YT YouTube2Text

YouTube2Text → Transcripts → OpenRouter State Of Models: Jev, Open-Weights, and Tokenomics

OpenRouter State Of Models: Jev, Open-Weights, and Tokenomics — Transcript

by IndyDevDan · 5,243 words · 765 segments · language en · Watch on YouTube

Open in transcriber Summarize with AI Download SRT/VTT

Full transcript

0:00What's up engineers? Andy Debb Dan here.
0:02With the release of the new Jev system 1
0:04model class and all the overnight
0:06clones, [music] this is a great time for
0:08us to break down the state of AI models
0:11as we enter the last quarter of 2026.
0:13[music]
0:14Why is that? Because it's important to
0:16understand the intelligence available to
0:18you and the trends around them so you
0:20can do [music] your best work at the
0:22right performance, speed, cost
0:24trade-off. I have five big trends we'll
0:26break down in this video. I'll throw
0:28them up on the screen now. We'll [music]
0:29use open router to explore these trends.
0:31If these ideas interest you, stick
0:33around and let's scale our compute so we
0:36can scale [music] our impact. Trend one
0:38tokens go up.
0:45Quick announcement. Early sign up for
0:47the phase 3 [music] Aentic engineering
0:49product is officially available. Sign up
0:51to get on a private email list to
0:54understand what advantage is coming next
0:56[music] to you and every engineer who's
0:58on the edge. I can't wait to share this
1:01with you. This will be by far one of the
1:03most valuable things I've ever created.
1:05Look at Open Router's weekly token
1:07chart. Usage keeps going up, but you
1:09have to look at the numbers to
1:11understand how astronomically high token
1:14usage is now. Roughly one year ago
1:16today, there was 4.5 trillion tokens for
1:20the entire week. Guess what that number
1:22is today? It's not 10 trillion, not 20
1:24trillion, not 50 trillion, not 100
1:27trillion. Last week, there was a total
1:29of 146 trillion tokens used. It's not
1:33exponential growth, but it's close. Some
1:36people keep calling AI a bubble, but the
1:38growing usage numbers we can concretely
1:40see on Open Router shows demand for
1:42these tools that's nearly exponential.
1:45And you've likely seen this in your own
1:47work. The more tokens you can gain
1:49access to that's intelligent and cheap,
1:52the more use cases you keep finding.
1:54This is true for you, it's true for me,
1:56it's true for the entire engineering
1:58ecosystem. And we're all slowly learning
2:00how to make our tokens more economically
2:03valuable. This is the art and science of
2:05tokconomics. Now, you'll notice
2:06something really interesting here. There
2:08are barely any anthropic models in the
2:11top LLM leaderboard. That's because
2:13OpenR does not count tokens from other
2:15language model providers. Anthropic is
2:18targeting a massive $2 trillion IPO
2:20despite losing almost $50 billion. And
2:23this makes us ask an uncomfortable hard
2:25question with all the great growth
2:27Anthropic is seeing with all that token
2:29spend coming into Anthropic. Is this $2
2:32trillion IPO value justified? And the
2:35real question is, can their explosive
2:37growth continue? Will we keep seeing the
2:39token spend that we see on Open Router
2:42come through to the model providers? If
2:44Anthropic has any chart that looks just
2:46like this and their usage, the answer
2:48will be yes. They will make $200 billion
2:51in annualized recurring revenue by the
2:53end of 2028, which is their projection.
