0:00

Today on the AI Daily Brief, how

0:02

should we be pricing agents? Before

0:04

that in the headlines, a new EO

0:06

on AI education. The AI Daily

0:08

Brief is a daily podcast and video about the

0:10

most important news and discussions in AI. To

0:12

join the conversation, follow the Discord link in our show

0:14

notes. Welcome

0:24

back to the AI Daily Brief headlines edition, all

0:26

the daily AI news you need in around five

0:28

minutes. We kick off today with something

0:30

of an unexpected one. President Trump

0:32

has signed an executive order to boost

0:34

AI education and workforce training. A White

0:36

House fact sheet said, America's

0:44

youth needs opportunities to cultivate the skills

0:46

and understanding necessary to use and create

0:48

the next generation of AI technology. Early

0:50

training in AI will demystify this technology

0:52

and prepare America's students to be confident

0:54

participants in the AI assisted workforce. propelling

0:57

our nation to new heights of scientific and economic

0:59

achievement. Preparing our students to be

1:01

leaders in AI technology also requires investing in

1:03

our educators, providing them with the tools and

1:05

knowledge to both train students about AI and

1:07

utilize the technology in the classroom. Now,

1:10

in terms of specifics, the executive

1:12

order establishes a task force on AI

1:14

education. Led by the director of

1:16

the Office of Science and Technology Policy,

1:18

the committee includes the secretaries for agriculture,

1:20

labor, energy, and education alongside numerous White

1:22

House advisors. The task force was

1:24

instructed to establish a prize called

1:26

the Presidential Artificial Intelligence Challenge, which

1:29

will, quote, encourage and highlight student

1:31

and educator achievements in AI, promote wide

1:33

geographic adoption of technology advancement, and

1:35

foster collaboration to address national challenges with

1:37

AI solutions. Alongside the

1:39

task force is instructed to develop

1:41

online resources for K -12 students in

1:43

foundational AI literacy via public -private partnerships. The

1:46

order also contemplates the use of various federal grants

1:48

to advance the goal of AI education. Teacher

1:51

training is addressed with the Secretary of Education

1:53

instructed to prioritize the use of AI

1:55

in discretionary grant programs for teacher training, and

1:57

the Department of Labor is instructed to

1:59

direct funding towards AI apprenticeships and certification

2:01

programs designed to provide high school

2:03

students with workforce credentials for AI

2:05

careers. As with most EOs,

2:07

the details are left to government officials to

2:09

execute. However, White House Staff Secretary Will Sharf

2:11

gave the elevator pitch, stating, The basic idea

2:13

of this executive order is to ensure that

2:16

we properly train the workforce of the future

2:18

by ensuring that school children, young Americans, are

2:20

adequately trained in AI tools. On signing

2:22

the

2:25

order, President

2:27

Trump

2:29

said, Now,

2:32

part of the impetus for moving on AI

2:34

education is likely pressure from China. Numerous

2:37

provinces in China have introduced mandatory AI

2:39

education at all gray levels, K

2:41

through 12. No doubt the rest

2:43

of China will quickly follow after the central

2:45

government recently placed AI at the center

2:47

of a comprehensive education overhaul. The Education Ministry

2:49

said that promoting AI will help, quote,

2:51

cultivate the basic abilities of teachers and students

2:53

and shape the core competitiveness of innovative

2:55

talents. Moving over back

2:58

into business world, open AI

3:00

is now forecasting $125 billion

3:02

in revenue by 2029. According

3:04

to investor documents viewed by the information,

3:06

OpenAI sees this revenue boost coming

3:08

as revenue from agents and other products

3:10

overtake those from chat GPT. These

3:13

numbers would put OpenAI firmly in the top

3:15

tier of tech companies. NVIDIA

3:17

recorded $130 billion in revenue last

3:19

year, while META recorded $160 billion.

