# AI Token Efficiency: The New Enterprise Metric

**Podcast:** The AI Daily Brief (Formerly The AI Breakdown): Artificial Intelligence News and Analysis
**Published:** 2026-06-04

## Transcript

Today on the AI Daily Brief, how companies are becoming AI token efficient.
Before that in the headlines, Chat Shippity becomes the fastest app to ever reach a billion users.
The AI Daily Brief is a daily podcast and video about the most important news and discussions in AI.
All right, friends, quick announcements before we dive in.
First of all, thank you to today's sponsors, KPMG, Robots and Pencils, Assembly, and OutSystems.
To get an ad-free version of the show, go to patreon.com slash ai daily brief, or you can subscribe and Apple podcasts.
If you want to learn more about sponsoring the show, send us a note at sponsors at ai daily brief dot AI.
And one more quick thing.
If you are looking to get up to speed fast on AI, you might have heard that episode that I did with Nufar about a week ago on the four AI hires that executives need to make right now.
Nufar is now offering a four week executive AI sprint called executive catch up.
There are just a couple days left to register.
You can find out about it at aiexecutivecatchup.com, and there will also be a link in the show notes.
We kick off today with another big checkmark from my 2026 predictions.
Although, honestly, I have to say this was the most gimme of all those predictions.
ChatGPT has officially hit a billion monthly active users.
That is, according to new estimates from data analytics firm Sensor Tower, who looked at monthly active users in May.
Now, the milestone has been a long time coming, and there's actually been a fair bit of digital ink spilt over it.
Specifically, back in April, the Wall Street Journal made a very big deal of OpenAI's failure to hit this milestone as their end-of-year target for 2025.
That article also highlighted a failure to reach monthly revenue targets and was part of a fairly negative news cycle for OpenAI.
Ostensibly, the narrative was that ChatGPT had hit a growth plateau as Claude and Gemini gathered steam, but in reality, and as listeners of this show well knew even back then, for those paying close attention, the narrative already seemed out of date by the time it was published.
OpenAI did have a rough end of the year as Claude Code took the world by storm.
Those issues led to Sam Altman calling a Code Red in December, and Fiji Simo declaring the end of side quests in March.
By April when the article was published, however, OpenAI was already well into the middle of their resurgence.
Codex was seeing a spike in popularity, and the release of GPT-5-5 had for many folks, the first time in a long time that OpenAI's state-of-the-art model was in the Vibes lead compared to its Anthropic pair.
And you have to think that for most people, now that the milestone has been reached, the five-month delay on reaching the billion-user milestone wasn't all that big a deal.
ChatGPT is still by far the fastest app to reach a billion users, taking just three and a half years.
That's better than TikTok's five years and significantly faster than the eight years it took YouTube and Instagram.
Around 12% of the global population is now logging into ChatGPT every month, making OpenAI's flagship product dramatically different to everything else in the industry.
Now, SensorTower's report did capture how dramatic the rise of Claude has been.
The company has seen 640% user growth over the past year.
but that still puts them only at 56 million monthly active users.
In other words, they're still around just 5% of the consumer use of ChatGPT.
But that also shows just how valuable the business audience is, given that Anthropic are officially ahead of OpenAI in the revenue race right now.
It is worth noting that CensorTower also found that most users aren't making a hard switch from ChatGPT to Claude, despite Claude's rising numbers.
ChatGPT users who installed Claude in the first quarter ended up using ChatGPT 5% less.
That's not a nothing number, but it certainly suggests that people are adding Clot as a second chatbot rather than as a direct replacement.
Now, why do all these numbers matter?
Well, frankly, it's because they're going to be thrown around a lot in the horse race narrative when it comes to Wall Street IPOs.
But for anyone who's not invested in one to the exclusion of the other, it just shows how incredibly dynamic and fast growing both of these companies are.
Speaking of big AI milestones, bots and agents have overtaken humans in web traffic for the first time.
According to Cloudflare's data, bots now represent 57.5% of web traffic that flows through their service.
