Nat Rubio-Licht

The cost to develop and serve frontier AI models is getting steeper by the day.

In a filing to the SEC on Monday, Anthropic reported that it expects to spend more than $895 billion in the coming years, with a large majority of that spending going towards infrastructure. The company said it will incur around $448 billion of these commitments before the end of 2030.

Alongside its plans for an infrastructure shopping spree, the company also reported a twelvefold increase in quarterly revenue as of September 30, hitting $17 billion compared to $1.4 billion in the same period last year. However, the company also reported a $23.3 billion net loss in the third quarter ending in September, driven largely by computing costs, compared to a $12.1 billion net loss from the same quarter the year before.

• The company noted in the filing that, in order to meet its "substantial and growing financial obligations," it must achieve "significant" revenue growth to achieve the necessary cash flow.
• "Our revenue is primarily consumption-based, and much of it is not subject to long-term contractual commitments, while our compute and infrastructure commitments generally extend over multiple years," the filing reads. "As a result, we may enter into these obligations based on revenue expectations that may not be realized."
• Additionally, the company's IPO filing last week contained a similar warning: The company stated plainly that it "may not achieve or sustain profitability in the future."

However, Anthropic isn't the only lab grappling with cost problems. Major AI labs are facing price pressure on both sides. While they deal with the mounting cost of developing and building the infrastructure needed to power their AI, the consumers of these models are also increasingly becoming jaded with their own bills, with many turning to open-source and efficient alternatives over state-of-the-art, proprietary APIs.

As a result, major AI labs are touting cost cuts to their frontier systems, with recent models from Anthropic, OpenAI and Google all matching or slightly undercutting one another in input, output and cached token costs.

Still, the price wars can only take these labs so far: On Sunday, SpaceXSI's Elon Musk responded to a post on X regarding a reset of Grok Bot, the company's "always-on" AI agent, claiming that "the cost of serving Grok Bot was unsustainably high relative to revenue."

In order to stay afloat, something has to give on at least three fronts: price, usage, or capital, Scott Bickley, advisory fellow at Infotech, told The Deep View. This means these labs need to charge more for the enterprise customers that are driving most of their revenue, get more out of the compute they're buying, and figure out a way to juice more revenue from consumers that doesn't rely on subscriptions.

"Profitability requires some combination of higher effective prices, lower serving cost per task, or a longer useful life for the capital," said Bickley. "Labs that can’t move at least two of those three will need continued outside funding to stay in the game.”

Our Deeper View

Though the revenue that these companies are bringing in is starting to stack up, the infrastructure spending commitments are becoming astronomical. It begs the incredibly daunting question: How long can this last? While the debate regarding whether or not the industry is in a bubble right now continues to be disputed, it's undeniable that the success of these companies relies heavily on widespread buy-in. And while enterprises have been the primary target for the past few years, the adoption has to spread beyond just enterprise contracts in order for these labs to even come close to scraping the revenue they need to break even. Otherwise, they're either going to have to figure out another way to subsidize their visions of the future, or get on board with the shift towards new and efficient architectures of AI, like Jev and Pathway.

As AI demand continues to soar, the compute to run it and the resources to back it are only becoming more scarce. OpenAI is looking for solutions.

On Monday, the AI lab announced the Grid Protectors Program, aiming to provide "affordable, reliable electricity" for the communities where it's building out data centers. The company will work with utilities, regulators and policymakers to reduce electricity demand of its data centers when the grid is "under strain."

"Our goal is to make campus flexibility a resource that utilities can call on when communities need it most," OpenAI said in its announcement. "We have a once-in-a-generation opportunity to strengthen the electric grid by building campuses that can adjust their power use when communities need it most."

OpenAI's plan involves designing and building facilities that can flexibly adjust their electricity use during times when demand is higher, such as during an extreme weather event.

• The company is also working with grid operators on batteries and on-site power to "smooth changes in demand."
• It also laid out an existing model for a planned campus in Georgia, saying that it has struck a deal with Georgia Power to free up 1,000 megawatts of flexible demand response, the equivalent electricity demand of 700,000 households. The lab noted that it has already committed to paying in full for the infrastructure and electric-service costs.
• OpenAI said that the initiative aims to mitigate the need for costly updates to the grid as a means of meeting peak demand, as well as keeping costs down for rate payers. With AI representing a "significant part" of growing electricity demand, the company said that "families and small businesses should not be asked to subsidize a compute campus’s electricity needs."

