Cheaper Models and Restricted Access Are Weakening the Frontier AI IPO Story

The return of Anthropic’s Mythos and OpenAI’s limited GPT-5.6 preview point to the same shift: frontier model access is no longer simply a product-release decision by a lab. It is becoming a mediated process in which the U.S. government has a say over who gets access first, and possibly who does not get access at all.

Alex Kantrowitz described the immediate Anthropic news as a “major de-escalation” in the company’s confrontation with the Trump administration. Citing Semafor, he said the U.S. government had lifted its block on Anthropic’s Claude Mythos 5 model, allowing Anthropic to release it to more than 100 U.S. institutions, including major companies and government agencies. The same letter, he said, was silent on Fable 5, a weaker version of Mythos that had briefly been the most powerful AI model available to consumers, though people close to the talks said they were moving toward releasing Fable as well.

The unresolved question was whether anything substantive had changed in the model or its deployment. Kantrowitz said there were not yet details about “actual structural fixes” or whether the government-requested anti-jailbreaking safeguards had been added.

Ranjan Roy treated that uncertainty as the central issue. The difficult part, he said, was understanding “who has control of what and when” in the emerging regulatory landscape. He did not see a clear plan in the government’s approach to Anthropic. To him, the move looked like an expansion of the prior Project Glasswing structure: first a set of 20 to 30 companies, now more than 100 institutions, without an obvious explanation of what had fundamentally changed.

Kantrowitz connected Anthropic’s release to OpenAI’s announcement of GPT-5.6. OpenAI, he said, appears to have watched what happened with Mythos and Fable and chosen a narrower, government-aware rollout from the outset. In OpenAI’s own release language, the company said it was beginning with “a limited preview for a small group of trusted partners whose participation has been shared with the government before releasing more broadly.” OpenAI also said it did not believe “this kind of government access process should become the long-term default.”

That caveat mattered. Kantrowitz read the rollout as an admission that OpenAI does not want the Anthropic conflict repeated, even if it is complying for now. The immediate consequence is a market in which access to the most capable systems may depend on being inside the approved group.

For Kantrowitz, the risk is not only procedural. If companies or countries cannot rely on access to frontier AI from U.S. model providers, they may move more quickly toward open-source models, model routing, and non-U.S. alternatives. That could weaken the return on the enormous investments frontier labs are making. He stopped short of saying the whole business model is definitely in question, but said that given the scale of money being raised and the stakes attached to the frontier, “maybe it is.”

Roy separated the safety argument from the market-structure argument. He noted that Anthropic itself had raised alarms and invited government scrutiny. OpenAI now appears to be following a template Anthropic helped establish. He was not convinced that withholding a frontier model from a full-scale audience is necessarily bad; there is an argument, he said, that the most qualified and responsible companies should test and push the limits of a powerful model first.

Kantrowitz accepted that case in principle. He has argued before that if a system is powerful enough, caution may be warranted. But he emphasized the second-order effects: access restrictions can become a mechanism by which government and model companies pick winners and losers. The approved firms get capabilities their competitors do not. The labs may be able to charge more to that small group. And the broader economy shifts away from a market in which any customer can buy the best available tool.

Roy’s response was that the “government picking winners and losers” frame should not surprise anyone watching the administration’s broader approach. He cited other examples of the government weighing in on tech-industry outcomes, including who gets to buy TikTok and how Stargate rolled out. In his view, frontier AI is receiving the same top-down treatment. He did not call it good; he said it was “a very bad thing” for the global AI landscape, especially relative to China. But he was less shocked than the venture-capital and tech-commentary circles reacting to it.

Kantrowitz’s concern was that the U.S. could end up disadvantaging its own model companies. He cited Aaron Levie’s argument: if the U.S. remains at the frontier and heavily regulates access to intelligence, the U.S. can gain an economic and geopolitical edge by controlling who gets the most advanced systems. But if the U.S. delays releases while another player, specifically China, keeps moving and reaches comparable capability, then the delays advantage the other player’s models and tech stack.

That possibility intersects with a change already happening inside enterprise AI buying. Customers are no longer automatically routing everything to the largest closed model. They are experimenting with cheaper models, open-source systems, and routing layers. Kantrowitz suggested that if the “harness” around the model becomes the competitive advantage rather than the underlying model, then a restrictive policy could ice out frontier labs and push more of the value toward open source and China. A robust open-source ecosystem could be beneficial in many ways, he said, but it would also limit the ability to control safety outcomes. That is why, in the same environment, some people are talking about banning open-source models — a move Kantrowitz said he does not think would be good either.

