Token-maxxing is dead, OpenRouter's token usage plateaus | Hacker News Reader

Token-maxxing is dead, OpenRouter's token usage plateaus

openrouter.ai

There are a few problems with using OpenRouter's public statistics to make statements about the wider field:

- It only tracks companies that have decided to use OpenRouter to help them switch models - which biases towards companies that like open weights. Anyone using OpenAI or Anthropic directly is much less likely to be represented.

- OpenRouter don't share the number of unique users represented in those numbers. I worry a lot about whales - are there a few giant users of OpenRouter who, if they change their preferred model, instantly affect the global leaderboards?

In this case I suspect the numbers are affected by the release of Claude Haiku 5.5 and GPT-6 Luna - both excellent inexpensive models, which may well take market share from some of the models that OpenRouter is serving. OpenAI and Anthropic have the most mature client support so I expect it's more likely people will switch away from OpenRouter if those are their daily drivers.

Happy to be convinced otherwise if anyone has useful data that I've missed.

Luna is incredible for the price but I find 6 to be worse than 5.6.

I had the same thoughts. I like OpenRouter a lot and still use them, but these numbers make me think more about how much more crowded the space has become. In the last year, you've had startups completely pivot into this niche (Respan and Requesty off the top of my head) who have lots of usage, infra companies like DeepInfra have really leaned in hard to the "serverless interface for open models" positioning, some of the popular harnesses that previously were complimentary to OpenRouter now sell their own gateway/subscription (OpenCode, Nous), and the subscription discount for Claude/Codex is wildly lopsided vs API pricing for solo devs.

Developers are no longer as excited about this stuff, and non developers either don’t care since it doesn’t solve their issues or quite openly hostile to this tech… companies are also starting to back off since other than some very limited cases it’s just extra costs that no one wants to see.

Are those just your opinions or can you back them up with something more?

Just some observations from largish work environment

Re: Companies starting to back-off, I've seen several articles like this one https://futurism.com/artificial-intelligence/enterprise-ai-s... (yes I realize this source is generally anti-ai)

Regarding regular people being ambivalent or hostile, I feel like that's pretty obvious, at least in the western world. Across all social media, outside of tech-chambers, the anti-ai sentiment is very loud. Nearly everyone posses a misplaced hatred for data centers, and this both fuels and is fueled by ai-hate. There's massive backlash from gamers and inevitable walk-backs whenever a video game company proposes using ai for anything.

Eh, anecdotal, but this tracks with the perception I have both from current and former coworkers.

There is a dwindling subset that are still very excited for it, but the the majority it either became a boring tool. A growing minority harbors some resentment for it as it made the acting of worse less tolerable.

"Companies"? A huge proportion of Openrouter tokens are coming from Hermes. That is not "companies", that is "AI bros using Hermes to do things that could be done with a dozen lines of python" and inflated by the massive overhead that is the system prompt hermes generates for each request. Even with just a few tools and skill files it can reach 20-30k tokens.

I'm not kidding when I say that if anything comes out which substantially improves Hermes token efficiency,or causes a large number of Hermes installations to fail or stop being used (like a major security vulnerability), Openrouter and some of their providers will be in trouble.

Agreed, OpenRouter is competing in an increasingly crowded market and even a lot of cloud providers like Neon, Vercel and Cloudflare are rolling out their own gateways to vertically integrate with their hosting. It's hard to say what the true net usage of LLM look like in the world!

I agree with you about the sources of bias, and I’ll add one more: the acquisition of OpenRouter has led to a fairly clear step off the gas pedal in terms of marketing & mindshare.

Nonetheless, the AI market is supposed to be a rising tide that lifts all ships. That’s where the pricing is - indefinite perfection. If we’ve suddenly moved to a world of zero-sum competition, that’s significant.

Or maybe Codex/Claude Code got better and more people are just using those?

Or more people are using harnesses that do what openrouter does so they don't need it anymore?

Or they are using other GPU providers directly to run their open source models?

I don't think this data is conclusive. It's a very unique subset who are using openrouter.

> Or maybe Codex/Claude Code got better and more people are just using those?

Anthropic’s revenue growth has leveled off, too. Not the exponential hockey stick that was predicted.

I don't think you can draw an accurate narrative from two data points but I'm going to do it anyway. This and the slowing in revenue growth at Anthropic are not terrific signs. This should be the era of exponential growth for the AI industry. Even if the unit economics are sound at these prices, companies really need more revenue to service their current debt obligations and for now that can't come from price it has to come from volume. Furthermore, this is OpenRouter's statistics which partially represents the cost conscious section of the market and this doesn't line up with an induced demand story.

Again, two data points do not make a full story but if you're an AI bull this is not what you want to see.