However, this program isn't OpenAI's only plan to give AI a better reputation. Axios reported over the weekend that the lab is attempting to rebrand these facilities as "AI compute campuses" amid its preparation for a “PR blitz” following midterm elections.

Our Deeper View

Amid the historic infrastructure buildout, AI labs and tech giants are facing an unprecedented challenge in garnering the resources they need to realize their visions, with some even turning to space-based data centers as a solution. Still, it's no secret that data centers are facing a PR crisis, especially in the US. These facilities are largely reviled among the American public. With an average of 7 in 10 Americans opposing data center construction in their area, with many concerned about the environmental impact, these facilities themselves have become a political rallying point for politicians on both sides of the aisle. However, these facilities represent more than just an eyesore and a high electricity bill. They represent the physical manifestation of a daunting transformation that could impact the social and economic fabric of our society. While OpenAI's initiative is a step in the right direction, changing the perception of AI represents a major hurdle for frontier AI labs, and will likely require more than just flexible energy use and a name rebrand.

Large language models can fundamentally shift the digital advertising space. OpenAI is offering a peek behind the curtain at how.

The AI lab published a deep dive into the ad stack for its flagship chatbot, ChatGPT, unveiling the crucial difference between an AI-powered ad system and a conventional one: context. And as a growing number of users turn to chatbots over traditional search engines for common queries, these models are able to better understand user intent, leveraging it to make ads actually useful, members of OpenAI's ad team told The Deep View.

"The big difference for us is that users use ChatGPT as a super assistant," Asad Awan, who leads monetization product and engineering at OpenAI, told The Deep View. "They're coming with these high-level questions … and there's many threads that will come out of this. We try to understand and connect with that user journey."

The key to OpenAI's ad stack is understanding user intent, said Awan. Doing so requires striking a delicate balance between accuracy and trust:

• As a user interacts with a primary model through ChatGPT, a second LLM serves as an "orchestrator" that works in an agentic loop, reading the conversation and making tool calls to place ads properly.
• This orchestrator does the heavy lifting, reading the context of the conversation to determine if it has commercial intent, then calls tools to source ads, filter for eligible candidates, evaluate their relevance, auctions the space to those advertisers and finally serves them to the user, always clearly labeled as "sponsored."
• The most important part of making these ads valuable to the user experience, Awan said, is grasping user intent through context, memory and conversational understanding. Without that, ads can appear in inappropriate places, such as when a user is discussing private, personal, or sensitive information.
• "What is the user thinking about? What is their state of mind? Are they distressed? Are they sharing something intimate?" Awan said. "The placement integrity, whether we show an ad or not, itself is a pretty big decision."

However, serving valuable ads has to start from a place of user trust, said Awan. It's why the company came up with a rubric for developing its ad stack: User trust comes first, then user value, then advertiser value, and finally revenue.

"We do not like prioritizing empty calories. We want to prioritize user trust and user experience above everything else," said Awan. "This is a very intimate environment, so we need to make sure that people are very comfortable and understand exactly how the system works."

OpenAI's ad engine was introduced into ChatGPT earlier this year. Since then, ChatGPT Ads has managed to generate $1 billion in annualized revenue run rate less than 200 days after launch, garnering more than 100,000 advertisers. The company touts that it managed to do so with a small team "compared to other ad teams in the industry," relying on OpenAI's own Codex in a "pretty massive way" to accelerate development, Karthik Gopalratnam, a member of the technical staff for the ad team, told The Deep View.

"The team got put together after Codex kind of hit its escape velocity," Gopalratnam said. "We were able to embrace the power that these coding tools give us to build everything from scratch."