3:21

OpenAI's 2029 forecast isn't anywhere near

3:23

terminal growth. With the company predicting

3:26

they'll be able to achieve 40 %

3:28

growth to hit $174 billion in revenue

3:30

for 2030. In one sense,

3:32

the company is extrapolating from very

3:34

little information. OpenAI achieved 3 .7

3:36

billion in revenue last year, which was

3:38

almost 300 % year -on -year growth. Predicting

3:40

what the next 30 -fold increase in revenue looks

3:43

like is speculative to say the least. But the

3:45

components of this stratospheric growth say a lot

3:47

about where the company sees the AI industry going

3:49

in the medium term. The first

3:51

point that comes from OpenAI's numbers is that the level

3:53

of hypergrowth is actually a necessity for the company. They

3:56

expect to burn 46 billion in cash

3:58

over the coming four years, and that

4:00

2029 figure of 125 billion is actually what

4:02

they need to hit profitability according to

4:04

their own estimates. Part of how

4:06

they get there is through cost moderating. The

4:08

company is predicting that inference costs will come

4:10

way down over the rest of this decade.

4:12

Last year, gross profit as a percentage of

4:14

revenue was only 40%. OpenAI is modeling

4:16

that ratio increasing to 70 %

4:18

by 2029. Interestingly, that's

4:20

still lower than the average gross margin

4:23

of cloud software companies, which is currently around

4:25

74%. OpenAI expects another revenue

4:27

boost to come from new product lines. At

4:29

the moment, revenue generation is simply

4:32

two buckets, consumer chatGPT subscriptions and

4:34

API usage from developers. The

4:36

company sees packaged agents like Operator, though,

4:38

being a big seller as they get

4:40

more performant. They see agent revenue hitting

4:42

$29 billion up from $2 billion this year.

4:45

Simon Smith from Click Health wrote, Hidden

4:47

in plain sight with OpenAI's revenue projections is

4:49

the fact that it's launching agents separate from

4:51

ChatGPT. Now,

4:58

still the 2029 projections view chatGBT

5:00

subscriptions as a major component of

5:02

revenue, representing 50 billion in

5:04

sales, which seems to imply either paid

5:06

user numbers in the billions, or

5:08

the ability to charge premium prices and push

5:10

large -scale enterprise deals. The figure

5:12

is separate from API usage, which OpenAIC

5:14

is growing to 22 billion by 2029,

5:16

once again up from 2 billion today.

5:19

Underpinning all of the numbers is massive

5:21

expectations of growth in users by the end

5:23

of the decade. OpenAI expects

5:25

to hit 3 billion monthly active

5:27

users, 2 billion weekly active

5:29

users, and 900 million daily active

5:32

users by 2030. They're currently

5:34

at 500 million weekly active users,

5:36

up 60 % since December.

5:39

For some comparison, Gmail has somewhere in the

5:41

range of 1 .5 to 1 .8 billion

5:43

monthly active users, and Facebook claims 3

5:45

billion monthly active users. Speaking

5:48

of OpenAI, the company's ImageGen API

5:50

is now available. Developers

5:52

can now integrate OpenAI's image generation model

5:54

into their own products. The feature

5:56

has been one of OpenAI's most successful rollouts

5:58

ever, with the Ghibli trend powering 700 million

6:00

uses in its first week of availability.

