All-In Analysis of the Next Phase of AI: Model Convergence and Value Shifting to Workflows

Original Title: Anthropic IPO at Risk, Meta's Muse Pop, Token Prices Fall, Open Source Gains Share, Alignment Fails
Program: All-In Podcast
Participants: Chamath Palihapitiya, Jason Calacanis, David Sacks, David Friedberg

Editor’s Note: Over the past two years, competition in the AI industry has revolved around one core variable: who has the strongest foundational models. Models like GPT, Claude, Gemini, and the rapidly emerging DeepSeek and Qwen have continuously reset benchmarks, making "cutting-edge capabilities" the most important pricing basis for model companies. However, as the speed of model iterations increases and inference costs continue to decline, discussions are shifting from "whose model is the strongest" to "when there are enough good models, how much is the strongest model worth?" As model capabilities gradually transition from scarce commodities to more readily available foundational abilities, a more critical question arises: what is truly scarce in the AI industry in the long term—the models themselves, or the products, workflows, and distribution capabilities built upon them?

In the latest episode of the All-In Podcast, tech entrepreneurs Jason Calacanis, Social Capital founder Chamath Palihapitiya, Ohalo CEO David Friedberg, and Craft Ventures co-founder David Sacks discussed the recent flurry of model releases, the expansion of open-weight models, the decline in token prices, and the business models of leading model companies like Anthropic and OpenAI. Rather than focusing on "how much performance the next generation of models has improved," this conversation is more noteworthy for attempting to answer where the profit pool in the AI industry might migrate after the rapid diffusion of model capabilities.

The four hosts of All-In Podcast discuss the commercialization of AI models, open-weight models, and the business models of leading models.

In this discussion, the hosts effectively deconstructed "models becoming stronger" into a set of more fundamental structural questions: as the performance gap between models narrows, why would companies still pay a premium for cutting-edge models? As open-weight models can handle an increasing number of everyday workloads, is the token itself moving towards commoditization? And as "brains" become cheaper, will value concentrate back on agents, enterprise workflows, and vertical applications?

First, model competition is shifting from capability scarcity to cost efficiency. In the past, there was often a significant performance gap between the most advanced models and the second tier, and companies were willing to pay a substantial premium to achieve higher accuracy, stronger inference, and more stable outputs. The current change is that model releases are becoming increasingly dense, the performance of open-weight models is continuously approaching that of closed-source cutting-edge models, and API providers are also continuously lowering prices. This means that for many tasks that do not require extreme capabilities, "the best" is transitioning from a necessity to a choice that can calculate input-output ratios. Model performance remains important, but the ratio of price to performance is beginning to outweigh simple benchmark rankings.

Second, tokens may be transitioning from high-margin products to standardized foundational inputs. The logic of the previous round of AI commercialization was relatively straightforward: model companies provided intelligence, developers purchased tokens, and then packaged them into applications externally. However, as similar tasks can be transferred between multiple models and price differences reach several times or even dozens of times, companies will naturally allocate different tasks to models at different cost levels. Simple summarization, customer service, internal processes, and ordinary coding tasks can be assigned to cheaper models, while only a few tasks such as scientific research, complex engineering, and high-value financial decisions continue to call upon the strongest models. The real challenge here is not whether cutting-edge models have value, but how much revenue must rely on cutting-edge capabilities to be viable.

Third, the competitive boundaries of model companies are being forced upward. Chamath likened foundational models to "brains," while referring to the tools, browsers, memory, permissions, and task planning systems built around models as harnesses. This distinction is important: as the gap between "brains" continues to narrow, the factors that ultimately determine product experience may increasingly not be the models themselves, but whether they have "limbs" and can truly complete tasks. Products like Meta Muse thus take on another layer of meaning—users do not care which model is being called behind the scenes; they care about whether it can organize emails, book flights, process documents, and complete work. The cheaper foundational models become, the more economically viable these products are.

Fourth, the business models of cutting-edge models may thus polarize. On one end is the massive but highly price-sensitive general intelligence, supported by open-weight models and low-cost models; on the other end are fewer but willing to pay a high premium for Frontier Intelligence, such as complex research, mathematics, engineering, biomedicine, and highly competitive professional scenarios. Therefore, the key question for companies like OpenAI and Anthropic is no longer just "can we continue to improve the models," but whether they can prove that the incremental value created by the most advanced models is sufficient to cover their higher training, inference, and capital expenditures in the long term.

If this discussion can be compressed into a single judgment, it is this: the core contradiction of AI is gradually shifting from "is the intelligence strong enough" to "how is the intelligence priced, allocated, and productized."