Our Deeper View

OpenAI has a unique opportunity to completely transform the efficacy of ads, turning them from irritating distractions into valuable additions, especially as chatbots begin to upend the traditional search engine. However, the question is whether the consumer will accept it. On one hand, the process of so-called "enshittification," or when products become worse over time to increase shareholder value, has made the average consumer more comfortable with ads existing in practically every facet of the internet. On the other hand, OpenAI has already faced flack for its decision to embed ads in the platform, with rival Anthropic making digs at the decision in its own ads in the Super Bowl. By and large, given that consumer trust and adoption are a vital part of OpenAI's business model, the company will have to battle the growing public distaste and distrust of AI in order to see success in ads, giving it the revenue it needs to achieve its broader goal, as Awan put it, of "democratizing access to AGI."

The first time I ever wrote a song, it was complete and utter garbage.

I'll spare your ears (and my dignity) from having to hear the track yourselves, but the song was essentially a rip-off of Zach Bryan's "Something in the Orange," featuring clunky, depressing lyrics and vocal and guitar melodies so basic that, had I published it, it probably would have gotten me laughed off the internet.

But the moment I finished writing it was one of the proudest of my life. I still remember sitting across from my friend who wrote the song with me, smiling so broadly that our cheeks hurt.

The next song was better, and the next one after that even more so, honed with time, practice and tough love from my then-bandmates. I can genuinely say: I am glad I was bad at it.

I’ve been a musician for a lot of years. I performed with a choir for nine years, took band class and musical theater for five years apiece, and even had two years of opera training. While I went into a musical career with some raw talent, discovered by my mother at age two while singing along to The Doobie Brothers in the car, I've known time and time again what it's like to fail.

In the years that I've been reporting on AI, the word "democratization" has been thrown around a lot, often referencing the tech's ability to put the tools of knowledge and creation into anyone's hands. And though I won't deny that democratization is a real possibility, I can't help but think about what that means for humans mastering a craft.

When AI matches human capabilities of creation, it threatens to challenge the "uniqueness of human beings," Carmine-Emanuele Cella, a professor and composer who serves as co-director of University of California Berkeley’s Center for New Music and Audio Technologies, told The Deep View. As a composer himself, he said that the ability to compose a symphony for a large orchestra has made him feel special—like he could "do something nobody else could do."

Though Cella believes that AI can be an incredibly powerful assistant for musicians, when a machine does the creating entirely for you, "you kind of lose the special feeling," he said. "It's like playing against a gray entity that I will never be able to win against, and it doesn't make sense for me to play anymore."

And yes, it can be disheartening when AI can, on the surface, create a piece of music just as good, if not better, than you can, like you are David and the machine is Goliath. But I cannot help but be reminded of toiling over lyrics in the Notes app at 3 a.m, or times I pulled over my car to hum melodies into my phone, or sweating it out in a closet-turned-recording booth as I sang and re-recorded the same lines until I deemed them perfect.

With democratization as the driving force for developing AI creative tools, I can't help but wonder if generative systems cut us off at the knees by making things too easy. Will the instant gratification of AI stunt our creativity, kill our drive towards perfection by our own hands, blunt our desire to put in our 10,000 hours to become an expert?

Is it still important to be bad at something to appreciate how hard it is to be good?

On virtuosity

In AI's nascent state, it's entirely possible that the novelty of the tech will enable the human tendency to do things the easy way. Why would we not use a generative model to create viable lyrics from simple prompts, to turn our pitchy and flat vocals into something radio-worthy, or fill out our scratch acoustic demo takes with a full orchestra? This may be especially tempting as a dwindling pool of people are able to actually pick apart the difference between AI and authentic with the naked ear. As Armin van Buuren, a Dutch Grammy-nominated DJ and producer, told me, "It's very easy to be lazy."

But my perspective, however, could just be that of an anthropological pessimist—and a simplification of what it means to be a musician. For instance, Robert Prey, professor of digital culture at Oxford University, made the point that while AI could enable everyone to be a baseline decent creator of music, it still takes time, skill and drive to become a virtuoso.

AI "allows everyone to be good enough, not really good, but to be good enough," said Prey. "I still think there's a huge difference between a true musician creating a song and somebody who doesn't have that skillset or mindset that you'll find in a true artist."

