Without Caching, Your AI Agent Costs 10x More - George Cameron, Artificial Analysis

AI Engineer · 19 min · 180 sentences · from YouTube's caption track

Each timecode opens YouTube at the start of that sentence. Line anchors (#s42) are the cue ids in the WebVTT, and every line carries its start and end seconds. All transcripts has every talk, and the whole corpus as one file.

00:01[music]
00:12Hey everyone.
00:13Thanks for coming.
00:14I'm George.
00:15I'm one of the co-founders of artificial analysis.
00:19Today I'll be talking through optimizing the cost of intelligence.
00:24To give a quick background to who we are at artificial analysis, we're a independent AI benchmarking company.
00:31We benchmark across the AI stack.
00:33So the agent layer without coding agent benchmarks and other agent benchmarks, language models for their intelligence, image, video, speech, and music models, as well as the hardware and inference layers.
00:47So that we can benchmark the tradeoffs between all these layers.
00:52When you're making decisions, you're not just choosing a model, you're also choosing you're making an inferences decision.
00:58How do I access the model?
00:59How much am I paying for it?
01:01What's the speed that I'm that I'm getting?
01:03What's the cash hit rate?
01:04We believe strongly that all of the kind of layers of the stack are interrelated.
01:08We're all making decisions across them.
01:10And so that's why we benchmark the full stack of AI.
01:15Kind of leading companies kind of reference artificial analysis and our results.
01:21I think but what we're most proud of here, when we take a look at these references from OpenAI, from Anthropic, from Google, and and Nvidia, they're not just referencing
01:34our raw intelligence benchmark numbers.
01:39It's not just a bar chart.
01:41We We're big fans of our intelligence index, but we're kind of most interested in talking through the tradeoffs that we're making between these technologies.
01:51And so that's why we're proud that all of these are kind of two by two charts showing tradeoffs between uh, choices developers need to make.
02:02Um, sometimes not when they're when they're getting things working, uh, but really when they're rolling it out in production.
02:11We're also quite widely referenced, um, by kind of traditional and, uh, and new media.
02:18We are based in San Francisco.
02:21We've got a team of 50, um, and we are hiring.
02:25Um, so we're hiring across uh, four deployed engineers, um, AI research staff as part of our member of technical staff team, and ML engineers.
02:36Um, and so, you know, catch me after or please reach out if you if you might be interested, um, in joining the the world of AI benchmarking.
02:47I think before we understand cost, we need to understand the kind of landscape of AI and the metrics to to understand and to optimize.
02:57And so, starting with understanding the AI landscape, I'll first introduce our intelligence index.
03:05It's we view it as the best single number to understand the general general intelligence, uh, of language models.
03:14It's a synthesis metric of 10 different evaluations covering different capability areas of AI.
03:21That includes knowledge work.
03:24It includes scientific reasoning.
03:27Uh, it includes coding and terminal use, uh, and it includes other capabilities including kind of knowledge, uh, hallucination and other relevant capabilities and considerations, um, regarding intelligence that people use when when choosing between language models.
03:47What we say is for the current use cases being built and the next use cases.
03:51It's not necessarily to measure and
03:53[clears throat]
03:54and benchmark progress towards AGI.
03:56It's really to help developers make decisions for the next use cases that they're building uh, using AI.
04:05And so of course Opus 5.5 from Anthropic is currently leading the index.
04:11I think to many in this room that won't be a big surprise.
04:14But also to discuss today is that there's as you can see here many other models.
04:18We have over 500 models on the artificial analysis website.
04:22Um and many kind of current and and and presented on this chart.
04:27And I think what we'll be talking through today is a bit that that kind of takeaway that there's also reason to use models other than the most intelligent model in the world.
04:39We also publish individual eval results.
04:41So we released AA analyst agent which is a finance and accounting focused focused benchmark.
04:49Thought I'd share that this being New York.
04:52And we share the individual results for those wanting to go to the next layer of specificity when thinking about their model selection and and selection of other technologies because
05:04it's not always the case that models exactly follow the generalist view of the intelligence index.
05:12And so here we kind of highlight that the Gemini series of models is actually very good at navigating complex complex spreadsheets.
05:19Of course Anthropic series of models is is of course as well.
05:25To step back a bit looking at our intelligence index and the leading model over time from the top 12 or so labs here.
05:34Um firstly I think big takeaway it's kind of hectic.
05:37There's many dots and I think probably that resonates for a lot of us in the room who have kind of felt this over the past few years where there've been
05:46huge number of model releases and the number of model releases in is increasing.
05:51And in a sense kind of the the pace of progress in our view is also accelerating and I think that is shown in this chart here.
06:00And so I think the takeaway here is that AA is moving quickly, progress is moving quickly, and so kind of we might feel exhausted trying to keep up with it, but
06:12um it makes sense to to it to invest that time and energy because if you want to stay on on top of AI and what AI's capable of, um then staying on top of releases
06:22um is a a big part of that.
06:25The next view won't deep dive, this could be a whole talk, but open weights models continue to keep within 3 to 9 months of the leading proprietary models.