Now a big chunk of this is of course AI data scrapers, but the growth in web agents has also been dramatic over the past year.
This is creating some challenges.
The rise in bot-based browsing means a drop-off in website ad revenue, and there's also been a sharp rise in malicious automated traffic.
Cloudflare now classifies 37% as bad bots that ignore web-crawling rules in robots.txt.
And yet, this was an entirely inevitable outcome.
In an interview at South by Southwest back in March, Cloudflare CEO Matthew Prince predicted that bots would overtake human web traffic by next year.
He said, For a long time, the internet was about 20% bot traffic.
Google was the largest, but you had a whole bunch of other things, including hackers and spammers and all kinds of miscreants that were online.
With the rise of Gen.ai, it's just an insatiable need for data.
We're seeing a rise where we suspect that in 2027, the amount of bot traffic online will exceed the amount of human traffic that's online and will continue to grow after that.
In an X post on Wednesday, Prince wrote, Well, that happened faster than I predicted.
Thought it would be end of 2027, then early 2027, but agentic traffic growing so fast that bots have now passed human traffic online for the first time in the internet's history.
And of course, for the entrepreneurs out there, the great agent adaptation is on.
I'm pretty sure that we're going to need to deliver the AI Daily Brief via MCP and API fairly soon.
Next up in the headlines, Meta is taking on the Enterprise with the launch of a new business-focused agent.
Sort of.
We'll get into why Enterprise is maybe the wrong way to look at this.
Meta unveiled the new agent at the WhatsApp-focused conversations conference in London on Wednesday.
They said that the product will be built on top of existing business messaging services, allowing users to do things like automating appointment booking or closing sales.
In the future, Meta expects the agent to be able to conduct market research and manage calendars across an organization.
In a recorded message, Mark Zuckerberg said, As our models advance, your agent will take on more and eventually help you run your whole business.
The easy comparison for the headline writers out there.
is that it seems like Meta is attempting to deliver open claw for small businesses.
Meta's head of product, Naomi Gleit, told Reuters, This is definitely an enterprise play.
We actually want to take actions now.
We actually want it to be able to complete the payment, to process the booking, to place the order.
Meta said that the agent is already in testing in some regions, and they already have a million businesses using it.
The agent will initially be available for free, but Meta will shift to a paid subscription over the coming months.
Alongside the business agent within Meta's apps, the company will also launch a broader business agent platform.
The platform will allow businesses to build custom agents for other operations, and will include connectors for hundreds of non-meta platforms including Shopify and Zendex.
Leek commented that one of the big missing pieces in the AI landscape is a unified platform that caters to the smaller end.
The number one thing I hear she said, especially from small businesses, is I just want to go to one place that can do all the things.
And this is where I want to draw the distinction with the enterprise language.
This is definitely a product for business, a B2B product, But I think that the confusion is that when we say Meta is building an agent for the enterprise, it's natural to think that the implication is rolling out some agent across a 5,000-person company.
Analyst Patrick Moorhead summed up the feelings of many when he said, I'm so tired of Meta's we're getting into B2B, take us seriously this time trope.
Every fiber of their soul is consumer ads.
He then went on to list a whole slew of B2B and commercial fails, but I don't really think that what they're talking about is a big enterprise sort of play.
That's not really what Meta is talking about.
This is for five-person companies, and those companies are already using WhatsApp and Messenger as the core part of their business stack.
When Zuck announced it, he said, now a clothing shop in Birmingham or a bakery in Sao Paulo can offer the same always-on, highly personalized experience as a major brand.
Writes Five Points Capital, the biggest problem with AI right now is usability.
If you're a restaurant owner, you're too busy to learn how to set up AI agents, and you don't want some AI consultant coming in to charge you 20k for something you're not even sure will work.
Meta business agents will just work, like an iPhone.
That convenience and simplicity is what small business owners desperately want.
And I think that this is much closer to the right analysis.