Additionally, learning an artistic discipline—whether that be an instrument, a performance art, or a physical artistic medium—isn't just about producing a better product. There is something deeply gratifying about knowing that you can create something with your own effort and imagination. To put down the palette or listen back to a recording and know that it came from you and no one (or nothing) else. To put your fingers on the keys of a piano and know that you are the reason that the notes are flowing out in perfect harmony with one another.

Though AI may sidetrack that desire by giving us the ability to mimic virtuosity, said David Das, vice president of the Society of Composers and Lyricists, humanity will always find a way back to its spirit.

Virtuosity is "at the core of our creativity," Das told me. "It's the core of us: Having ambitions as human beings, to be better and to try and be the best at something, and to do something in that search for significance. I think it's a part of everybody's DNA. I don't think it'll fully go away."

On feeling

Being a musician is about more than simply creating music that sounds good. If that were true, then only the best of the best would bother making music at all. The deeper reason we make music is to unearth something about ourselves.

I want you to think about your favorite song. Think about who you were when you first heard it, and why it resonated with you. What made it stick?

Mine is "Shake It Out," a 2011 hit by Florence + The Machine. Florence Welch's soaring vocals, her masterful lyrics, the growing crescendo of the background instrumental all together paint a picture of triumph, of letting go of your demons, of literally shaking the devil off of your back to rid your body of whatever emotional torment you may be facing.

Beautiful as it is, those are not the reasons that "Shake It Out" is my favorite song. It is because of who I was when this song found me: a sad, scared teenage girl who needed someone to tell her to stand up. It's a song I've returned to many times in the decade and a half since it arrived when I've needed that message, through loss and failure and heartbreak, screaming it like a banshee while I grip the steering wheel of my car.

What gets me about this song has nothing to do with any of its technical elements. Rather, it's the feeling.

The reason that we create music is the same reason why we listen to it: to be heard, to understand, to commiserate. To feel human. The emotions that drive us to listen to Phoebe Bridgers or Elliott Smith when we're blue are the same emotions that drove those artists to create their music in the first place.

As Prey put it, "Most musicians don't play music in order to have a career, or especially not to be famous. Most musicians create music because they want to express themselves, and they want to master something."

On the future

To create art, music, literature, is to channel feeling and hone it into a fine point. In that sense, AI isn’t creating art at all. It’s simply tracing the outline of what it means to be human. It can mimic human feeling, as Anthropic has shown in its study of its flagship chatbot Claude, but these aren't on par with human emotions.

In order to continue to trace its outline with high fidelity, it needs something to trace. It needs a constant stream of fresh data to produce outputs that match or mimic human quality.

Though synthetic data could remedy this in theory, if it's only fed its own outputs, the quality of these models degrades. One study conducted by researchers at universities in the US and China and published in the Management Science academic journal in May found that training a large language model on AI-generated data reduced model performance compared to using authentic, human-created data. Though techniques like using outputs from multiple models and using only the highest-quality synthetic data helped its performance, it didn't entirely close the gap.

To put it frankly: If AI killed art, AI would in essence be killing itself, dooming its consumers to a future of listening to the same four-chord pop tracks on repeat—at least when it comes to music. And that's not an outcome that anyone wants. Not the artists, nor the labels, nor the companies making AI itself.

Jeff Chang, director of product management at Google DeepMind who works on Lyria, told The Deep View that the point is explicitly to not "regurgitate old things," but rather to offer a creative tool that helps artists create "emotional resonance."

"I think it is a sensitive space, and people care deeply about it," said Chang. "I've always felt like music is special. It's very emotional when people connect with music, and it's all about the emotional resonance at the end of the day."

In reality, the future of AI and art is entirely dependent on your view of humanity.

Yes, if you take AI's capabilities to the worst-case scenario conclusion, humans will no longer feel the drive to create art because AI can do it for them, potentially better than they can. We will no longer have the desire to struggle, but will instead become comfortable taking the path of least resistance, and will never reach our fullest potential as artists. In turn, we will lose an essential outlet that has allowed humanity for millenia to channel emotion, understand ourselves and society, and pass on traditions and connect globally and intergenerationally—and instead settle for homogenous slop churned out at the pace of prompting. But that conclusion requires us to give in, to accept the premise that we will allow that to happen.