In this sense, the subjects of this article are no longer just OpenAI, Anthropic, or a specific new model, but a repricing of the entire AI industry value chain: as models themselves become increasingly accessible, the truly scarce assets may revert to users, scenarios, workflows, and the ability to convert intelligence into actual value.

Below is the original content (for ease of reading and understanding, the original content has been reorganized):

TL;DR

AI model competition is shifting from "who is the strongest" to "who can achieve sufficient levels at a lower cost," essentially moving model capabilities from scarce commodities to standardized supply.

The continuous decline in token prices indicates that the profit margins of foundational models themselves may be compressed, and the AI value chain is shifting from "selling intelligence" to "selling results."

The biggest impact of open-weight models is not to replace cutting-edge models but to take over a large number of general workloads, forcing companies to recalculate whether the "premium for the strongest model" is worth it.

In the future, cutting-edge models are more likely to focus on high-value complex tasks in research, engineering, and finance, with the key not being to cover all scenarios but to prove their irreplaceability.

As model capabilities converge, competitive advantages will increasingly shift to agents, tool calls, memory, permissions, and workflows, essentially moving value from "brains" to "execution systems."

The cost structure of enterprise AI will shift from "uniformly calling the strongest model" to layered routing, matching different tasks with models at different price and capability levels.

The core business risk for OpenAI and Anthropic is not that models stop progressing, but that a significant portion of existing revenue is ultimately commoditized by cheaper models.

In the next phase, what may truly be scarce in the AI industry is no longer the models themselves, but the users, scenarios, distribution, and the ability to convert intelligence into actual output.

Video Highlights

Models Are Getting Stronger, But More Importantly: They Are Getting Cheaper

In recent years, there has been a very clear logic in AI model competition: whoever has the strongest model possesses scarcity.

Each time GPT, Claude, or Gemini crosses a capability threshold, it brings a new wave of product and capital narratives. When the performance gap is large enough, companies are willing to pay a noticeable premium for "the strongest intelligence."

However, All-In believes this logic is becoming more complex.

Friedberg cited a series of recent model updates during the program. Although there are discrepancies in some model types and release dates in automatic transcription, the underlying trend is clear: model updates are becoming extremely dense, and performance improvements are occurring alongside cost reductions.

For example, DeepSeek released V4.1 Flash on September 10. This is a 552 billion parameter MoE (Mixture of Experts) model, but only about 8 billion parameters are activated during input, and about 16 billion parameters are activated during output. DeepSeek also lowered API prices and stated that the new KV Cache design will significantly reduce memory and storage requirements.

OpenAI is also moving in the same direction. The GPT-6 Sol and Luna launched on September 22 have API prices reduced by approximately 50% compared to the previous GPT-5.6 promotional prices.

Anthropic launched Claude Opus 5.5 on the same day. Its input price per million tokens dropped from $5 for Opus 5 to $4, and the output price fell from $25 to $20; for Cache Read, which is frequently used in agent scenarios, the price dropped from $0.5 to $0.2. Anthropic stated that, based on comprehensive token usage efficiency calculations, the typical task operating cost could decrease by about 40%.

Individually, each company's price reduction can be explained as architectural optimization or economies of scale. However, viewed collectively, the signal is clearer: AI intelligence itself is undergoing rapid price compression.

Open Weights Begin to Capture Volume, Tokens Are Becoming Commodities

In addition to price declines, another change is occurring: an increasing number of tasks do not necessarily require the most expensive models.

Here, it is necessary to distinguish a concept that is often confused. The program frequently uses "open source," but strictly speaking, open-weight does not completely equal open-source. The former typically means that model weights are downloadable and can be deployed independently, but licenses, training data, or complete training code may not be fully open.

For enterprises, the commercial implications are very direct.

If an internal summary, customer service classification, code assistance, or routine agent workflow can be completed to a sufficiently good level using a model that is ten times cheaper, then it will become increasingly difficult to explain to the CFO why they should continue to call the most expensive frontier model.

Data from Vercel's AI Gateway has already shown a similar trend. Its Production Index, released in September, indicates that the proportion of open-weight models processed by Vercel AI Gateway has risen from 7% in December 2025 to 56% in August this year, surpassing half for the first time.

Meanwhile, the average token cost has dropped to less than half of what it was five months ago, with a 23.2% decrease in unit token price in just August alone.

It is important to emphasize that this is the production traffic data from Vercel AI Gateway, and it cannot be directly extrapolated to claim that "56% of AI tokens globally come from open models." However, it at least provides a sample from a real production environment: developers are increasingly migrating workloads to cheaper models.