During the event, Meta said that they currently have 200 million businesses already using WhatsApp around the globe, and have reached $2 billion in annual revenue for paid messaging services on the platform.
I actually think that this is one of the more unique and valuable roles that Meta could play, so I frankly am excited to see what they do with this.
For now though, that is going to do it for today's headlines.
Next up, the main episode.
One of the most important AI questions right now isn't who's using AI, it's who's using it well.
KPMG and the University of Texas at Austin just analyzed 1.4 million real workplace AI interactions and found something surprising.
The highest impact users aren't better prompt engineers, they treat AI like a reasoning partner.
They frame problems, guide thinking, iterate, and push for better answers.
And the good news?
These behaviors are teachable at scale.
If you're trying to move from AI access to real capability, KPMG's research on sophisticated AI collaboration is worth your time.
Learn more at kpmg.com slash US slash sophisticated.
That's kpmg.com slash US slash sophisticated.
This episode of the AI Daily Brief is brought to you by OutSystems, a leading agentic systems platform built for the enterprise.
Organizations all over the world are building, orchestrating, and governing agentic systems on the OutSystems platform and with good reason.
OutSystems' open and unified platform allows teams to architect, deliver, and scale governed agentic systems with agility.
Teams of any size and technical depth can use OutSystems to build, deploy, and manage AI apps and agents quickly and cost-effectively without compromising reliability and security.
With OutSystems, you can rapidly launch ideas from concept to completion.
It's the leading agentic systems platform that is unified, agile, and enterprise-proven, allowing you to accelerate growth, reduce operational friction, and deliver real enterprise impact with AI.
OutSystems.
Build your agentic future.
So coding agents are basically solved at this point.
They're incredible at writing code.
But here's the thing nobody talks about.
Coding is maybe a quarter of an engineer's actual day.
The rest is stand-ups, stakeholder updates, meeting prep, chasing context across six different tools.
And it's not just engineers.
Sales spends more time assembling proposals than selling.
Finance is manually chasing subscription requests.
Marketing finds out what shipped two weeks after it merged.
Zencoder just launched Zenflow Work.
It takes their orchestration engine, the same one already powering coding agents, and connects it to your daily tools.
Jira, Gmail, Google Docs, Linear, Calendar, Notion.
It runs goal-driven workflows that actually finish.
Your stand-up brief is written before you sit down.
Review cycle coming up?
It pulls six months of tickets and writes the prep doc.
Now you might be thinking, didn't OpenClaw try to do this?
It did, but it has come with a whole host of security and functional issues, which can take a huge amount of time to resolve.
Zencoder took a different approach.
SOC 2, Type 2 certified.
Curated integrations.
Tighter security perimeter.
Enterprise grade from day one.
Model agnostic and works from Slack or Telegram.
Try it at zenflow.free.
Today's episode is sponsored by Bolt.new.
Bolt.new is agentic engineering on multiplayer mode.
Designers, product managers, and engineers build in the same environment.
And the design system agent keeps every screen on brand.
No more Frankenstein UIs stitched from a dozen prompts.
Whether you're shipping internal tools, moving from prototype to production, or replacing a legacy admin panel, Bolt.new takes your team from concept to deployed app.
One personal recommendation, hit plan mode before you build.
I had a project I'd half described in three different prompts and plan mode made me actually think through it with Bolt.new before a single line got written.
It saved me from rebuilding the same screen probably about four times.
Build better apps faster.
Start with the link in the description.
Welcome back to the AI Daily Brief.
Today we're diving deeper on the big AI theme of the moment, which is token efficiency.
Now, you might have heard this term coming up a lot more recently.
Matthew Berman recently tweeted, everyone is talking about token efficiency now.
I made an argument on Twitter yesterday that every AI business is now and for the foreseeable future a token efficiency business.
In other words, every company that is selling services or products around AI is somehow and in some way going to be trying to help companies be better at allocating AI budgets effectively to get the most value from the raw capability that the AI represents.
Now, there are a ton of stories right now about advanced early AI adopter companies shifting their strategies as token consumption goes way up in the agent era.