06:35And so open weights models remain an option.
06:38Uh and so I think that's that's the takeaway for the talk today and and it flows through to a lot of um kind of cost implications uh which sometimes open weight models can have favorable
06:49uh characteristics.
06:53Next, let's dive into cost and how to think about cost.
06:56And I want to discuss think talking about cost from the perspective of being kind of mid-late 2026 because I think if we were doing the if I was doing this talk even six or nine months ago, I think it would be a different talk.
07:10AI has moved a long way.
07:12And so we'll go into the drivers, but to kind of start with a little bit kind of the answer of why we're so conscious of cost is because these models
07:23are costing many dollars um to complete individual tasks.
07:29That wasn't the case uh 12 months ago.
07:32Um people were spending more than ever on AI and can spend more than ever on a single task um on AI.
07:39And so here we present Opus and Fable costing above $5 to complete a task.
07:46But on the other side, um there's also models that cost 5 cents and less to complete a task.
07:53And so what we're seeing is there's orders of magnitude, multiple orders of magnitude uh difference in cost uh between the between the models.
08:02And that means that the cost decision and the model decision uh matters.
08:08And what I would say is that yes, you might say oh the uh the models on the left here are less intelligent and that's totally true.
08:16But they're still very good.
08:18Uh uh Luna here, GBD5 um sorry, GBD6 Luna is 7 cents uh per task um and it's much better than than GBD4 was um originally, right?
08:30Which kind of and and and way better than GBD3.5 Turbo, which really kind of like opened up um th- this wave of kind of language models.
08:38And so, I would say that there's two orders of magnitude difference between the leading proprietary models and much cheaper, but still very good, very capable models.
08:53A next way to kind of consider looking at cost is uh from a time perspective.
08:59And what we've done is we've bucketed models by their intelligence and looked at when that intelligence tier was first reached and then how did the kind of cost of trying to access the intelligence with the cheapest model develop over time.
09:15And what we see here is that within 1 year, accessing that level of intelligence, a 30 on the intelligence index or a 50 on the intelligence index, essentially reduces
09:26by an order of magnitude, kind of 10x, within around a year.
09:30So, picking the the light purple here, um once we reached a 30, cost a bit over a dollar.
09:37A year later, it fell to around 10 cents.
09:42And then now we're well below uh the 10 cent.
09:45We're in the kind of um 1 cent to 10 cents territory.
09:48And so, I think it's there's also a time dimension here whereby we should be um you know, making choices for now, but also conscious that that is going to change very quickly
10:00in terms of the intelligence that we can achieve and and that maybe $1 million um spent for a big gigantic can you know, is likely we can access that level of intelligence at 100k
10:11and then in 2 years potentially around 10k.
10:18Now, here we have a build up of the cost per task.
10:21This is a lot more complex than where we were 12 months and definitely 24 months ago.
10:28I think 12 months ago we're talking about reasoning models.
10:30We're saying that we have to pay for reasoning tokens and um you know, the the labs were just talking about the output token price um and so we needed to do an analysis
10:41um as a as a community to really understand the cost because we have to count reasoning tokens.
10:45Now with agents it's become way more complex.
10:48Way more complex.
10:49So, to understand the build up of a cost per task, here we use our AA briefcase agent technology workbench mark.
10:56It looks like agents in an environment of thousands of files completing knowledge work tasks like creating reports, creating PowerPoints, creating marketing videos.
11:05And what we can see here is on average Opus is costing $21 to complete a task.
11:12And it does a great job.
11:13No arguments there.
11:15Focusing on cost though, what we can see here is that uh answer tokens, you know, cost around 16 cents.
11:22Reasoning tokens, um oh sorry, uh it's like 2 cents.
11:27[laughter]
11:28Um and then we have then we have reasoning tokens a bit more, but then really where the you know, building up to the single language model call, the first kind of column that hits the ground and I'll talk about the second there.
11:39Um so each call is 13 cents, right?
11:42But with agentic trajectories we're having to think about the number of steps, of course.
11:47And so it took 165 steps to complete the task.
11:51And so that's where that's a new dimension that we need to think about when it comes to agents.
11:56And then thinking to kind of what goes into a single language model call, you can see there's answer tokens, reasoning tokens, input tokens and then with input tokens you need to think about how many of those input tokens uh kind of cash hits
12:09um and so get that roughly 10x uh discount on an input token price.
12:15And what we can see with the build up is the biggest driver of the kind of input into what a single language model calls is the number of input tokens
12:26and then secondly the amount of or proportion of those input tokens that are cache hits.
12:32And then the third biggest driver after that is of course or it's actually the biggest but it is the number of turns per task.
12:38And so we were talking a year ago around the number of output tokens, how much reasoning, the reasoning verbosity.
12:45I think today we really need to be talking about input tokens, uh cache hit percentage, um and then the number of turns per task is the key metrics that we should be thinking about when thinking about the cost
12:57um of running of running agents.
12:59And I think just to really emphasize the number of cache hits, if we don't have that then and and we don't have any cache hits, right?