Roy cautioned against collapsing all of that into one story. Much of the current move toward open-source or smaller models, he said, is driven by cost and task fit, not necessarily regulatory concern. The work being routed to open source today is often different from the work a company might want Mythos or Fable for. The frontier systems are being discussed in the context of finding security vulnerabilities or pushing scientific innovation. By contrast, many open-source deployments are about finding the cheapest adequate model for repeatable business tasks.

Still, both saw a real convergence: restricted access makes customers look elsewhere, and cheaper adequate models give them a reason to stay there.

OpenAI’s GPT-5.6 release was presented as a three-model family: Sol, Terra, and Luna. Sol is the flagship model, Terra is positioned as a balanced model for efficient high-volume work, and Luna as a fast, affordable model for everyday work.

The naming drew immediate skepticism. Kantrowitz called Terra and Luna “the worst branding mistake ever made,” because both names are associated with cryptocurrencies that collapsed. Roy initially agreed, noting that one had fallen by 99.999% and Terra by 98%, making them “poster children” for the crypto fallout. Kantrowitz said the problem was not just that they crashed; he characterized the episode as a scam in which people invested believing the assets were stable and lost heavily.

Roy later softened slightly after Kantrowitz explained the apparent celestial theme — OpenAI trying to create a naming system analogous to Anthropic’s Sonnet, Opus, and Fable. Roy moved from “net negative” to undecided on the names, though both still treated the crypto association as a serious branding problem.

The branding issue was secondary to the rollout. None of the three GPT-5.6 tiers was publicly available, according to Kantrowitz. Access was limited to a set of companies shared with the government. Roy found that notable because OpenAI’s own positioning implies different risk levels: Luna is “fast and affordable for everyday work,” while Sol is the flagship. If risk scaled cleanly with model size and cost, Roy said, one might expect Luna to be broadly released while Sol remained gated. Kantrowitz joked that “nimble fighters” can still be dangerous; Roy extended the joke into the substance of the concern, imagining lightweight systems “exploiting software left and right” and disrupting financial infrastructure.

On capability, Kantrowitz said OpenAI ran GPT-5.6 through Exploit Bench, a cybersecurity test, and that its performance was basically on par with Mythos. The notable distinction was efficiency. Mythos used around 300,000 tokens on the test, while GPT-5.6 used around 100,000. Kantrowitz described GPT-5.6 as at least two times more token efficient than the Mythos preview and said it was priced at about half the cost.

Roy said the attention to price itself is a shift. In earlier phases of the AI market, most people did not know or closely track input and output token costs. Now price is becoming a headline feature rather than an afterthought. That matters because the enterprise story around AI revenue had been built around customers using the newest, most expensive frontier systems. A model family that makes efficiency central may be good for buyers, but it complicates the story that frontier labs can keep growing primarily by selling maximum-capability models at premium prices.

The strongest version of the frontier-lab business story depended on expensive models being indispensable. Roy said the reported annual recurring revenue growth shown to the market from roughly September or October through February or March was built on customers “cranking on whatever the latest, most expensive frontier model was.” If customers realize they do not need the Ferrari because the Honda Accord is good enough, the IPO narrative takes a serious hit.

Kantrowitz framed the problem through competitive substitution. Up to now, buying from the leading closed labs had a “nobody gets fired for buying IBM” quality. The leading models were easy to plug in, obviously strong, and available from the companies everyone was watching. But if frontier access is restricted, customers have an added incentive to look elsewhere. If they look elsewhere and find cheaper or open-source models that are close enough, the closed frontier companies may face slowing growth at precisely the moment investors expect near-flawless execution.

Roy agreed that expectations have been built by strong execution. He said the major labs have executed “nearly flawlessly” over the last six to 12 months, which is part of why expectations are so high. But he introduced another reason a restricted release could appeal to labs: it may reduce distillation risk.

He described reports of “distillation swarms,” where competitors send millions of prompts to a frontier model in order to understand its behavior and recreate aspects of it in an open-source system. If broad release accelerates that process, then limiting access to the 100 most profitable enterprise and government customers could protect the frontier model, preserve premium pricing, and reduce the chance that competitors learn too much too quickly.

Kantrowitz said his “hot take” is that Anthropic’s safeguards around Fable were not only about preventing misuse. He believes they were also about preventing competitors from distilling the model “in places that provide the most value.” Anthropic, he said, effectively indicated that if it believed a customer was using its technology to build a competing model, it would block access. The company backed off somewhat after criticism that preventing AI research with its models was anti-competitive, but Kantrowitz said the concern was real.