Not terrific signs for the AI industry or the economies that have come to depend on them, perhaps.

But it is only two points of unclear data among a sea of exponential growth stories. It is just that the total debt amount and rates environment mean the industry has to hit an extremely small target which prices everything toward perfection.

I think part of the issue is that SWE work is slowing down on needing more context and token throughput (see Opus 5.5) and new industries that _can_ adopt AI right now just don't need insane amounts of tokens flowing through it, for instance Legal (in the context of people using LLMs for Legal correctly), where you kind of just need the best model and context is maybe a few dozen PDFs relevant to your task.

I do find it pretty extraordinary that Anthropic's revenue growing from $4.8bn in Q1 to $11.5bn in Q2 to $16.8bn in Q3 is seen as a disappointment, but they did indeed stop doubling on a per-quarter basis.

I don't have the numbers handy, but they spent drastically more than that each quarter. Growing revenue is all well and good, but profit does eventually matter, especially if debt dries up.

I don't think that's true in 2026 - they claim that two of those three quarters were profitable.

You may be thinking about their 2025 numbers, which have been widely commented on recently, often in a way that failed to separate them from 2026.

> ...they claim that two of those three quarters were profitable.

This is the same company that claims -among so many other things- that:

* 100% of its employees believe that the over-arching project they're working on is likely to kill all humans either by the end of the decade, or in ten years

* not only is it completely incapable of working on said project safely, it is impossible for the company to stop working on said project unless they get special laws and regulations made just for them (the implication here being that they play a very large role in creating said laws and regulations, given that they're the SMEs and all)

* that their super-duper LLM-based vulnerability scanner was so powerful that access to it absolutely had to be tightly controlled lest The Bad Guys use its awesome power to pwn every bit of software on earth [0]

I don't trust their claims. If we still had a well-functioning investigative press, I expect we'd see the majority of the big claims coming from the Major LLM Manufacturers pretty quickly either exploded or substantially revised into much smaller claims.

[0] ...that this pretty neatly feeds into their long-running narrative of Imminent And Extreme Danger is just a coincidence, honest!

I rate the financial claims they make to their investors as more likely to be true than the blustering hype that comes out of them the rest of the time.

Lying to your investors is fraud.

> * that their super-duper LLM-based vulnerability scanner was so powerful that access to it absolutely had to be tightly controlled lest The Bad Guys use its awesome power to pwn every bit of software on earth [0]

Their actual claims were demonstrably true. Just look at how the rate of CVEs has gone up. Or, alternatively, look at what can be done now with the open weights models that are only a few months behind the SOTA models.

> * not only is it completely incapable of working on said project safely, it is impossible for the company to stop working on said project unless they get special laws and regulations made just for them (the implication here being that they play a very large role in creating said laws and regulations, given that they're the SMEs and all)

Even if you think this is all BS/regulatory capture: https://en.wikipedia.org/wiki/Prisoner's_dilemma

The way around everyone defecting in the prisoner's dilemma is to have enforcers.

> they claim that two of those three quarters were profitable.

no they didn't. What they said was "positive adjusted operating income". The most important word there is "adjusted", ie. they are not profitable if you include all costs

Do we agree that they spin the delivery there to have people hear "we were profitable"?

They don't seem to do this by accident, and the only reason I see would be to deceive investors or potential investors.

I think it's a combination of it not being a GAAP revenue number and the timing. If you have flexibility in the report of the revenue number people assume you will choose the most flattering one especially before an IPO. If this is the most flattering number they can demonstrate, people are going to worry about what the not so flattering data says.

The other part is how big the debt load is downstream. Anthropic is one of the revenue engines that has to feed the rest of the supply chain, clouds, SPVs, semiconductors, etc. They have to be pulling very large numbers to justify everyone else's finances.

I think the growth is real and tremendous and that's probably more important than this blip in revenue growth. But things tend to look peachy until they're not and so everyone is on watch for cracks in the foundation.

> This and the slowing in revenue growth at Anthropic are not terrific signs.

The CEOs of both Broadcom [0] and Intel [1] publicly noting that the various LLM companies are going to have a difficult time paying for the things they need to pay for and Oracle [2] continuing to prohibit LLM-generated code in OpenJDK due to both copyright and licensing concerns [3] are also not terrific signs.

[0] <https://www.reuters.com/world/asia-pacific/broadcom-ceo-says...>

[1] <https://www.inc.com/moses-jeanfrancois/intel-and-broadcom-ce...>

[2] ...that is, the "law firm with a great big sales department and a software development division tucked into a broom closet somewhere" company ;)...