2:55That's how they're able to make this
2:57insane IPO claim. So, we have a couple
3:00really, really interesting narratives
3:01going on here, and it all depends on
3:03language model usage continuing to go up
3:06nearly exponentially. If this trend
3:09continues, we can put our fist down and
3:11say there is no bubble. But if this
3:12trend does not continue, I think the
3:15bare case for Anthropics IPO is going to
3:17be true and then it'll likely follow
3:19with OpenAI and then the House of Cards
3:22starts falling down. I'm more in the
3:24optimistic camp, but pretty soon we're
3:26going to see which one of these two
3:27cases is true. Regardless, what's clear
3:30here is token usage is going up. What
3:32does this trend mean for us? I think the
3:34key question every engineer needs to ask
3:36is, is your token spend going up because
3:38you're getting more useful work done or
3:40is it going up because you're trying to
3:42use more tokens and you can't find a way
3:44to use them efficiently? Are we getting
3:46more productive use out of our token
3:48spend? This is tokconomics. It's not
3:51enough to just use more tokens. I always
3:53say on the channel, scale your compute
3:54to scale your impact. But it really
3:56matters what you're doing with these
3:57tokens. A lot of engineers are opting
3:59for cheap models and they sacrifice
4:01their own time and their own business's
4:03performance while other engineers are
4:05able to leverage these models properly
4:06to get the job done at cheaper prices.
4:08You want to be the engineers in that
4:10second bucket. You want to be utilizing
4:11these models without sacrificing your
4:13performance. And the Jev model by Typesa
4:16is a huge contributor to that. We'll
4:18talk about that more in a second. You
4:19can see the astronomical growth here.
4:21257% up week overweek coming in at 2.6
4:25trillion tokens. More on that in just a
4:27second. And this is the game we're all
4:29ultimately playing. More tokens does not
4:31equal more value. It's highly
4:33correlated, which is why I like to say
4:34scale your compute to scale your impact.
4:36But burning tokens to feel more
4:38productive is not progress. If it was,
4:40vibe coding would be the ultimate thing
4:42to do. It'd be the best place to spend
4:44your time. But it's not. Generating code
4:46is the easiest part of software
4:47engineering. The hard part is everything
4:49else. I like to aim for a steady rise in
4:52my compute usage versus an astronomical
4:55dump of tokens. Now, this brings us to
4:57the interesting question of given the
4:58entropic numbers and given the narrative
5:00around language models, who's winning
5:02this? What model provider? What model
5:04lab is actually winning the AI [music]
5:06race?
5:12So, who's actually winning the AI race?
5:14Just by glancing at the market share on
5:16Open Router, you see Open Weights and in
5:19the top three spot, we see DeepSeek,
5:20Google, and OpenAI. Now, the more
5:22fascinating part here for me is this. If
5:25we hover over, you know, a weekly
5:26breakdown, notice who you'll see in the
5:28top three spots. Deepseek, OpenAI,
5:31Google, Deepseek, OpenAI, Google,
5:33Deepseek, OpenAI, Google, Deepseek,
5:35OpenAI, Google. Almost the entire year
5:38with others, which is, you know,
5:40everyone else that's not in the top 10
5:42popping up sometimes in third place. So,
5:44why is this? You know, everyone keeps
5:46saying that uh Google has fallen out of
5:47the AI race and they're not relevant. I
5:50have no idea what they're talking about.
5:51Uh Google with the Gemini Flash series
5:53is one of the most effective workhorse
5:56models available. It's one of the most
5:59useful models available in my model
6:01stack. You can see this. It's one of my
6:02favorite models in that A tier that can
6:04do 90% of everything you're trying to do
6:06with your language model. And inside of
6:08my PI coding agent, Gemini 3.8 Flash is
6:11my default. And there's a great reason
6:12for this. It's because it has the best
6:14trade-off between performance, speed,
6:16and cost. For everything it does for
6:18you, for everything this thing gives
6:19you, it's just a fantastic model. Now,
6:21there are a bunch of competitors in the
6:22A tier right now. This is definitely the
6:24most contentious place to be for a
6:26language model right now. But that's a
6:28great thing, right, for you and I, the
6:29engineer, the intelligence explosion is
6:31continuing, right? We just had the new
6:33uh cloud opus 5.5, we just had sonnet
6:35drop, we had the 6 soul and the six luna
6:39model drop. And uh, you know, there's
6:41more compute available than ever. We
6:43have to give Luna a big shout out here.