6:02

Pricing is set to $40 per million

6:04

output tokens, with OpenAI saying that this

6:07

translates to around $0 .02 per low -quality

6:09

image generation, .07 on

6:11

medium settings, and .19 for high -quality

6:13

images. And while indie developers

6:15

are getting their first opportunity to make use

6:17

of the model, OpenAI says that the range of

6:19

corporate partnerships is already extensive. Adobe,

6:22

Airtable, Wix, Instacart, Godaddy, Canva, and Figma

6:24

are already using the mode or experimenting

6:26

with its integration into their apps. Lastly

6:29

today on OpenAI, and lastly in general in

6:31

the headlines, details have been emerging about

6:34

OpenAI's open source model throughout the week. The

6:36

company is targeting an early summer release and aims

6:38

to produce a reasoning model that tops benchmarks. TechCrunch

6:41

sources say that Aiden Clark, OpenAI's VP of

6:43

research, is leading the project, which is still

6:45

in its early stages. They added

6:47

that the plan is to release the model

6:49

with few limits on commercial use, contrasting with open

6:51

models from Google and Meta that have prohibitions

6:53

above a certain scale. The goal seems

6:55

to be to compete with DeepSeq and other open

6:57

-source models out of China, which have relatively

6:59

few restrictions on how they're used. Sources

7:02

say the model will be text -only, operating

7:04

on high -end consumer hardware, and

7:06

potentially have the ability to turn reasoning on and

7:08

off. Later in the week, further

7:10

reporting added that the model could include

7:12

unique architecture touches to achieve those lofty benchmarking

7:14

goals. OpenAI leadership are discussing plans

7:16

to allow the open model to connect to

7:18

the company's closed models in order to quote unquote

7:20

hand off more difficult queries. The feature

7:22

was reportedly suggested during one of the recent

7:25

developer forums the company is conducting to source outside

7:27

input on the model, and the idea is

7:29

apparently gaining traction within the company. While

7:31

accounts the model is still in the planning phases

7:33

and training hasn't begun, but there is a

7:35

lot of excitement about this one and we will have to keep

7:37

an eye on it. For now that is

7:39

going to do it for today's AI Daily Brief Headlines edition, next

7:42

up the main episode. Today's

7:44

episode is brought to you by Vanta. Vanta

7:47

is a trust management platform that

7:49

helps businesses automate security and compliance,

7:51

enabling them to demonstrate strong security

7:53

practices and scale. In today's business

7:55

landscape, businesses can't just claim security,

7:57

they have to prove it. Achieving

7:59

compliance with a framework like SOC2,

8:01

ISO 2701, HIPAA, GDPR, and

8:04

more is how businesses

8:06

can demonstrate strong security practices.

8:08

And we see how much this matters every time

8:10

we connect enterprises with agent services

8:13

providers at Superintelligent. Many of these

8:15

compliance frameworks are simply not negotiable

8:17

for enterprises. The problem is

8:19

that navigating security and compliance is

8:21

time -consuming and complicated. It can

8:23

take months of work and use up valuable time and

8:25

resources. Venta makes it easy and

8:27

faster by automating compliance across 35 -plus frameworks.

8:29

It gets you audit -ready in weeks instead

8:31

of months and saves you up to

8:33

85 % of associated costs. In fact,

8:35

a recent IDC white paper found

8:37

that Vanta customers achieved $535 ,000 per

8:39

year in benefits, and the platform pays

8:41

for itself in just three months. The

8:44

proof is in the numbers. More than 10

8:46

,000 global companies trust Vanta, including

8:48

Atlassian, Cora, and more. For a

8:50

limited time, listeners get $1 ,000

8:52

off at Vanta.com slash NLW.

8:55

That's V -A -N -T -A dot com

8:57

slash NLW for $1 ,000

8:59

off. Today's episode is

9:01

brought to you by KPMG. In

9:03

today's fiercely competitive market, unlocking AI's potential

9:05

could help give you a competitive

9:07

edge, foster growth, and drive new value.

9:10

But here's the key. You don't need an

9:12

AI strategy. You need to embed

9:14

AI into your overall business strategy to

9:16

truly power it up. KPMG

9:18

can show you how to integrate AI and

9:20

AI agents into your business strategy in

9:22

a way that truly works and is built

9:24

on trusted AI principles and platforms. Check

9:27

out real stories from KPMG

9:29

to hear how AI is

9:31

driving success with its clients,

9:33

at www .kpmg .us slash ai. Again,

9:36

that's www .kpmg .us

9:38

slash ai. Today's

9:42

episode is brought to you by Super

9:44

Intelligent and I am very excited today

9:46

to tell you about our consultant partner program.

