consider to be middling DevEx. Most things work, but only on their platform, and there are some weirdnesses, like the fact that anthropic disabled its admin API key creation, so you can't programmatically create keys against a budget.
I saw another comment thread on this recently. What I gathered was they hid it in the Cloud Platform, so for those who jumped the cloud hoops already, maybe it's alright. (However when I read what Openrouter offers, as you just laid out, I feel Google can't hold a candle to Openrouter DevEx.)
However, if there's anyone who is aware, was there a brief point where getting a Gemini API key was simple (enough)? I'm asking because, in true Google fashion, the AI product went through a lot of changes.
You would think that these things would be standard fare, but they really are not. Other providers need metadata from a separate catalog, don't report pricing, don't really report available credits / usage either.
So yes, OpenRouter does have a good DevEx.
And if it's not worth it, then are you spending enough where it even effects Openrouter's bottom line?
With enough devs playing with hobby projects I'm not doubting it's profitable, only that $7 billion seems way too high.
It also has way better uptime than the underlying platforms, even for proprietary models like Claude. When Claude APIs are having issues, OpenRouter Claude still keeps working because they can route to AWS Bedrock instead of Anthropic etc. This effect is even bigger with open-weight models because they typically have 5-10 providers.
It seems strategically not in their interest.
And per the rankings [0] 18.2% of the tokens sold through went to OpenAI [1]. 3.8% for Anthropic though. Google's up there at 23.4% primarily because GCP/Vertex/whatever they're offering is a billing nightmare and the only way to set a hard cap and sure you won't wake up the next day with a $10k bill is via OR.
[0] https://openrouter.ai/rankings#market-share
[1] Although gpt-oss 20b and 120b are indexed as 'openai' because they're the creator even though they don't serve gpt-oss via openrouter. But neither are in the top 20 of models this month, while Luna and Sol are.
> **Google's** up there at 23.4% primarily because GCP/Vertex/whatever they're offering is a billing nightmare and the only way to set a hard cap and sure you won't wake up the next day with a $10k bill is via OR.
OpenAI absolutely lets you set a hard or soft cap. I've set them and they've worked, I think one time I saw a $-0.01 balance. Their docs were probably written by ChatGPT. Later in that article it says:
> A spend limit can monitor spend without enforcement, or you can enforce it as a hard limit so API requests fail after spend reaches the limit.
And they still allow the most certain limit, which Google does not: have $0 on your deposited balance and don't have auto-reload enabled, you get 4XXed instead of going into debt (or at least by at most a couple cents).
Although for fun, go ask Gemini how you can set a cap on GCP spend. It'll try to write you a pub-sub client.
And OpenRouter can steer traffic away from those labs over to their own model.
The major labs don’t get personalized data about the user to train on, since it’s aggregated.
Which is just some of the reasons why strategically it doesn’t seem in their best interest to allow a middle person.
> OpenAI/Anthropic/etc don’t establish the direct customer relationship.
But they also don't have to pay to support and maintain the direct customer relationship and all the billing, support, liability, etc. The credit card merchant fee (a not insubstantial % at scale) is paid by Openrouter (or Stripe now) and the bill Openrouter pays to OpenAI is transferred probably via wire or something with a lot less fees. Meanwhile if you're paying OpenAI via credit card directly, OpenAI is paying Stripe or whoever.
> And OpenRouter can steer traffic away from those labs over to their own model.
The inverse is also true. If you've deposited money on a Openrouter account, you've taken a very big step across their proprietary moat. Would the proprietary companies rather lose them as a customer forever, or be able to entice them back, which can be done with a single change of a JSON parameter?
> The major labs don’t get personalized data about the user to train on, since it’s aggregated.
(Edit: Looked it up, Openrouter does send a pseudonymous but stable hashed userid upstream, stable across different API keys in the same account [0]. That makes de-identification a lot easier.)
First, a lot of the corporate use is getting ZDRed, so there's nothing to track without going back on zero data retention.
De-identification was trivial for normal web traffic before LLMs. Think of all the crap people's harnesses and claws send without de-identifying. Many coding agents can't help themselves from scanning your home directory.
Sure, the type of person to comment on HN probably has good hygene, but the typical vibe coder doesn't even know what they're leaking.
[0] https://openrouter.ai/docs/cookbook/administration/user-trac...
It's an administrative burden to sign up very every model provider, and there are many independent inference providers now that serve only open source models.
OpenRouter provides a useful service by allowing easy prepaid model access with much higher rate limits, and they also aggregate different model providers to route queries by price, latency, etc.
I can imagine just the opposite outcome from the same scenario: as people settle into their favorite but commoditized models, competition for marginal inference cost will take over. A company like OpenRouter that promises the cheapest tokens by the minute becomes essential on the low-cost margin.
I think that OpenRouter and equivalents get pushed out of the market only if the froth calms down (as you posit) and winning models stay proprietary, perhaps with their own unique API surfaces.
Okay, I can see that, but if their value is just cost optimization, their ability to mark up the tokens becomes increasingly difficult as well. Or, people will build a router themselves to avoid paying the markup, possibly with reduced features, but someone will open source it. Heck, Claude or GPT can probably one-shot it these days. Either way, I think the whole OpenRouter model is going to struggle unless the market stays frothy.
Also, you don't need to mark up tokens much if you're a commodity volume business. Think of Costco and their margins & membership fees. Not everything has to be high margin, not everything has to be a SaaS subscription.
Just because some Higher Ups in Stripe bought this, doesn't mean its that valuable.
It can easily be, that Stripe can just afford it and think that they are part of the big boys and thats just what companies cost today.