Roy then raised a challenge from Bill Gurley: if a company says it is on the verge of AGI or ASI, why can’t its model recognize espionage or illicit distillation in real time? If the system can cure cancer in a few years, Gurley asked, shouldn’t detecting distillation be easier? Roy found the point compelling because distillation is existential to the labs’ businesses; relying on withholding access suggests they have not solved a problem they urgently need to solve.

Kantrowitz noted that Anthropic had recently accused Alibaba of distilling its models, reinforcing that the issue is not theoretical. But even if limited release helps defend the frontier, the business tradeoff remains: the fewer customers who can access the best model, the more reason the rest of the market has to build around substitutes.

That tension colored the discussion of OpenAI’s potential IPO timing. Kantrowitz cited reporting that OpenAI had hired bankers and lawyers with an eye toward a public offering as soon as the third or fourth quarter of this year, but was leaning toward waiting until next year because it would not hit the $1 trillion valuation Sam Altman was seeking. He also cited the context of SpaceX’s IPO: a massive offering at a $1.77 trillion valuation that later pulled back.

Kantrowitz offered a speculative read: OpenAI may know it is important to beat Anthropic to public markets, especially given Anthropic’s recent growth, and may be signaling delay while preserving the option to move quickly if it sees an opening. Anthropic’s Fable problems could make that opening more attractive.

Roy thought IPO-related leaks are often purposeful and found Kantrowitz’s theory plausible as a strategic possibility. But he leaned toward waiting. Two or three months earlier, he said, the market had a clear direction: huge ARR numbers, enthusiasm about state-of-the-art closed models, and a dominant narrative around the latest frontier system. Since then, the “vibe shift” has been unmistakable. People such as Bill Gurley and Coinbase CEO Brian Armstrong were, in Roy’s phrasing, “bragging about not using the frontier models.” Instead of talking only about the newest Anthropic or OpenAI benchmark, the market was suddenly discussing Qwen, GLM-5.2, and model routing.

Kantrowitz challenged the wait-and-see view. If OpenAI waits too long, he said, it may spend heavily while customers learn to “efficiency max” their usage. The growth curve could flatten as buyers route work away from expensive models.

Roy’s answer was that, in his view, the curve may already have changed. The last few months would have to be shown in any IPO disclosure. Instead of going public on a simple story of an “insane” ARR curve driven by premium model usage, he thought OpenAI might be better off allowing newer business lines to mature and going out with a stronger, more diversified story.

The cost shift is no longer just a social-media meme. Kantrowitz cited reporting from The Information on customers lowering their Anthropic and OpenAI bills. One example was Ensemble Health Partners, a hospital software provider planning to spend up to $100 million on AI this year. Ensemble had success switching to an OpenAI model that was one-twentieth as expensive as the company’s more advanced models. The model powered a tool that writes appeal letters to insurance companies that refuse reimbursement for care provided by Ensemble’s hospital clients. Ensemble sends around 15,000 such letters a month, and the company’s chief technology officer said the switch would save nearly $700,000 a year.

Kantrowitz said this is what happens when smaller “flash” models become good enough. Companies inspect their spending, decide the current pattern makes no sense, and move work to cheaper models. More powerful systems may remain in the stack as a top-layer decision maker rather than the default worker.

Roy said the most important detail is that the cheaper replacement in the example was still an OpenAI model. That makes the issue more subtle and more threatening to the IPO narrative. The problem is not only that customers might leave OpenAI or Anthropic for open source. Even within the same provider’s product suite, moving from expensive frontier models to cheaper models weakens a story built on premium-model consumption. A healthy platform story would say customers should use the full suite efficiently. But the market narrative around the labs has depended on the most expensive models carrying the growth.

Kantrowitz raised the counterargument: Jevons paradox. If model use gets cheaper, customers may use much more AI overall. A category they barely spent on in 2023 could become a large and growing budget line, even if the per-task model is cheaper.

Roy said he believes that long-term. Over the next decade, he expects “massive increases” in token consumption and the “agentification of everything.” But timing matters. The frontier labs face immediate pressure to show growth and justify valuations. In a lower-pressure environment, he said, the product should naturally recommend the right model for the job. He has seen Claude sometimes explain which of its models are best suited for which part of a task when asked. But the companies do not fully bake that into the core product because, in the short term, it is not good for the business.