9:48

The New Super Intelligent is a platform that

9:50

helps enterprises figure out which agents to adopt

9:52

and then with our marketplace go and

9:54

find the partners that can help them actually

9:56

build, buy, customize, and deploy those agents. At

9:59

the key of that experience is what we call

10:01

our agent readiness audits. We deploy a set

10:03

of voice agents which can interview people across

10:06

your team to uncover where agents are going

10:08

to be most effective in driving real business

10:10

value. From there we make a

10:12

set of recommendations which can turn into

10:14

RFPs on the marketplace or other sort of

10:16

change management activities that help get you

10:18

ready for the new agent -powered economy. We

10:21

are finding a ton of success right now

10:23

with consultants bringing the agent readiness audits

10:25

to their client as a way to help

10:27

them move down the funnel towards agent

10:29

deployments with the consultant playing the role of

10:31

helping their client hone in on the

10:33

right opportunities based on what we've recommended and

10:35

helping manage the partner selection process. Basically,

10:38

the audits are dramatically reducing the time to

10:40

discovery for our consulting partners and that's

10:42

something we're really excited to see. If

10:44

you run a firm and have clients who might

10:46

be a good fit for the agent readiness

10:48

audit reach out to agent at bsuper .ai with

10:50

consultant in the title and we'll get right back

10:52

to you with more on the consultant partner

10:55

program. Again, that's agent at

10:57

Bsuper .ai and put the word consultant

10:59

in the subject line. Welcome

11:01

back to the AI Daily Brief. Today,

11:03

we are talking about agent pricing. And

11:05

while this may seem like it's just insider

11:08

baseball or only relevant for startups that are

11:10

trying to figure things out, I actually think

11:12

that it's much more important than that. The

11:14

agent pricing conversation is not just

11:16

about how startups and tech companies

11:18

are thinking about things. but also

11:20

implicates how enterprises are imagining and

11:22

considering agents. Are agents software

11:25

to be procured in the same way

11:27

that SAS was before? Are they

11:29

digital employees to be hired? How does

11:31

the answer to those two questions impact how they

11:33

should be paying for them? It

11:35

is absolutely the case that agent

11:37

pricing will shape and have a deterministic

11:39

impact on agent and AI business models,

11:41

which will have impact on things like

11:43

company design, all of which is to

11:45

say this is actually an important conversation.

11:48

Now, the specific context that prompts

11:50

us to have this conversation today is

11:52

that Windsurf has sort of kicked off

11:54

a price war on AI coding assistants. Their

11:57

standard Pro Tier is now priced at

11:59

$15 per month with an allowance for

12:01

500 prompts before requiring a top -up. They're

12:04

also getting rid of their Flow Action

12:06

Credit System, which charges for tool calls

12:08

within an agentic workflow. The company says

12:10

this means you can only pay per user

12:12

prompt. Tech and Enterprise subscriptions also

12:14

saw a price drop to $30 per month and

12:16

$60 per month respectively. CEO

12:18

Rob Hu made it clear that this was a direct

12:20

shot at cursor, although he didn't mention

12:22

the rival company by name, his announcement post

12:25

stated, and

12:30

in

12:32

a very pointed comparison he added, and

12:44

not charging for tool calls is a big deal

12:46

as agentic coding gets more complex. Windsurf

12:48

is charging a flat rate while complex cursor

12:50

prompts can easily balloon to cost a dollar

12:52

or two each. CEO Rob

12:54

here is very much hustling. He

12:56

spent hours this week replying to potential customers commenting

12:59

on how they want to see the service

13:01

structured in price. Of course, this all

13:03

comes in the background of a rumored OpenAI acquisition.