Chamath summarizes this change as "models beginning to cluster." His judgment is that more and more models are approaching a level of capability on a wide range of tasks that makes them interchangeable. Thus, the question is no longer just "which model scores the highest," but rather: how much more are you willing to pay for that last bit of performance difference?

Cutting-edge Models Won't Disappear, But Must Prove Their Value

This is also the real business challenge faced by OpenAI and Anthropic.

David Sacks expressed a more optimistic view in the program compared to other hosts. He believes that even if open-weight models take away the vast majority of tokens, it does not mean that Frontier Intelligence loses its commercial value.

The reason is that there will always be a set of tasks for which there is a high willingness to pay for the "best." For example, drug development, complex mathematics, cutting-edge engineering, cybersecurity, or highly competitive financial trading scenarios. For these tasks, if a stronger model can improve the success rate even slightly, the economic value it creates may far exceed the token cost.

Friedberg shares a similar view. The life sciences research organization he is part of uses Anthropic's models because the capability differences among models are still significant for some high-difficulty scientific tasks.

Therefore, what is truly worth paying attention to is not whether "open models will eliminate Claude or GPT," but a more specific question: how much of OpenAI and Anthropic's revenue comes from truly irreplaceable frontier tasks, and how much comes from ordinary tasks that can eventually be migrated to cheaper models?

If the vast majority of revenue comes from the former, model companies still have strong pricing power. However, if a large portion of revenue is simply due to enterprises temporarily accustomed to using the strongest models, then as enterprises optimize AI costs, this revenue may face increasing price pressure.

This is why the customer structure and revenue quality of model companies may be more worth observing in the future than simple token growth.

From a capital market perspective, this is particularly important. Investors need to judge not just whether "Claude or GPT can continue to grow," but how sustainable the frontier intelligence premium behind this growth really is.

As "Brains" Become Cheaper, Value Begins to Shift to "Hands and Feet"

Chamath raised another judgment worth noting in the program. If we view models as "brains," the true determinants of the final product experience are increasingly becoming the agent harnesses that wrap around the brain.

The so-called harness can be understood as the entire system that enables the model to "do things," including browsers, tool calls, memory, task planning, permission management, code execution, and the ability to connect with different software. The same model, when placed in different harnesses, may perform completely differently.

This means that as foundational models converge, the focus of competition among AI companies may begin to shift upward.

The core question in the past was: who has the smartest model? The next question may increasingly become: who can make this model actually get the work done?

Meta's Muse is a representative of this trend.

On September 8, Meta launched Muse, positioning it as a personal AI agent: it not only answers questions but can operate browsers and applications through an independent virtual computing environment to complete tasks such as sending emails and arranging travel.

In the program, Chamath mentioned that he used Muse to organize his personal email and tried to have it book flights and hotels. He believes that the most important significance of such products is not which underlying model is used, but that it packages complex AI capabilities into software that ordinary users can directly use.

This also explains why a decline in token prices does not necessarily mean a shrinkage in the value of the AI industry. On the contrary, cheaper intelligence may make more products economically viable. A thinner profit margin at the model layer does not equate to the entire AI stack earning less. Value may simply be redistributed from the model itself to companies that own users, workflows, and final products.

The Next Round of AI Competition May Not Be About Whose Model is First

This change is still in its early stages.

There are still capability differences between cutting-edge models and open-weight models, and different benchmarks do not yet represent all real enterprise tasks. The assertion made by All-In that "models are converging" is currently better understood as a business judgment rather than a completed industry fact.

However, price changes have become clearer. Open-weight models are entering more and more production environments, API unit prices are continuously declining, and closed-source model vendors are also actively lowering prices. Meanwhile, products like Meta Muse are beginning to attempt to push AI from chat windows to agents that can actually perform tasks.

What is truly worth observing next is not just how many points the next generation of GPT or Claude has improved on benchmarks.

More importantly, there are four variables:

First, whether the share of open-weight models in real production tokens can continue to increase; second, how quickly the unit cost of intelligence can decline; third, how many irreplaceable complex tasks high-priced frontier models can retain; fourth, whether agents can truly form stable usage frequency and commercial revenue.

If the first three trends continue to develop, AI models may increasingly resemble the foundational computing power in cloud computing: still important, but with prices becoming increasingly transparent and standards becoming more unified.

And what may truly determine the value of companies in the next phase is no longer just "who has the smartest brain," but who can turn increasingly cheaper brains into products that people are genuinely willing to use and pay for.

[Video link]