6:45This is a really, really powerful model
6:46that frankly is keeping OpenAI in the
6:49LLM rankings on Open Router, right? A
6:51lot of this volume is Luna. These three
6:54companies are sitting at that
6:55intersection. They're sitting at that
6:57sweet spot, the workhorse A tier model
7:00level where it's good enough to do
7:01useful work. Deepseek, Gemini, and
7:03there's Luna and Terara is here as well.
7:05This one is less popular, but they're
7:06sitting at that great workhorse level
7:08where cost, speed, and performance
7:11really intersects. As mentioned, Open
7:13Router traffic only tracks models going
7:15through Open Router. When I'm looking at
7:17the anthropic IPO coming up, I'm asking
7:19myself, this is like a fun thought
7:21experiment you can run. How many tokens
7:22of this 65, let's round down, 60 billion
7:25of the rumored current annualized
7:27revenue run rate that Anthropic has is
7:29coming from raw token spends. Like
7:31what's the total token count from
7:33Enthropic? So you can quickly easily do
7:35this math. Let's just assume that $60
7:37billion run rate. What if it were all
7:39Opus 5.5 tokens at Opus 5.5 pricing?
7:42Just to put things into perspective, it
7:44looks like open weights models are
7:46winning and these cheap fast performant
7:48models are completely dominant. But if
7:50you just do a little bit of on the
7:51napkin math, you'll find something very
7:53different here. If all the 60 billion
7:55run rate where Opus 5.5 tokens, that
7:57means that Enthropic is processing 65.2
8:00trillion tokens per day. As we just
8:03looked at, we know this Open Router is
8:04doing 140 trillion per week. That's
8:07insane. That means they're doing
8:08something like 450 trillion tokens per
8:11week. And if you don't like my
8:12simplified numbers here, right? Assuming
8:13that this is all Opus tokens, which
8:15they're not, cut it in half, right? It
8:17doesn't really matter. You can cut it in
8:18half again if you want. That means
8:19they're doing 32 trillion per day or 228
8:22trillion per week, which is still 1.5x
8:25the total volume on Open Router. So,
8:28don't fully believe this market share
8:30here. You have to really really
8:32understand that the AI labs are getting
8:34massive, massive direct traffic. And of
8:37course, you know, Anthropic is known for
8:38being an enterprise leader. So, they're
8:40seeing these enterprise token spending
8:43numbers. Might be number seven down
8:45here, but don't let this fool you,
8:46right? [laughter]
8:47These open router numbers are really a
8:49representation of individual engineers
8:51and small to medium-sized businesses
8:53spending tokens. I'm sure Open Rider has
8:55some enterprise, but the real enterprise
8:57are fully bought into Anthropic, OpenAI,
9:00Google, and maybe a couple others. Who's
9:03winning the AI race? Right? You see a
9:04bunch of openweight models here on Open
9:06Router. Who's really winning? Everyone
9:08is winning. Again, let's scroll back up
9:10to the top models. It's just spend
9:12across the board is going up. It's going
9:14up massively. Something like 50, 60, 70,
9:1680x of increase over the entire year. I
9:19remember looking at token volume a year
9:21ago on open router thinking, "Wow, these
9:24numbers are insane. They're insane.
9:26They're insane." But they just keep
9:27growing. So, who's winning? Everyone is
9:29winning. Token spend is going up across
9:31the board. Chinese open weights and now
9:33some close weights models are completely
9:35dominant here in terms of this specific
9:37model that's executing. But we have to
9:39give a lot of credit to the frontier
9:41model labs for paving the way. And then
9:43we have to give credit back to open
9:45weight models coming out of China for
9:46making it affordable. But then we have
9:47to give credit back to open AAI and to
9:50anthropic because guess how these models
9:52are getting trained. They're getting
9:53trained from distillations from the
9:55state-of-the-art is a very popular very
9:57well-known thing that's happening now.