13:05

The deal isn't finalized, but according to well

13:07

-sourced reporting, OpenAI is looking at a $3

13:10

billion price tag to buy the company. Alongside

13:12

the price drops, Windsurf announced users will

13:14

get another week of free and unlimited usage

13:16

of the newly released GPT -41 and O4

13:19

mini models. It is a

13:21

question, I think, about whether Windsurf remains

13:23

profitable at these price levels or if

13:25

they're subsidizing users to gain market share. In

13:27

their announcement post, the company said that their

13:29

infrastructure engineers have been able to optimize GPU

13:31

usage and that they are delivering on their

13:33

promise to, quote, pass savings back to our

13:36

end users. Obviously, I'm now

13:38

watching closely to see if Cursor fires back

13:40

to commence a full -scale pricing war in one

13:42

of the fastest growing verticals in tech history. So

13:45

as I said, that is the specific

13:47

genesis of the conversation today, but

13:49

this agent pricing conversation is happening more

13:51

broadly. Aaron Levy from recently

13:53

wrote, nailing AI agent pricing is a

13:55

super important topic for AI companies right

13:57

now. There are two dynamics at play

13:59

in AI. pricing land. The first dynamic is

14:01

that the models themselves are getting cheaper and

14:03

cheaper to run. But the other trend

14:05

is that the use cases for customers are requiring

14:07

more and more inference. We've seen

14:09

examples of deep research using up to 100x

14:11

the compute of a standard query before. AI

14:14

coding agents similarly can consume enormous

14:16

resources depending on the task. So

14:18

even as the inference gets cheaper per token, the

14:21

total inference goes up dramatically. A

14:23

normal thing to do with resource pricing is to

14:25

effectively shift the cost of the resource onto

14:27

the customer. It's a clean model, but there are

14:30

many areas where a key customer use case

14:32

may be technically possible but unaffordable today, even though

14:34

they will become affordable tomorrow. So,

14:36

do you wait to solve the problem when it's

14:38

economically practical to scale, or lean in now and

14:40

bet on the cost improving? The

14:42

answer probably would be different in any other

14:44

technology category in history. But the implications

14:46

of AI model efficiency improvements is that software

14:48

companies can afford to price AI in

14:51

a way that anticipates the cost curve over

14:53

time. This allows you to unlock

14:55

use cases today that may be otherwise less

14:57

economically attractive, but where you know they soon

14:59

will be. It's definitely a bet, but

15:01

one that increasingly seems like it will pay off. This

15:04

is all thanks to the constant AI breakthroughs coming

15:06

from the frontier labs as well as open weights

15:08

model providers. And this doesn't appear to

15:10

be slowing down anytime soon. So

15:12

Aaron here in this particular case is

15:14

not discussing heuristics or models or frameworks

15:16

for pricing, but instead just the changes

15:18

in cost of goods sold, and specifically

15:20

the countervailing forces driving it in two

15:22

directions. cheaper to run models,

15:24

but more inference -required use cases.

15:27

At the beginning of April, Mani Badina, the founder

15:29

of Paid, wrote a post on the growth

15:31

on Hingeblog called a new framework for AI

15:33

agent pricing. The analysis comes

15:36

after looking at dozens and dozens of

15:38

different AI agent startups and Mani

15:40

divided their pricing models into four different

15:42

quadrants. They are per workflow, where

15:44

you pay per completed workflow, per

15:46

agent outcome, where you pay per completed

15:48

objective, So both of these are

15:50

outcome -based, one based on successfully completing

15:52

a workflow and one based on successfully

15:54

completing an objective, however many workflows

15:56

it took. And then another category

15:58

is what Manny calls activity -based. So

16:01

that's, for example, a fixed monthly fee

16:03

per agent. This is sort of a

16:05

replacement model for FTEs, or then a

16:07

pay -for -usage, a per -action agent, or

16:09

a consumption model. As Manny points

16:11

out, these different models are good for different

16:13

types of companies. The FTE replacement

16:15

model, a price per agent, he

16:17

says is best for AI agents

16:20

handling broad responsibilities or entire job

16:22

functions with consistent predictable workloads. He

16:24

points out that the advantage is that you get

16:26

to draw from the headcount budget or labor budget,

16:28

which is at least 10x larger than the software

16:30

budget, but that in many cases this

16:32

leaves you subject to other companies that just

16:34

charge less. The consumption model or

16:37

price per agent action, he suggests

16:39

is best for agents performing varied,

16:41

discrete tasks with unpredictable frequency or

16:43

volume. the price per agent workflow,

16:45

aka the process automation model, he

16:47

says is best for agents that execute

16:49

multi -step processes with clear intermediate deliverables,

16:51

and the price per agent outcome, aka

16:54

the results -based model, he says is

16:56

best for AI applications with predictable performance

16:58

and clearly defined success metrics in markets that

17:00

already expect it. Now, I want

17:02

to go back to his FTE replacement comparison

17:04

to a post from last year from Y

17:06

Combinator called vertical AI agents could be 10x

17:08

bigger than SAS. This was actually on their

17:10

podcast. And basically the argument

17:12

here, is that if agents actually

17:14

are replacing big swaths of

17:17

human labor, then they're competing for

17:19

labor budgets, not software budgets, with

17:21

labor budgets being radically higher than

17:23

software budgets. The interesting tension

17:25

that we're already seeing over here at

17:28

Superintelligent as companies get introduced to

17:30

agents is whether they're going to be

17:32

priced in reference to the equivalent

17:34

human labor or in reference to the cost of goods

17:36

sold with some amount of margin. I think

17:38

the agent companies in general are trying to

17:40

keep the comparison to the comparative human labor.