Walmart, as we discussed this week, has started to cap usage of their internal AI tool because employees were using it too much.
Uber, as we discussed just yesterday, has set a $1,500 a month limit on spend on tools like CloudCode.
And the whole issue of token cost is starting to come home to roost for the big labs.
In their enterprise event on Tuesday, OpenAI's Sam Altman said that AI budgeting had recently become a, in his words, huge issue for some companies, even though cost was something that, quote, never came up earlier in the year.
Now, again, none of this is particularly surprising if you look at the underlying dynamics.
The move from assisted AI to deploying lots of agents to do things for us.
has meant a significant increase in the amount of AI being consumed, represented by the number of AI tokens being used.
However, the number of AI tokens being consumed are limited by the number of AI tokens that get produced, which is limited by a whole supply chain of things like power and inputs and components.
And unfortunately for all of us, we are in the very early days of the build-out of that infrastructure and are very likely to be, over the course of the next half decade at least, living in a situation of some sort of token shortage.
And what happens in a market economy?
When there's more demand for something than there is supply of something, the price goes up.
Or, which is manifesting in the case of the labs, as shifting people off of subsidized per seat-based plans and over onto API pricing, meaning that they are paying for all of the tokens they're consuming.
And because that consumption can be effectively unlimited, that's why you're seeing companies start to impose caps.
Now, part of the reason that the media is so interested in this story is speculation around how this could slow revenue growth for both OpenAI and Anthropic.
heading right into their IPOs, and then further, how a slowdown in revenue growth and perhaps an underperforming IPO could change the capital market's appetite to continue to put money into those companies, which could have downstream impacts on that AI build-out, which could make the problem worse, etc., etc.
However, for our purposes today, we're not interested in the market discourse side of the conversation.
What I want to focus on is how companies are actually adapting and getting more token efficient.
Now, part one of this is a simple recognition.
that the efficiency and cost of intelligence are just as important as the raw underneath intelligence when it comes to AI in practice.
Perplexity's CEO Aravind Srinivas recently argued on CNBC that the single metric that would determine the winner of the AI race was which company can provide the most token value per watt per user.
He continued, whoever is able to maximize this particular objective really well by balancing accuracy, latency, cost, privacy, and intelligence altogether, they're going to win.
That's what's going to win long term.
Again, when it comes to AI in practice, it's not just raw intelligence, but the efficiency with which that intelligence is delivered that's going to really matter.
Now we're starting to see efficiency considerations show up in other areas of the discourse like benchmarking as well.
Up until this year, pretty much the only things people cared about when it came to benchmarks was the highest overall number in raw intelligence.
That's what state of the art meant.
However, as we've moved into the agent paradigm, even the benchmarking companies themselves are spending a lot more time on the efficiency of intelligence as well.
For example, increasingly the most important chart from artificial analysis is not just their leaderboard score, but their intelligence versus output tokens used for quadrant chart.
This one is a little bit easier if you're looking at it, but for those of you who are listening, I'll try to describe it.
In the Y column, we have the raw score on the artificial analysis intelligence index.
That's the aggregate score across all of artificial analysis's tests that has at the moment Claude Opus 4.8 on max setting and 5.5 on extra high setting up at the very top scoring between 60 and 62.
On the x-axis is the output tokens used in all of the tests that represent the artificial analysis intelligence index, with fewer obviously being better.
The top left quadrant then represents a combination of highest scores and best token efficiency, and paints quite a different picture than just the intelligence index alone.
Specifically, while Claude Opus 4.8 is now slightly above GPT-5.5 in terms of its intelligence index score, Claude achieves that score while using about 80 or 90% more tokens, meaning it's significantly less token efficient, and actually placing both Opus 4.7 and 4.8 outside of the most attractive quadrant.
The release of Gemini 3.5 Flash also saw a lot of this discourse around it as well.