It could be that Stripe really really like the brand name and all the positive notion of it.
It could be that Stripe doesn't want to build this from scratch in a timefrime of 1 year or 2 because Stripe might be too corporate to be able to do startup stuff.
Stripe has quite a high motivation to leverage agents thoguh because they are preparing for Agents which will buy through stripe. They already provide the SKU backend and support the agent payment stuff.
I don't see how stripe adds any value here (and I've had such terrible experience with stripe automatically breaking my stuff I am worried I can't trust openrouter now) but I can see stripe wanting to be in the middle of any two people giving each other money on the internet and this is squarely (lol) that.
The other thing is convenience and centralized security from using one gateway to access everything. It's a lot better than having to deal with N accounts with separate limits and monitoring. And giving your payment details to one company instead of 20 is obviously safer.
Investors like them because the pricing is inherently usage based so there's zero risk of clients using more tokens than what they paid for. Guaranteed profit as long as they can keep a modest amount of customers.
If compute is constrained and expensive, OpenRouter is what you'll use to get around the constraints at individual providers.
[0] https://x.com/dwarkesh_sp/status/2084333160075055122?s=20
Can’t explain the valuation since everything in this space is rationally overvalued, but I don’t think OpenRouters valuation is that surprising, all things considered.
Just guessing. The frenzy around popular, good already, and successful services with the corporate crap flowing from this acquisition announcement too is appalling. The "what's best for you, the user" heavy emphasis when this would be inherently evident in any honest service forecasts the opposite.
Some highlights from one of my agents asked for a no bullshit evaluation:
"By buying OpenRouter, they own the routing layer that decides which model gets called and at what price."
"Stripe wants to be the economic infrastructure for AI — meaning they want to be the toll booth and the traffic cop for the entire AI economy."
"... insider market intelligence that OpenAI, Anthropic, and Google don't fully have. Stripe would now have it."
"The $7B+ price is absurd on any financial metric, but this isn't a financial acquisition — it's a strategic positioning play to own the platform layer of the next computing paradigm."
"140x revenue multiple: At ~$50M revenue, this is not a financial valuation. It's a strategic land grab — buying the chokepoint before someone else does."
I do not feel a particularly strong smell of 'best for the user' here for some reason... More like the usual 'how do we squeeze out more for our PE folks from this' kind of scent.
Using something like OpenRouter (or any of the AI Gateways) is better than tying yourself to one LLM provider that can rug pull on pricing or change models in a way you don't like.
The value is in the network effect I think. OpenRouter is popular and has a good head start over anything Stripe could build internally.
We are basically back to 2020 trying to get people to understand that an NFT is NOT the underlying asset it abstractly represents…
Please go on as I do not get your point.
Tokens don’t have an inherent value and have no other properties that would make they tradeable.
Another way to think about it: a company that spends $1M on tokens is burning that money hoping they will be able to recoup by generating a better product. If it would be tradeable, they could spend $1M to generate tokens, produce absolutely no product, and just resell all those tokens to get their money back. That’s obviously nonsense, that’s not at all what tokens are. And they cannot transform the tokens they got into compute, you need to consume hardware and energy to mint tokens, you cannot convert tokens into hardware
No one is talking about trading tokens.
The idea being that getting your tokens from Provider A is no different than Provider B, especially if they both offer the same model. You can change one value in a request to openrouter and suddenly be hitting a different provider but offering the same tokens, because they offer the same model and the same settings.
You're describing a commodity. However, the commodity isn't the tokens, but the compute capacity, i.e., serving a model. And compute capacity isn't a currency — at least not, until you can acquire compute capacity from one party and exchange it with another party.
dgellow is absolutely right: tokens aren't a currency, nor can you trade them, nor are they fungible. The original claim that "tokens are just a type of currency" [1] makes no sense.
Surely stripe if anyone have learned to harness the cash flowing through their system. Hell, they could be emitting bonds on expected token consumption bills!
https://news.ycombinator.com/item?id=34423387
are there more examples of this? a bit of an HN anti-portfolio. good reminder to others.
love that Dropbox has became textbook example of viral marketing
But such is normal with normies providing startup critiques, if you just look at every announcement and say "this will fail", you'd be right 99% of the time.
So nothing really has changed in this community's vibes.
I mean - just looked at this downvoted comment essentially saying the same thing - https://news.ycombinator.com/item?id=49119464. And also the intensity of the replies.
(i know much of it is a crap shoot and easy to quarterback on Monday morning.)
Isn't the trend increasingly to let an orshestrator agent decide which model to use in terms of problem difficulty vs. size and thinking budget?
Once you fix those parameters there's little to optimize over, no? I mean you want to find the cheapest one across providers but that's not a recurring decision.
In the second case OpenRouter is very useful; in the first case, maybe it’s an area they move into in time - providing and owning/optimising the orchestrator?
Convenience. Paying only one provider, but get access to different models hosted around the world. Need a model for a project where latency matters? EU hosted and ZDR for another customer? Cheapest for personal use?
You get everything under one roof.
One great example is Open Banking. I would support an Open Router that mirrors Open Banking, unfortunately the Open here doesn’t carry the same connotation.
Kudos to the team for succeeding in executing around a clear DX issue during the chaos of providers, though long term I am not really a fan of it staying as is even if Stripe gets to be the one carrying it forward.
Open banking is slightly different, as it's more about account access (although with read+write access, you can kind of build it into a payment scheme as well, although it's usually clunky).
The brazilian central bank has a page about it[1]
[1] https://www.bcb.gov.br/estabilidadefinanceira/openfinance
[1] https://www.mckinsey.com/br/~/media/mckinsey/locations/south...