17:42

and that makes sense because those budgets again are much

17:44

larger. The challenge is a couple pieces. First

17:47

of all, if you're trying to say, my

17:49

agent does the work of a junior developer

17:51

and a junior developer would cost $100 ,000, but

17:53

my agent's only going to cost $40 ,000, it's

17:56

almost for sure that someone who can provision that

17:58

agent for much cheaper is going to say, well,

18:00

screw that. It doesn't matter that the human

18:02

equivalent was $100 ,000. The thing only cost me

18:04

a couple thousand dollars to run, so for $5

18:06

,000, you can have it. Basically, there

18:08

is likely competitive pressure which pushes

18:10

prices down. The other thing is that

18:13

some of the use cases that agents will be

18:15

used for won't be able to be priced compared

18:17

to their human labor because a priori the cost

18:19

of human labor would have been so high previously

18:21

that no one would have ever considered doing the

18:23

thing. I have a very specific example

18:25

from our agent readiness audits. As part

18:27

of the audits we do voice agent interviews

18:29

which can be across a handful of leaders or

18:31

it could be across every single person inside

18:33

a department. Right now we're live with

18:35

a big pharma company with all 200

18:37

members of a specific department of theirs And

18:39

the problem if we had tried to

18:41

price that against what it would have cost

18:43

for McKinsey to interview all 200 people

18:45

in that department is that we know a

18:47

priori that the cost of McKinsey interviewing

18:50

200 people like that would be so

18:52

astronomical that they never would even

18:54

consider that. So we can't sit

18:56

there and say, hey, if you ask

18:58

McKinsey to do this, it would be

19:00

$300 ,000 and we only want to charge

19:02

you $100 ,000 because they'd say, yeah, but we

19:04

never would even consider paying that because it's so

19:06

out of range as to be irrelevant for our

19:08

actual planning. And so we are

19:10

left in our circumstance having to price it

19:12

closer to the reference point of the cost

19:14

of goods sold because what we're doing is

19:16

fundamentally opening up new opportunities that simply weren't

19:19

possible before agents. Now, one

19:21

really interesting countervailing note comes from

19:23

Signal. They wrote, was talking to

19:25

someone else deep in AI and we started

19:27

riffing on how agent pricing could spiral into

19:29

totally uncharted territory. Like if

19:31

someone builds a quote -unquote perfect agent for

19:33

legal diligence or biotech and is protected

19:35

by real modes, data evaluation frameworks, workflow

19:37

lock -in and trust, they could charge

19:39

above human rates and still be the

19:42

obvious choice. 24 -7, few

19:44

errors, infinite scale. Kind of

19:46

ridiculous that the floor for AI labor is maybe

19:48

heading toward zero, but the ceiling could possibly

19:50

go infinite at least in the near term. Now,

19:53

this is one of those things that really

19:55

requires a lot of imagination around what the

19:57

moats would be that would even justify this. But

19:59

I think a relevant takeaway, no matter

20:02

what, is that agents aren't just going

20:04

to be competing on cost. They're

20:06

permanent and perpetual availability. They're massive scalability.

20:08

These are things that make them better than the

20:11

human equivalent, not just a better choice because they're cheaper.

20:14

Again, to round up with the same example of the

20:16

voice agent interviews we use for the agent readiness audits,

20:18

if you wake up with insomnia at 1 .30 am,

20:21

you're welcome to do your interview then. You

20:23

don't have to schedule it with us, you can do it

20:25

at your own convenience. And that is

20:27

a massively better experience than if we

20:29

were doing traditional voice interviews with our analysts.

20:32

Anyway, all of this I think is interesting,

20:34

and not just in a theoretical way, but in a

20:36

way that will shape the future of the AI

20:38

and Agent space. For now, though, that's gonna do it

20:40

for today's AI Daily Brief. Appreciate you listening as

20:42

always, and until next time, peace!