9:59So there's a circular optimization
10:01that's happening at least from you and
10:02I's perspective right from the engineers
10:04perspective cost going down,
10:05intelligence going up. It's a great
10:07thing. Everybody's winning. There's no
10:08one model wins. There's no one AI lab
10:10wins. This is the optimal scenario we're
10:12in here. But something super exciting
10:14has happened recently. You know about
10:15it. You know, it's in all the headlines
10:17right now. Jev, we have a new class of
10:19models to work with. What's old is new
10:22again. We have zero shot classifier
10:24models back on the scene with Jev
10:26[music] leading the way.
10:33Trend number three is a old new model
10:35class led by the dev model. So, first
10:38off, 2.68 trillion tokens up 200%. This
10:41is the fastest grower outside of Mimo
10:44Flash and the rumored uh Space Bunny
10:47alpha model which is rumored to be a new
10:49Minimax model. So Jev is seeing
10:51explosive growth here. One of the crazy
10:53things about Jev and the usage numbers
10:56behind Jev on open router specifically
10:58is that Jev uses a fraction of the
11:00tokens that language models do when
11:02they're inside of agents. And it is
11:04still 12th place this week. This just
11:07came out. It's brand new technology or
11:10it's old technology made new and highly
11:13optimized. You've seen all the Jev
11:14clones come out. Everyone talking about
11:16fine-tuning models. Every other engineer
11:18saying that they've done this before.
11:19Yeah, this is classifier models. I
11:20myself personally have worked on a lot
11:22of these traditional classifier, you
11:25know, scikitlearn models. But it's not
11:27that simple. And let me show you the
11:28chart that proves it. Jev from Typesafe
11:30is one of the most scalable models I've
11:33ever seen. And it's one of the most
11:35reliable models. This is a really
11:37important chart that not many people
11:39will pay attention to. Uptime Jev just
11:42doesn't go down. Now, this is in
11:44contrast to many, many, many, many, many
11:46other models with much worse, you know,
11:49latency and much worse uptime. Speak of
11:52the devil, I'm filming this right now
11:53and Claude just had another outage, a
11:56blip in the system. But it's insane to
11:58me that uh everyone thinks that they can
12:00create a Jev clone when they're not
12:01really understanding that this is a
12:03super generic classifier that's zero
12:06shot that can be applied to many many
12:07different scenarios. Now definitely if
12:09you have a bunch of training data, you
12:11can outperform Jev with a small model.
12:13But what you won't outperform is Jev
12:15across many many different use cases.
12:17Contrary to popular belief, Jev is not
12:20easy to replace. Engineers that were
12:21with the channel last week know that I
12:23am actively embedding Jev into my PI
12:26coding agent. I'm already benchmarking
12:28this model. Last week we covered 10
12:30levels of Jev for a Gentic engineers. Uh
12:32when you finish this LLM breakdown,
12:34check out that video to understand how
12:35to use Jev for real production use cases
12:37all the way up to powerful agentic use
12:39cases. But you can see here just with
12:41the Jev model just one additional model
12:43and a couple of tool calls I am
12:44outperforming on both speed and cost.
12:47You can only see cost here. I need to
12:48add speed to this PI agent comparison
12:50UI. You can see here, you know, on a
12:52relatively simple 400 500K input um
12:55prompt. I'm saving 5 cents and
12:58specifically I'm saving about 20%. The
13:00percentage is the thing to pay attention
13:01to as I'm benchmarking using Jev, a
13:04classifier model inside of PI. You
13:06really want to keep your eye on this.
13:07I'm using the new Sonnet 5.5 model here.
13:09By the way, this new classifier model is
13:12groundbreaking. Why is it
13:13groundbreaking? Why does Jeb matter so
13:15much? It's because this is a new species
13:17of models that you can use alongside
13:20traditional language models and agents.