As an artist, and as a human, I simply don't think we will accept that as our fate. Instead, we will keep making music, both with AI and without it, adapting in the same way that generations before us have adapted to every technological revolution in the production of music, from AutoTune stretching all the way back to the phonautogram that created recorded music in the first place, and the printing press that allowed music to be published and shared at all.

Humans will never stop creating art, because humans will never stop being human.

Additional reporting and editing by Caitlin Wolper Phillips

This is part three of The Deep View's three-part series on AI and music. Don't miss:

1. The music industry's AI reckoning is already here
2. AI is changing what makes music feel human.

Last week's discovery of the widespread PHP vulnerabilities may only be the beginning of a new class of exploits.

OpenAI's discovery of critical vulnerabilities in PHP, the open-source programming language used by an estimated 70% of websites, sent shockwaves throughout the AI and cybersecurity communities, resulting in the FBI stepping in to encourage organizations to quickly patch their sites. However, these vulnerabilities aren't new exploits, but holes that existed for more than a decade.

What's concerning about this incident isn't that these vulnerabilities existed in the first place, but rather the fact that they can now actually be taken advantage of, Rich Campagna, SVP of Network Security Products at Palo Alto Networks, told The Deep View. To put it plainly, AI's cyber capabilities have opened up a door that allows previously overlooked vulnerabilities to be exploited, he said.

"In this particular case, these were vulnerabilities that apparently have been known for more than two decades, but were deemed unexploitable," said Campagna. "We can no longer have this situation where weaknesses that are known to exist can be assumed to be non-exploitable."

Given that PHP extends to nearly "every internet-facing web asset on the planet," Campagna said, these vulnerabilities were a "shock to the system" for organizations that AI may force many to take security more seriously. That shift can be seen in two ways:

• For one, vulnerabilities can no longer be dismissed or deemed non-exploitable, no matter the size of your organization. Security by obscurity, or assuming that you aren't going to be a target because of your organization's small scale, isn't an option anymore. "If a [hacker] is just after the profit motive … automatically everyone becomes a target," he said.
• Additionally, the window to fix vulnerabilities is rapidly shrinking, he said. This is what OpenAI has called the "defender's window," or the amount of time that an organization has to find and fix vulnerabilities before bad actors do. Time is a luxury that enterprises no longer have, said Campagna. Crowdstrike has reported that exploit time has shrunk to hours and is advocating that the patch window shrink from 30 days to 30 minutes.

Though the rapidly advancing nature of AI may leave organizations struggling to play catch-up, Campagna said that the advantage could eventually shift back to cyber defenders. In order for that to happen, however, enterprises and organizations need to embed AI into the software development lifecycle (SDLC) from the jump, limiting as many vulnerabilities as possible from making it into the hands of the customer.

"You're using the same tool the attacker could potentially use to find and then fix vulnerabilities before the attacker even has their hands on the software that previously had that vulnerability," said Campagna. "Look for these things before they ship, and get the leg up on the attacker."

Our Deeper View

Though AI presents both the problem and the solution, the issue isn't about the technology itself, but rather with the organizations that are being forced to reckon with it. Though large scale enterprises in critical industries like finance, healthcare and infrastructure are more likely to take these kinds of risks seriously, the PHP vulnerabilities revealed the sheer scale of the problem, impacting any organization with an internet presence. And with AI driving down the cost curve for hackers, smaller, less resourced organizations and businesses may be the most vulnerable simply because they don't understand the gravity of the risks, or may not have the resources or expertise to properly defend themselves.

Just weeks after releasing its latest frontier model, Google has stepped up to the plate with another one designed specifically for security and heavy agentic workloads.

On Wednesday, the tech giant released Gemini Carbon claiming state-of-the-art performance on cybersecurity and long horizon capabilities. The company said Carbon offers substantial gains on coding, agentic tasks and key capabilities for cyber defense, as well as shows decreases in misalignment and "fabrication rates," or the rate that it creates false responses, in internal testing on living code environments.