13:09Cache hit percentage is zero, caching isn't working, maybe we've got our own setup with VLMs, we haven't enabled prompt caching, um then instead of $21 per task, this costs over $200.
13:21So that's just to really emphasize how much caching matters here when we're talking about agents and also when talking about inference providers, how much you really need to to care about that.
13:30We started recently publishing benchmarks around cache hit percentage on our website.
13:34And next consideration is um you know different reasoning efforts.
13:38And so here we have um kind of column um column charts showing the different reasoning efforts um for Opus 5.5 and the implications on these metrics.
13:51And so I think pretty much the key takeaway here is when we're talking about reasoning efforts, we're not just talking about um uh kind of reasoning the number of reasoning tokens
14:02um for an agent single agentic call, it does impact the number of turns.
14:07Um And of course then you have reasoning tokens and then those reasoning tokens feed throughout the agentic trajectory into kind of much more input tokens um because of the way that like message histories work
14:21uh when it comes to comes to using agents.
14:24And so even with it when using the same model, very important to think about the reasoning effort because here we can see that at low Opus 5 can cost 5.5
14:34can cost a dollar per task and at max it can cost $21.
14:39[clears throat]
14:41And here's kind of um a breakdown going from Opus 5.5 to GLM 5.3 Flash, a much cheaper but still extremely good, very capable model um that's that's open weights.
14:55And I think when we look at this when we look at this breakdown, this is really just to show that one, there's big differences between as we saw between the leading models and the and much cheaper models and two is that all of the different
15:08drivers that I was talking about play a role in driving cost differences um between models.
15:16And then thirdly, of course, that input token prices and your cache hit percentage um are the biggest driver um of cost when doing long uh horizon agent tasks.
15:30So we've talked about the cost and I think there's no points for knowing that GLM 5.3 Flash is cheaper than Opus 5.5. I think a lot of people in this room know that Lunar is cheaper than than Astra.
15:44Great.
15:45So the challenge for us as builders is making the trade-off decision.
15:49That's the hard thing.
15:51How do we understand our trade-off curves for our tasks?
15:56And I think a starting point that we think is useful to use is the generalist um view of intelligence versus cost per task and starting with the Pareto.
16:07So starting with okay, my task um where does it sit across this uh this intelligence Pareto in terms of the model that's needed.
16:20If I'm able to spend more or wanting leading intelligence, I want to look to the right of the Pareto.
16:26If I want cheaper, then I should let look to the left of the Pareto.
16:30The Pareto being this this line depicted in this in this chart here.
16:35And when choosing and wanting a cheaper model, let's take a look at the at the Pareto and seeing, "Okay, I'm willing to my budget is less and so what's the model that at that point offers the best intelligence
16:50for that that dollar?"
16:51Or from the other perspective of I want this level of intelligence, what's the cheapest way I can get it?
16:57How to think about navigating that decision and asking yourself those questions?
17:02A way that we think's helpful is to bucket tasks into two categories.
17:07So first is unlimited intelligence tasks, whereby more intelligence can always deliver a better result.
17:14An example being, "Analyze drivers of the costs in my business and how do I optimize them?"
17:20Kind of smarter will always deliver a better result there.
17:24Next, the second bucket is limited intelligence tasks and here is when more intelligence won't give you a better result.
17:32So the question, "How much did I spend on subscriptions last month?"
17:35I think a cheaper like Luna will answer that question as long as it's set up with the right MCPs.
17:41Currently, it'll tell me $80.
17:43Using Astra will also tell me $80.
17:45$80.
17:46I'm not getting any benefit from more intelligence.
17:50Um and so then I should focus on what's the cheapest way what's the level of intelligence needed, the ceiling, and what's the cheapest way that I can get that level of intelligence.
18:01So that's limited intelligence.
18:02Unlimited intelligence, the question is, "What's my budget?
18:05How much am I willing to pay for increased intelligence?"
18:12In terms of how we suggest using artificial intelligence, we suggest to kind of and how to navigate these decisions.
18:19I think it's understand the landscape, the metrics, then compare models based on priorities, shortlist models, and then for kind of smaller prototypy workloads, use the shortlist, test on your own,
18:34uh you know, in your own context and your own agents.
18:37For more advice users, particularly when rolling out in production, we recommend creating your own benchmark to take that shortlist to five models and try and make um the decision optimize for the decision
18:50um because of the kind of magnitudes of cost that we're dealing with now.
18:53Um we recently released a product called Optima um that allows users to come to Artificial Analysis, upload their context, and we will um together with you using a lot of our um knowledge that we've distilled into this kind of Optima agent,
19:08we generate you a benchmark um and give you your own trade-off curves between your view of intelligence for your task um and the cost to help you help you make these trade-offs.
19:22So, that's the cost of intelligence.
19:24Um thanks very much for for listening.
19:26Come check us out our website at artificialanalysis.ai, and um as mentioned at the start, we're uh we're hiring quite broadly across roles.
19:34So, thank you very much.
19:36[applause]