13:22As viewers of the channel know, I think
13:23ins ors. It's not about this model or
13:26that model or that provider or that
13:28model provider or that one or this tool
13:29or that tool. It's about using them
13:30together. Think ins. All right. It's
13:32Deepseek V4.1 Flash and Jev. It's Claude
13:36Opus 5.5 and GPT6 Astra together.
13:39Combine compute. Don't select compute.
13:40You want to be using all these models at
13:42the right time in the right combination.
13:44And Jev is a new species of model that
13:46we can now use. Interestingly, Open
13:49Router has realized this. They have this
13:50new, if we go to models, they have a new
13:53decisions category specifically for
13:55models like this, uh, shout out Jev for
13:57being one of the first ones. And now
13:59everyone else is is coming out of the
14:00woodworks to create their own Jev clone
14:03or classifier model that does roughly
14:05the same thing. I think they're going to
14:06find that what Jeb actually gives you in
14:09addition to like a great implementation
14:11like they've really nailed this null
14:12choice score. They have more types on
14:14the way. They have more models on the
14:15way which I'm really excited about. What
14:17they've really nailed is the speed. So
14:19the latency is consistently great and
14:21they've nailed the reliability right
14:24which a lot of the leading model labs
14:26need to take after. I know Jev is much
14:28simpler but reliability has been a
14:29massive second priority, third priority
14:32for these model providers. So really
14:34really happy with Jev. This is an
14:35important release. I am right now in a
14:37flurry to explore Jev, looking at
14:40additional use cases, really figuring
14:41out how to use this model at scale
14:43broadly. As you can see here, I am
14:45actively rolling this into my PI coding
14:47agent. I have this in my default model,
14:49so my agent can make quick decisions
14:51when it needs to. As we covered last
14:53week, you can save a bunch of input
14:55tokens. And you can do this at scale
14:57with every agent you execute. My
14:59estimates are coming in. I'm saving
15:00about 20% token spend with Jev than
15:03without it. That's going to add up a lot
15:04day after day, week after week. So keep
15:06your eye peeled on these new classifier
15:08models. Experiment with them. Don't just
15:10put them in your code. Put them in your
15:12agents. Use tools together. This new
15:14model is going to be around for a while.
15:16That leads me to the next really simple
15:18idea, really important idea. Uh just to
15:20remind engineers, let's talk about quote
15:21unquote free tokens.
15:28Just want to leave a quick note here on
15:30models like Space Bunny Alpha. I'm sure
15:31it's some great new incredible model.
15:33It's rumored to be the next minimax
15:35model. One of the things I don't like
15:36about Open Router is that they do have
15:38this free model deal, which I know is
15:40like a great thing, especially if you're
15:42on a budget. You know, the problem with
15:44these models is that if it's free,
15:46you're the product. It's that same
15:47simple idea that's been true throughout
15:49all technology. You can be pretty
15:50confident that these model providers are
15:52retaining your prompts. They're reading
15:54your prompts. They're using them. They
15:55say they don't use them for training. I
15:57don't really buy that. That's kind of
15:58the trade-off here for models like this.