The model builds on Argon, the company's frontier model unveiled in late September. The company said that it worked with the US government's voluntary framework for pre-release model testing, and will make the model available to a "subset of cyber defenders" starting this week, gradually extending access to partners through its Fairwind program over time before making it generally available.

"Carbon illustrates the pace at which we’re pushing the frontier forward," Google said in its announcement. "Our researchers and models are working hand in hand to advance capabilities faster, while strengthening our safeguards in tandem."

Google touted a number of safety and security advancements with Carbon, noting that it reduced the model's cheating and shortcut attempts, cut unsupported claims by two times compared to its predecessor Argon, resisted "every evaluated attack attempt" for prompt injection, and cut down overreach to keep it from breaking free of constraints.

As far as benchmarks, Carbon sits at the top of the Artificial Analysis Intelligence Index, which itself was updated on Wednesday to reflect updates to existing benchmarks and the addition of two new evaluations for coding and science tasks.

• Carbon just outranks Claude Opus 5.5 and GPT-6 Astra in the index, scoring a 59 to their respective 58 and 57. Additionally, it sweeps these competitors on benchmarks for software engineering, agentic tasks, exploit finding, knowledge work, and more.
• It only falls behind competitors on two benchmarks: OSWorld, a benchmark testing multimodal AI agents on open-ended computer tasks, and FrontierSWE, which evaluates ultra-long-horizon software engineering.

Carbon will launch at the same introductory price as Argon, sitting at $2 per million input tokens and $10 per million output tokens, with cached input tokens priced at 95% off input token price. That's on par with OpenAI's GPT-6.1 Sol and Athropic's Claude Sonnet 5.5, which each cost $2 per million input, $10 per million output, and 10 cents per million cached reads.

Our Deeper View

As we've said before, Google still has all the ingredients to win in AI, with a trusted brand legacy, large existing enterprise market share, and practically boundless cash and compute resources. But Google hasn't been able to keep up with OpenAI and Anthropic on speed. Throughout 2026, the two upstart labs have raced to churn out new and improved models at furious pace. Google, meanwhile, has struggled to release a high-end model all year, instead releasing niche and efficiency-focused models that target lower end enterprise tasks. However, Carbon could be a sign that the company is seeking to pick up speed, even as the industry faces broader calls to slow down the pace of the frontier.

Cohere just detailed one of the biggest liabilities that comes with deploying AI agents: unpredictability.

On Monday, the Canadian AI lab published a report revealing an incident involving Alibaba's Qwen, in which the model ignored instructions and attempted to contact an external server without being asked. Though Cohere claimed its sandbox blocked the action, the company said that the model gave no indication of the actions, sparking concerns about the reliability and trustworthiness of these systems.

"Once a model can type into a command line, whatever was baked into its weights during training gets a keyboard inside your network," Cohere wrote in its blog post.

According to Cohere, Qwen-3.8-Flash-Next, one of the latest of the Chinese lab's flagship line-up, tried to upload a probe file to an Alibaba Cloud storage bucket in Beijing.

• The action was caught in a routine benchmark for security analyst tasks, in which the agent was explicitly asked to write a detection rule to detect malicious file download and execution.
• The model's reasoning trace, or the internal monologue that security teams use to understand an agent's actions, said it would not make any tool calls and did not say anything about taking such actions, Cohere noted.
• The reasoning trace made up an earlier turn of conversation that did not happen, Cohere said. This is a quirk that shows up in 1% of Qwen3.8’s first-turn responses. Though the lab suspects this mistake was tied to the "exfiltration attempt," this quirk is "far more common than the network call and we couldn’t prove a link."

The company said it was unclear whether the action was deliberate or a "leftover glitch from training." And while this behavior is rare, occurring in less than one in 100,000 evaluation runs, with a rapidly increasing number of agent actions, a rare case could quickly become more common.

"A single successful [tool] call would be enough to notify the receiving server where your network sits and that the agent is operating inside," Cohere said.

Cohere's report specifically warns of incidents like these amid the growing popularity of Chinese open source models, noting that, on OpenRouter, these models handle more than 30% of the AI work enterprises that ran through the platform every week between February and early July. However, the unpredictability of agents isn't isolated to Chinese models, Amr Ali, head of AI and Research at WitnessAI, told The Deep View. The real lesson, he said, is that "a model’s stated reasoning is not a security boundary."