While the overall intelligence was much higher on Gemini 3.5 Flash than 3 Flash, the cost to run the tests was more than five times as much as 3 Flash, moving 3.5 from just at the edge of the most attractive quadrant to firmly outside of it.
All of this is finding its way into the popular discourse as well.
For example, YouTuber and AI entrepreneur Theo recently tweeted, Meanwhile, perception of token efficiency is also part of why Codex has become so much more popular among developers.
Biniam wrote, Codex has gotten noticeably better at token efficiency lately.
Same tasks that used to eat up a ton of tokens now feel way more reasonable.
Fundamental Analysis on X wrote, GPT-55 and Opus-48 sit around 1 point apart on the Intelligence Index.
60.2 versus 61.4.
Their token pricing is almost a match.
$5 input on both, $30 versus $25 output.
So why is there a 40% gap in the cost of running the full index?
And the answer, of course, as we just saw, is that the Opus models burned way more tokens to complete the index.
Fundy writes, That's the whole game now.
Per token pricing is the rate, and tokens to completion is the actual invoice.
A model can win on price per token and lose badly on price per task.
Because the reasoning trace, the restatement, the overthinking is the multiplier nobody printed on the spec sheet.
This is why the cheapest per token model is routinely the most expensive per outcome.
Researchers have a name for it called the overthinking task.
Smaller, cheaper models that ramble can cost more in total than a pricier model that's terse and converges fast.
The buyer-side implication is the part the market hasn't priced in yet.
A.
The flagship layer now competes on token efficiency, not just capability.
40% fewer tokens for the same score as a moat, and it doesn't show up in the pricing table.
Enterprises are learning that cheap model and cheap workflow are unrelated numbers.
Price per token was always a proxy, which means the real metric was always tokens times price times attempts to correct.
And if the new Microsoft models are any indication, this is very quickly going to cease to be a hidden consideration.
VC Tomas Tungus wrote, Microsoft put a new column on its latest model card, average token usage.
It will become a standard.
For example, he writes, MAI Code 1 Flash hits 71.6 on Sweebench Verified, using a third of the token's Claude Haiku 4.5 burns.
Benchmarks now ship on two axes, performance and the cost to get there.
Even the most valuable companies cannot afford state-of-the-art intelligence everywhere.
Model companies will compete on intelligence per dollar.
The app layer will compete one level up, on dollars per outcome, a closed ticket, a shipped PR, a resolved support case.
Every layer prices the way the customer thinks.
per result, not per token.
And so one of the ways that I think you're going to see adaptation is that the labs themselves are going to start to prioritize different things, not just raw intelligence, but token efficiency as well.
Certainly Microsoft thinks it has an opportunity to compete with their new frontier tuning approach.
In announcing the new models and their frontier tuning program, they gave the example of a collaboration with McKinsey, where when the model was tuned for McKinsey's tasks, The Microsoft model delivered the highest win rate, even outperforming GPT 5.5, while being 10 times lower in cost than GPT 5.5.
And it won't just be the big labs.
You're also going to see the agent labs and even app layer companies experiment with their own models, their own harnesses, and their own routing systems in order to get better token efficiency, which is exactly what I meant when I said that every AI business model is now, to some extent, a token efficiency play.
We saw this with Cursor's Composer 2.5, which completes coding tasks in the range of the state-of-the-art from both Claude and OpenAI, but with a radically higher efficiency.
Interestingly, we also just got something from legal AI firm Harvey along the same lines.
This week, Harvey tweeted, We partnered with Fireworks AI to train open-source models for legal.
Here's what we found.
One, hybrid legal agents can beat frontier models on quality and cost by routing selectively to a frontier advisor.
We tested a hybrid setup.
where GLM 5.1 served as the primary worker, routing tasks to Opus 4.7 as an advisor when needed.
GLM invoked Opus sparingly, just 0.83 times per task on average.
The hybrid setup beat Opus on both quality and cost.
They also found that post-training can push open models to frontier-level legal performance.