16:00That being said, you know, this model
16:01has jumped up into second place. So,
16:03it's clearly very incredible, very cost
16:05effective, but also it's just free,
16:07which really paints the picture of what
16:10the language model ecosystem
16:12appreciates, wants, and what we really
16:14need for models to be really effective
16:17at scale. We need them to be very, very
16:19cheap. We know cost is going down. We
16:20know intelligence is going up. But as I
16:22look at this trend, the usage of free
16:25tokens on Open Router, I'm constantly
16:27thinking usage is going to keep climbing
16:30and the cheaper tokens are while
16:32maintaining performance, the more use
16:33cases emerge. And so even though I don't
16:36like the free models on Open Router
16:38because I fully believe they're taking
16:40all your prompts, all your data, and
16:42training on them. You know, I'd rather
16:43pay than have them take proprietary
16:45information. But get it. The bigger
16:46trend that this paints for me is that
16:48models like this like point toward that
16:50future. It's a continuation of the
16:52trend. The cheaper models are, the more
16:54cost effective they are, right? Without
16:56losing too much performance, the more
16:57tokens will be spent. So, we can kind of
17:00see that across models throughout the
17:02market share. The cheaper models that
17:04are still effective tend to rise up,
17:06right? That's why once again, Deepc,
17:08Google, OpenAI are doing a really,
17:09really great job. This stealth model has
17:11exploded into near first place. It's
17:14pretty incredible how effective it is to
17:16release a lowcost model because it
17:18drives a lot of attention, a lot of
17:20usage solely based on its price which as
17:23I like to explore the open router usage
17:26data also shows a crack in this usage
17:28right because why is a lot of this usage
17:30coming to open router is because the
17:32models here are of a wide variety and
17:34they're also dirt cheap right some of
17:36them are even free because they train on
17:38your data so wherever you're getting
17:40your information sources from especially
17:42on these like large data uh analysis
17:44tools, you always have to understand
17:45who's providing the data, where do the
17:48stats come from, what are the users
17:49like. And you know, an open router, as
17:51valuable as this is, a lot of the usage
17:53comes from users that want price down
17:56over everything else. And that's in
17:58contrast to engineers using Claude, for
18:01instance, or OpenAI who want maximum
18:03performance, right? It's insane that
18:05Enthropic has these numbers or some
18:08numbers like this. Even if you drop it
18:09all the way down to Open Routers weekly
18:12numbers of 140 trillion, if OpenAI
18:14actually has these numbers, it's insane
18:16because it means that their tokens as
18:18they price them are much much more
18:21expensive than all the models on Open
18:23Router. Uh we do have that new Haiku
18:25model coming soon. We'll see what they
18:26do with that. Based on the Sonnet
18:28release, they're going to price it exact
18:30same. So, they're going to need to
18:31really create a price competitive Haiku
18:34model. I would love to see this come in
18:35under a dollar. that would make it
18:36potentially competitive with some of our
18:39workhorse models like the Gemini 3.8
18:41Flash. But at the same time, you know,
18:42I'm not really expecting Anthropic to
18:44move on their prices here. So, you know,
18:46that's kind of a caveat I like to
18:47mention here. I'm always cognizant of
18:50user usage data when I'm checking out
18:52these language model trends. As you can
18:54see here, just as a general trend, if
18:55your tokens, if your tool, if your
18:58products are very, very cheap to
19:00consume, but still provide value, users
19:02will show up. Now this is really hard
19:04thing to fully digest because you have
19:05to make money somehow off your tokens,
19:07right? This is tokconomics. So that's
19:09some insight into the market share. One
19:11last thing I want to point out and focus
19:12on is the top applications flowing
19:15through open router. Now this [music] is
19:17really fascinating.