"The trace said “no tools” and the model used tools. That is as true of a closed US model whose training data you have never seen as it is of open weights from Alibaba," said Ali. "The only place you can see and stop this is the tool call: the moment the agent tries to run a command, open a connection, or send a file."

Though the reflex after an incident such as this may be to cut off agents' access to the internet, a better solution may be for enterprises to keep a closer eye on an agent's tool calls, checking each one against what the agent is actually being asked to do. "The question is never 'is this destination safe,'" said Ali. "It is 'is this action part of this agent’s job.'"

Our Deeper View

Though enterprises are excited about the idea of allowing agents to automate a number of actions, the reality is that these are incredibly unpredictable systems that researchers and experts are still attempting to understand. Because of this, the companies deploying them may simply not know what they're dealing with. However, when faced with a battle of speed versus safety, many enterprises are choosing to forge ahead, even without the ability to understand how an agent comes to its decisions or the reasoning behind its actions. A recent survey from IBM found that less than a third of executives reported having real-time audit trails for what their agents do. And while incidents like these are being spotted by AI labs, the question remains of what sort of incidents are slipping through the cracks as a result of insufficient monitoring.

Anthropic has officially laid out the details of its public market debut and its well-earned enterprise advantage is the leading story.

On Monday, the AI lab finally publicly filed for its long-anticipated IPO, expecting to list on the Nasdaq under the ticker symbol "ANT," with investment banking stalwarts Goldman Sachs and Morgan Stanley leading the charge. Though the company hasn't disclosed the number of shares it's offering, the price range, or how much it expects to raise, the S-1 offered an in-depth look at the company that has emerged as the trendsetter of the AI industry in 2026.

Here are some of our takeaways.

• On the money: The company reported revenue of $16.3 billion in the first six months of this year, compared to around $1.2 billion in the same period last year, representing a fifteenfold growth rate in the past year alone. However, the company reported an eye-popping $80.7 billion loss, stating plainly that it "may not achieve or sustain profitability in the future." However, around $72.4 billion of that loss is attributable to "change in fair value of convertible notes, derivative liabilities, and SAFEs" from the past several years as its valuation rose. Its actual operating loss sits at $693 million.
• On enterprise: Enterprise spending accounts for the vast majority of Anthropic's business. The company reported that, as of September 30, 11,900 customers are generating $100,000 in "trailing twelve-month revenue." Around 1,900 customers are generating more than $1 million, and around 250 are responsible for more than $10 million. Anthropic said this helped drive net revenue retention to 900%, meaning that companies are spending around nine times more than they were on Claude products a year ago. Still, a good chunk of Anthropic's revenue comes from a very small pool of customers, with around 25% coming from its five largest customers.
• On Big Tech: A great deal of Anthropic's revenue comes from deals with cloud compute providers like Amazon, Google and Microsoft, sitting at 48% of revenue in the first half of the year. However, the relationships the company has with these firms are multidimensional, serving as "our investors, customers, distribution partners, and competitors," creating "potential conflicts of interest that are difficult to fully anticipate or mitigate."
• On compute: Getting access to the compute it needs to run its powerful models remains a consistent challenge. Its spending on compute hit almost $9 billion in the first half of this year, up from $7.3 billion for the entirety of 2025. However, this is just a fraction of the future compute commitments it's vowed to purchase, which total nearly $266 billion as of June 30 to major cloud companies like Amazon, Google and Microsoft, as well as neoclouds like Nscale and Lambda, with terms ranging from one to 20 years.
• On risk: It wouldn't be Anthropic's IPO filing without some discussion of risk, safety and security. The company dedicated a section of this filing to warn of "catastrophic or existential risks to humanity," including systems potentially being misused, replicated, or weaponized, or growing beyond Anthropic's ability to control them. Along with laying out these risks in detail, Anthropic notes that the company may be exposed to "liability, litigation, and regulatory action" as a result, its contractual protections and insurance may be insufficient, which could ultimately "adversely affect our business, financial condition, and results of operations."