Kantrowitz pointed to a previous controversy around GPT-5 routing queries to appropriate models inside ChatGPT. Roy distinguished between consumer chat routing and enterprise workflow routing. In a casual back-and-forth conversation, users may dislike being silently routed to a weaker model, especially if the result feels worse. In repeatable agentic workflows, by contrast, assigning the right model to the right task is exactly where routing becomes valuable. If a workflow has predictable components, customers should be much more deliberate about model selection.

Roy’s critique was that OpenAI and Anthropic should be leading customers into that portfolio approach, but the IPO pressure makes doing so dangerous in the near term. It would be product-correct and customer-friendly, but it would also teach customers to spend less on the flagship systems.

The same enterprise shift toward scrutiny is exposing billing problems. Kantrowitz cited another Information report about Vaudit, an AI bill-auditing startup, which found mistaken overcharges in bills from AI providers. Between March and June, Vaudit audited bills sent to 60 companies totaling $34 million, mostly for Anthropic Claude Code usage, and found about $1.7 million in mistaken overcharges.

Kantrowitz emphasized that $1.7 million on $34 million is not insignificant. The issue is not necessarily malice; it is opacity. AI systems consume tokens in ways many customers do not inspect closely. Tools such as Claude Code may go off and perform work on a user’s behalf, but the customer may not be able to easily tell whether the charge accurately reflects useful completed work.

Roy was less alarmed than he might normally be. If companies were not checking their bills, not digging into their usage, and sometimes not even knowing what they were spending, then errors were inevitable. He pointed to the example of Uber reportedly blowing through an entire year’s AI budget in a quarter as evidence that tracking has been weak. In that environment, billing mistakes can occur without implying bad intent.

The more interesting issue for Roy was how to treat failed or timed-out sessions. If a system consumes tokens while thinking but then times out or produces no answer, should the customer be responsible for those tokens? The reporting treated this as a billing error in some cases, but Roy said he had not necessarily thought of it that way before. In ordinary AI usage, failures and timeouts are often something the customer “just eat[s].” As usage grows, however, the question of who pays for consumed but nonproductive tokens becomes more important.

Kantrowitz said this happens to him multiple times a day: he asks a chatbot something, it thinks, and then gives up. Whatever happens behind the scenes may still have a cost. As these companies move toward major financial events, he said, those costs will add up.

The broader implication is that enterprise AI buyers are entering a new phase. In the first phase, the urgency was adoption. In the next, procurement teams, CTOs, and finance organizations will ask harder questions: Which model actually did the work? Was the expensive model necessary? Were tokens billed accurately? Did failed sessions count? Were routing and caching optimized? That scrutiny may be good for customers and for mature AI infrastructure, but it is another complication for labs trying to show simple, accelerating revenue growth.

Kantrowitz used SpaceX’s IPO performance as a reference point for the AI labs’ public-market timing. SpaceX, he said, had raised more than $85 billion at a $1.77 trillion valuation in the largest IPO ever, then traded up dramatically before falling back. By Friday, he said, it had shed about 15% for the week, closed around $153, and dropped from roughly a $2.5 trillion market capitalization to around $2 trillion.

Roy’s view was that $2 trillion is still frothy. He described the first week after the IPO as one of those market-cycle moments people remember: headlines about Elon Musk becoming a trillionaire and heading toward two trillion, SpaceX briefly surpassing Amazon’s market cap, surpassing Meta, and closing in on Microsoft.

The comparison was striking because, as Roy put it, SpaceX had $18.7 billion in total revenue and was losing $4 billion, while Amazon had $750 billion in revenue. He allowed that investors can tell a story about data centers in space or other expansive future businesses, but the scale of the valuation relative to current revenue made the run-up memorable regardless of what happens next.

For OpenAI and Anthropic, the implied lesson was not that a trillion-dollar AI IPO is impossible. It is that public markets may reward a grand future narrative, but they can also turn quickly when the valuation stretches far beyond the financials. If AI customers are already optimizing away from premium models, and if access to frontier systems is politically mediated, then the S-1 story becomes harder to tell cleanly.

Apple’s price increases produced the sharpest disagreement with corporate framing. Kantrowitz cited Bloomberg reporting that Apple raised prices on Macs and iPads after Tim Cook said soaring memory and storage-chip costs would force the company’s hand. Mac prices rose roughly 15% to 20%, while iPad prices rose 15% to 25%. Specific examples included the base MacBook Air rising $200 to $1,299, the base MacBook Pro rising $300 to $1,999, the entry-level MacBook Neo rising $100 to $699, the iPad Air rising $150 to $749, and the iPad Pro rising $200 to $1,199.