[3] <https://openjdk.org/legal/ai>

If you're looking for indicators one way or another, I have yet to see any economic productivity indicators that would show any correlation between productivity and LLM use.

I've been using Openrouter for API calls and Claude Max200 for CLI coding. Both for about 18 months now.

I have several apps hooked to openrouter for workload processing. Mostly OCR and some inference and some RAG + prompt + output stages. It works great.

I literally just got a note from Anthropic that my Max200 now includes API credit. On receipt I ported over my busiest app stage from openrouter/gemini to haiku, with openrouter acting as an overflow. $200 saved. Boom.

I don't think this is bullish for Openrouter. It took claude 2 minutes to build a primary/secondary API callout for my workload. That's not a very deep moat.

I will still use openrouter to try new models and A/B test "current hotness" for my assorted tasks, but the spend from me is gonna drop, by, well, $200/mo or so.

I'm actually gonna go against the grain here and say that token maxxing is not dead. In fact, it's just getting started. Maybe not in the short term. I mean, who knows how things are gonna shake out financially speaking. In the long term though, I think people are really just getting started figuring out creative ways to leverage these agents. And especially when they keep getting smarter over time, the surface of use cases is just going to grow and grow and grow. I think the present plateau is simply just a function of costs being a little too high for comfort at the moment combined with the technology just moving so fast that I think it's kind of outpaced us a little bit.

While we don't "token-maxx", our usage of Kimi K3, GLM 5.3, MiMo and Qwen goes up, up, up -- with Kimi K3 dominating. Our usage of Claude Code and OpenAI has been nonexistent for a year and a half...

Opus 5.5 is in a different category than all those Chinese models - as long as anthropic allows it to work on your stuff (which is a thing now too).

Depending on workload it could be like renting a bazooka to kill a few insects.

So you say, but I've done plenty of blind taste tests. The Chinese models are 99% as good for my purposes, and, I think, 99% of applied business programming purposes.

Isn’t Kimi k3 more expensive than Opus 5.5 in the benchmarks?

Of course, benchmarks don’t tell you how much it will cost on your workload.

GLM 5.3 was more expensive than Haiku and Sonnet 5.5 on a workload I was testing recently despite being much cheaper on the price sheet.

Most user here using Deepseek flash or GLM flash which are cheaper.

To be frank, Kimi's a better model...

I used to think that until I started using DeepSeek V4.1 Flash 1017 for the past 2 days (what a bad name). It is terribly fast! And smart. And cheap.

^ This. HN crowd is way too willing to presume the superiority of Anthropic and OpenAI frontier models, or to dismiss the Chinese ones as some distillation parlour trick.

While that is unquestionably going on, these models are legitimately innovative and stand on their own legs.

After getting a shock -$30 balance just cache writing a session into Astra, I started to finally go down the rabbit hole of token efficiency. There is a lot happening in the space of task routing, and portable memories, and self training classifiers on task success etc. It got particularly confounding tying to automate keeping track of all the different ways the different providers keep track of billing and credits, and account types, and incentive tiers..

I don’t think we can conclude that from this graph. Look at all the earlier points in time where it “flattened out” for about as many months. We tend to unconsciously overlook them because, with a sharp rise on either side, it obviously looks like noise.

I can’t make the same assumption here because I can only see what’s on the left, but that in no way implies that the opposite assumption is any more warranted.

The bigger problem is not token consumption plateauing, total token spent is lot lower than before. we are at %20 of our total peak spent. compared to couple months ago.

We are able to purchase similar level intelligence at much lower prices now. Even sonnet 5.5 is becoming an overkill for most of our agentic workload.

ps. our total token usage last 30 days is around 30B

I've had a hell of a time with more complex model setups in open router and have gone to just using the underlying APIs directly.

Elaborate on the setup please?

I use the api from the model provider?

I mean what is so complex about your setup

Is this 7 days rolling, or for the week that just started, compared to the last full week? Everything up for the day chart.

It is Mon-Sunday. They do a projection for this coming week (based of today usage) but it is often wrong. Headline is premature, only last week (12-18) was a little below (5-11). pretty meaningless and has happened for two weeks before (example: end of July)

Idk about others, but the usage that I'd usually put into OpenRouter is now going to OpenCode Go.

Absolutely sure this is bullish.

It is. The graph on their front page shows exponential growth.

Yeah, after scrutinizing the graph more closely, I'm actually surprised that people are even taking this headline seriously. There's been far larger "plateaus" in the past. And the growth in usage just this year, just in the last couple months, is just astronomical.

Dead just like it was in June?

What? Setting aside the valid arguments elsewhere, the graph doesn't even show that. Switch it to "log" and look at a straight line through the data points. The recent data points are within regular variation.

Reply on news.ycombinator.com