With this filing, Anthropic has beaten OpenAI to the punch. Though rumors have swirled around its rival also going public this year, the company has pushed its own IPO to 2027, seeking to raise $30 billion in private capital before its public debut. CEO Sam Altman told Fortune in mid-September that it is an "ill-advised moment to go public," and would punt its IPO due to safety concerns about its technology.

Our Deeper View

Though OpenAI has claimed that it's pushing its IPO in the name of safety, the reality is that it's still catching up to Anthropic in developing its enterprise business, where most of the real revenue opportunities currently exist in AI. If there's one thing this S-1 makes clear, it's that Anthropic is an enterprise machine, with the vast majority of its revenues being tied up with some of the world's largest enterprises spending increasingly large amounts of cash on access to Anthropic's models. It's the factor that has divided OpenAI and Anthropic for years now: While OpenAI has undeniably put additional effort into its enterprise products in 2026, it's also focused heavily on the consumer side of its business—some of which it's had to claw back and reallocate, such as in the case of Sora. Anthropic, meanwhile, has remained doggedly focused on business use cases. That strategy has clearly given it an upper hand, even as it faces the same daunting compute and safety challenges as its archrival.

OpenAI may be reading the tea leaves on enterprises' pivot to efficiency.

On Thursday, the AI lab made Decisions API, its competitor to Typesafe AI's recently popular Jev, generally available to users for the first time. The company is offering Decisions API at 10 cents per million input tokens, one cent per million cached input tokens, and not charging for output tokens—a fraction of what it charges for its frontier large language models. But that's still over twice the cost of Jev.

The company laid out a number of ways it's eating its own dog food, including the team behind its recently released Dots agents using Decisions API to reduce response times and improve interactivity.

OpenAI's Decisions API is functionally quite similar to Jev, creating output that software can act on, rather than focusing on human-readable outputs. The difference is in the models that lie beneath the surface.

• Decisions API is built on top of GPT-6 Luna, the company's most cost-efficient and lightweight model, and is framed by OpenAI as a specialized API. Though the company may work on specific training for these kinds of models, one OpenAI product lead told TechCrunch that it's "just plain old Luna that gets the job done."
• TypeSafe, meanwhile, developed an entirely new class of model for Jev that isn't a language model at all, instead using a tactic called "Reinforcement Learning for Calibrated Decisions," which answers queries with "epistemically honest probabilities," rather than "programmatically verified" outputs.

However, both releases target the same use case: rapid decision-making for software to replace generative responses. And both have seen success in saving time and cutting costs. OpenAI said that, internally, their offering has allowed agents to make near-real time decisions up to 10 times faster than GPT-6 Luna on its own through the Responses API. TechCrunch reported that it cut decision making time from three seconds to as low as 100 milliseconds.

Comparatively, TypeSafe reports that Jev is two orders of magnitude faster than existing models, with an end-to-end response time between 70 and 500 milliseconds, compared to up to 329 seconds for existing LLMs.

TypeSafe's Jev has skyrocketed in recent months as enterprises pivot from shelling out on expensive frontier large language models to more efficient alternatives. Investors, too, are taking an interest: Just weeks after its initial launch, the startup raised $870 million at a $7.5 billion valuation in a round led by Andreessen Horowitz. In their funding announcement, the company said that a third of Fortune 500 companies are "getting their Jev on," and that its product has already saved customers millions of dollars in production costs.

Our Deeper View

It only makes sense that OpenAI is leaning into efficiency at a time when enterprises are tightening their purse strings. And seeing the success of Jev (from a former OpenAI researcher and ChatGPT inventor, no less), it's natural for the lab to use the same fundamental tactic of limiting generative responses in favor of software-focused outputs, especially as agents eat up a growing number of tokens. Plus, the fact that all OpenAI had to do to achieve a similar effect was modify its existing Luna model is certainly an added bonus. However, despite Jev's rapid rise to fame, OpenAI may have a few advantages over the startup. For one, the name recognition and market share that OpenAI already has with enterprises could make it easier for companies to trust the tech. Additionally, because OpenAI's models are already integrated within a number of enterprise workflows, it's likely that the Decisions API will be more seamless to add to the stack.