With a little bit of post-training on Kimi's K2.6 model, they were able to move Kimi ahead of Opus on their legal agent benchmark, and to do so for 11 times cheaper than Opus alone.
Writes Patrick Oyo, Now, luckily for enterprise AI buyers, the infrastructure required for this sort of routing and even post-training is very quickly becoming productized.
Software development company Factory just released a new product this week called Factory Router, which they say picks the right model for every task automatically.
They write, A higher token build does not mean more work is getting done.
One-line fix?
Dock update?
Too often, routine tasks get routed to the priciest path out of fear of losing performance.
This only burns budget for no additional gain.
You wouldn't have Messi play Goalie.
Every model has different strengths, whether it's reasoning, speed, cost, or context.
Factory Router automatically picks the right model for every task.
And to show that this works, Factory says that Router delivered the same performance as Opus 4.7 at 20-25% lower cost.
Perplexity also announced a product this week in this domain.
They're calling it Hybrid Agentic Inference, and basically it's an inference routing system that intelligently distributes AI tasks between resources from your local machine and cloud servers.
Perplexity demonstrated the system at the Computex conference on Monday using their Perplexity computer agent.
The demonstration used local models running on Intel Core Ultra 3 hardware, so basically a relatively high-end consumer device.
Now, the ability to run AI models on local hardware obviously isn't novel, but what Perplexity is saying is new is the system's ability to split up tasks.
Perplexity's orchestrator can break a task down into components and assign them to sub-agents using a variety of different AI models.
The system can then determine which sub-agents need to run on the more powerful cloud inference and which can be completed on local hardware.
The process is all fully automated and requires no decision-making from the user.
And Perplexity pointed out that hybrid inference is especially useful when it comes to private information.
They claim their orchestrator is able to identify sensitive data and ensure it doesn't leave your computer.
Basically, the orchestrator was presented as a way to balance intelligence, accuracy, privacy, and cost when running fully agentic workflows.
And so just in a single week, you have a group of different products all being launched to help solve the problem of token efficiency.
And if you want some evidence that there's demand for this, look no farther than the recently released stats from Ramp, where their number one trending software vendor was China's DeepSeek.
Ramp lead economist Eric Karazian writes, In probably the biggest sign that companies are looking for cheaper alternatives to OpenAI and Anthropic, some are willing to use cheaper Chinese models, sending U.S.
data back and forth from China-hosted servers.
ERA also pointed out that three open-source model service providers made the list this month.
Glean CEO Arvind Jain captured the overall shift in an essay called Your Token Spend is an AI Architecture Problem, Not Just a Model Problem.
He argues that the four architectural levers that determine token efficiency are context quality, i.e.
it being too difficult for either the models to retrieve the right context for the enterprise task at hand, or for them to be confused by too many different buckets of conflicting context, which can just burn tokens before you even get to the actual task at hand.
Arvind also talks about model routing, where as he puts it, the goal is not to use smaller models everywhere, but to use the right level of intelligence for the job.
A third vector of token efficiency, he argues, is continual learning.
basically building systems that allow experimentation phases to happen once rather than every time.
He writes, When someone does useful work or writes something worth reusing, we document it so we do not have to recreate it from scratch every time.
Enterprise AI systems should work the same way.
If it doesn't, the system keeps paying the same exploratory cost again and again.
A system that learns from prior execution can reduce redundant reasoning, skip failed paths, and converge faster on the right workflow.
The result isn't just higher quality, it's lower cost on repeated work.
Lastly, he talks about harness design, which has been another big topic this year.
But to sum up, as I argued yesterday, it's pretty clear at this point that the big theme of the second half of 2026 is going to be how to put all of the exciting things that were uncovered at the beginning of 2026 into practice in a way that's actually cost-efficient and effective.
If you are building something in AI serving the enterprise, my guess is that in some way, shape, or form, that's part of your job even if you haven't identified it as such.
For our part, we will continue to track best practices in how companies are adapting.
But for now, that is going to do it for today's AI Daily Brief.
Appreciate you listening or watching as always.
Until next time, peace.