Kantrowitz had argued publicly that the move crossed from supply-and-demand pricing into greed. He cited Micron’s gross margin of 84.9% and Apple’s gross margin of 49.3%, while acknowledging that memory prices had risen. His objection was proportionality: a data center buyer installing massive amounts of RAM may have a clear reason to raise prices, but Apple putting relatively small amounts of memory into consumer devices does not obviously need to raise the entire device price by 15% to 25%.

Roy’s answer was direct: “Greed.” His tell was the HomePod mini. As an owner of multiple units, he said he did not know whether it had any storage, but it certainly did not have meaningful storage; raising that price by 30% suggested Apple was moving across the board rather than simply passing through component costs.

Both treated the hikes as unusually large for Apple’s consumer hardware. Roy compared it to airline pricing responding dynamically to oil costs: fares can move quickly because the input cost changes quickly. Apple’s products, by contrast, are long supply-chain products. Once prices rise, Roy suspected they will stay there.

Kantrowitz called Cook’s claim that the hikes were unavoidable “ridiculous.” Even if DRAM prices rise significantly, he argued, that does not account for a 20% increase in the entire device price. Apple will make margin on that increase. He also said Apple is partly responsible for the memory-market dynamics it now laments. He read a post arguing that Apple and hyperscalers had spent a decade squeezing memory makers below cost, contributing to consolidation in which a field of many players shrank to three survivors. In that telling, Micron’s high margins now reflect survival through brutal downturns and newfound pricing power, not a simple case of supplier gouging.

Roy asked whether Apple should do it anyway. Kantrowitz said no, calling it customer-hostile and bad business. The strategic reason is Apple’s services business: if Apple wants services to keep growing, it needs devices in people’s hands. Higher device prices could slow that installed-base growth.

Roy countered that Apple users are locked into the ecosystem. He noted that the premium iPhone model price had risen 65% over seven years, and he offered himself as an example: he does not buy each new iPhone because he does not feel he needs to, but he pays too much for services he forgets about. That may suggest Apple can tolerate slower device upgrades if the installed base keeps paying recurring services revenue.

Kantrowitz then offered a succession-related theory. If Tim Cook wants to set up John Ternus well, raising prices now creates a call option. If demand remains steady, Ternus inherits higher profits. If demand falls, Ternus can enter as the executive who cuts prices to revive demand and earns goodwill. Roy called it “Call option Cook.”

Still, both viewed the size of the increases as shocking. Kantrowitz rejected the idea that criticizing Apple’s pricing means treating it like a charity. A company can “get your bag,” he said, and still be “beyond reproach.” His position was that Apple historically presented itself as a “change the world” company making great products for as many people as possible; good profit makes sense, but maximizing beyond that can still be criticized.

The final subject was Om Malik, the founder of GigaOm, blogger, and venture capitalist at True Ventures, who died on June 24 after a long health journey. He was 59.

Kantrowitz described Malik as “one of the good ones” in tech: generous with time and advice, deeply read-in, humble, and blunt in a way that could be funny. Malik had appeared on the show, and Kantrowitz recalled being surprised that Malik knew Big Technology and knew his work. He also remembered sitting near Malik at an Apple event when another journalist asked how he managed to write for The New Yorker. Malik looked at him, smiled, and said, “Because I’m good.” Kantrowitz added: “And he really was.”

Roy said Malik had mattered to him even though they never met in person. When Roy was running a startup in 2013 and trying to get coverage in TechCrunch, GigaOm, and similar outlets, Malik responded to him and exchanged friendly emails. That stood out against the broader ecosystem. Roy also credited Malik as one of the first bloggers he read and one of the people who got him interested in “the whole new media thing.”

Kantrowitz said Malik “kicked off the modern tech blog era for sure,” but the tribute emphasized not only professional influence. He read from a condolence note Malik had written to Katie Jacobs Stanton after her father died. The visible message, shown on screen, said in part: “We all have a habit of mourning what we don’t have, instead of celebrating all the gifts we were bestowed upon. His life was a gift for you.” Malik wrote that although he did not know Stanton’s father, “it is clear his grandness lives in you,” and he hoped memories and lessons would serve as “a guiding light for the future.”

Kantrowitz said much the same could be said about Malik. His presence around the tech industry, he said, brought joy to many people.