19:23So there's only one big trend that I
19:24like to look for on top apps and that's
19:26where the PI coding agent is in relation
19:29to cloud code. For the longest time, Pi
19:31has been trailing cloud code in the
19:33fifth, sixth, seventh most used
19:35application through open router, but um
19:37it's been slowly catching up to cloud
19:40code usage. To me, this is a sign of a
19:42broader understanding of the agentic
19:45tools of harness engineering. The more
19:48engineers that realize how important it
19:50is that they own their agentic coding
19:52tool, that they can customize it, that
19:54they can control it, the bigger this
19:56number is going to get. And that's the
19:57trend I've been seeing very very slowly,
19:59very very smoothly. I'll be releasing my
20:01yearly predictions video coming up in
20:03December here. You know, just to give
20:04one of them away through open router
20:06usage, I expect Pi to take over cloud
20:09code finally for the first time ever in
20:11usage again on open router. As I
20:13mentioned, there are a bunch of caveats
20:14to the open router data as a reliable
20:17source because this number coming
20:19through cloud code is actually much much
20:20higher if you include anthropic
20:22memberships and the subscriptions. The
20:24numbers are much much much higher. This
20:26is still a very cool thing to see as
20:28engineers, you know, use more openweight
20:30models as they get hardware that they
20:32can deploy. I'm also really excited, you
20:34know, I'm going to be picking up one of
20:35the M5 Ultra Max Studios coming up here
20:38in October and I'll be benchmarking
20:40that. I'll be sharing all the stats
20:42around what models you can run on a 512
20:45unified memory Mac Studio. So, make sure
20:47you like, make sure you subscribe so you
20:49don't miss that video. That's going to
20:50send my personal token usage even higher
20:52and other engineers as well. again on
20:54these great openweight models that have
20:57been made available to us. You can see
20:58codeex generally lower usage here. Klein
21:01being the best I would say followup to
21:04pie coding agent. Kilo code not bad
21:06either. None of them get close to the
21:07simplicity and customization you get out
21:09of pi. So this is a trend I like to
21:10watch and keep my eye on. This is a
21:13large advantage that you can get as an
21:15engineer. If you control your agent
21:17coding tool and add more token spend
21:19here, you're going to have an advantage
21:21that other engineers don't have. Most
21:23engineers that watch this channel
21:24understand that by now, but if you
21:25don't, I'm going to leave a couple
21:27videos linked in the description where
21:28we utilize and really showcase what you
21:31can do with the PI coding agent. And one
21:32of them, I added several very, very
21:34powerful tools from Jev and coded it
21:36into the agent harness, added tools,
21:38modified the out of the box experience.
21:40You can see from that, by knowing how to
21:42harness engineer, I'm going to save
21:43about 20% token spend on every agent I
21:46boot up now. So, that's a really solid,
21:47very clear, significant advantage you
21:50get when you own your harness, not when
21:52you're renting them. We're renting
21:53enough stuff right now, right? We're
21:54renting intelligence. You want to rent
21:56as few things as possible, and you want
21:57to own as many as you can as it's useful
22:00for the time and investment you have to
22:02put into it, right? And I think owning
22:04your agent harness is absolutely worth
22:05it. So, those are the big trends I'm
22:08keeping my eyes on right now. The pie is
22:11growing. There's not a single winner.
22:13It's not open AI or anthropic. It's not
22:15closed weights versus open weights. It's
22:17everyone is winning, right? The pie is
22:18growing and the headliner chart is this
22:20one. Token spend is up something like
22:2280x year-over-year from 4.5 trillion
22:25tokens to 445 trillion tokens. It's not
22:29exponential, but it's close. Every time
22:31someone thinks something is slowing
22:32down, every time someone thinks there's
22:34a wall approaching, there isn't. Growth
22:36continues. I think the only question
22:37from a financial perspective is how
22:40harmful and how much is the US economy
22:44really banking on the anthropic IPO and
22:47the open AI IPO going well because this
22:49could pop the financial AI bubble. As we
22:52can see here, there isn't really an AI
22:54bubble. This is very raw useful
22:56technology. People point to crypto a lot
22:59saying that that technology was
23:00preemptive. I think crypto is also
23:02useful, but it was much harder to
23:04measure its usefulness. Now we know that
23:06crypto is also very useful, very
23:08relevant. But it's much easier to see
23:10here in the open router data, right? And
23:12again, this is just the public data.
23:14Enthropic openi, they have their own
23:15internal numbers that probably look a
23:17lot like this, maybe even steeper.
23:19What's old is new. Again, we have the
23:22Jev model creating space for new species
23:24of models, right? We have these new
23:26decision models coming into play here.
23:28These are very, very powerful models to
23:30help you make decisions in your work and
23:32in your software. Again, check out last
23:33week's video where we broke down 10
23:35levels of Jev for Agentic Engineers. The
23:37pie is actively getting bigger. There
23:40are no winners, but what there is is a
23:42bunch of options for you and I to scale
23:44their compute, to scale their impact. I
23:45think the big question for you and I
23:47going forward is constantly thinking,
23:49how can I increase my token spend in the
23:51most useful way? And so oftent times
23:53what I do is I throw state-of-the-art
23:56models at the problem I'm trying to
23:57solve. And then if I have to solve the
23:58problem repeatedly, I try to scale it
24:00down to cheaper, faster, more performant
24:02models. That's the big reason I like to
24:04maintain this model stack. Soon I'll be
24:06releasing in a gentically updated model
24:09stack that conglomerates a lot of the
24:11best benchmarks I look at for long
24:13horizon outloop agentic coding work. So
24:16you can, you know, play with that. You
24:17can analyze that to see how it compares
24:18to your model stack and how you think
24:20about the tier list of models. That's
24:23the name of the game. It's understand
24:24the technology, understand the
24:26intelligence available to you. So you
24:27can deploy it at the right performance,
24:29speed, cost trade-off for your work, for
24:31your use cases. I think when we look at
24:33models like Jev, we can expect to see a
24:36continuation of new species of models,
24:39not just language models to put in
24:41agents, but classifiers, decision
24:43makers, and other more specialized
24:45models that we can't even imagine right
24:47now. So big shout out to Jeff for paving
24:49the way. Again, check out last week's
24:50video to deeply understand Jeff. I think
24:53the most important thing every engineer
24:54can do right now is be prepared for the
24:57intelligence explosion to continue. That
25:00means using models like Jev to route to
25:02the right models to the right agents to
25:04the right workflows and it means
25:06understanding what model you need for
25:08the right time at the right price. You
25:10want to be managing your tokconomics for
25:12your inloop agentic coding work. And
25:13more and more the big unlock for
25:15engineers really focus on the signal is
25:17outloop agentic coding and more
25:20specifically outloop agentic
25:21engineering. For every engineer that
25:23made it to the end of this video, you
25:24know, big shout out to you. Thank you.
25:26That pre-signup for the phase 3 agentic
25:28engineering product is now available.
25:30Sign up is the first link in the
25:32description. There's going to be an
25:33optional quiz so I can optimize the
25:35product for you as soon as you sign up
25:37or as soon as you register if you're
25:38already an Aentic engineering [music]
25:40member. This will get you on the email
25:41list. We're going to be sharing
25:43exclusive big ideas and content from
25:46that next product that's coming [music]
25:47up. The Facebook product is consuming
25:49all of my time right now and making it
25:51the best thing possible for you and it's
25:53really turning out great. You know, it's
25:54crazy [music] to say this, but I think
25:55this is going to be the most valuable
25:56product I've ever created. Look out for
25:58an early December, late November launch.
26:01You [music] know where to find me every
26:02single Monday. Stay focused and keep
26:05building.

About this transcript

This page contains the full transcript of OpenRouter State Of Models: Jev, Open-Weights, and Tokenomics by IndyDevDan, generated from the public captions YouTube serves with the video. The transcript has 5,243 words across 765 segments, with the original timestamps preserved so you can click any line to jump to that moment in the embedded player.

What you can do with it

Use the transcript to take notes, quote the speaker, build a study guide, generate a summary with ChatGPT or Claude via the YouTube Summary tool, or export it as a timed subtitle file with YouTube to SRT. You can also re-open it in the transcriber to choose a caption track available on the video.

Free YouTube transcript tool

YouTube2Text is a free YouTube transcript generator — no signup, no daily limit. Paste any YouTube link and get the full transcript instantly, with timestamps, click-to-jump, caption language selection, AI prompts for ChatGPT, Claude, and Gemini, and exports to TXT, SRT, VTT, or Markdown.