Commodity or Moat? Three Bets on the Future of LLMs

In 2023, GPT-4 cost $30/$60 per million tokens. Today, smarter models cost $1-2. That’s a 95% price collapse in three years—faster than telecom (30 years), faster than airlines post-deregulation (20 years), faster than almost any market in history. When your marginal cost of production is near-zero, price competition is brutal and fast because there’s almost no cost floor to slow the descent.

But here’s the part that doesn’t get enough attention: the prices didn’t collapse uniformly. Three companies sell into the same market, use comparable architectures, compete on the same benchmarks—and their pricing is wildly different. Not a little different. Anthropic charges 3-5x what Google charges at the budget tier. OpenAI sits in the middle. Google bundles AI into a $250/month mega-plan that includes YouTube Premium and 30TB of storage.

Why? Because each company is making a fundamentally different bet about the answer to one question: will LLMs stay differentiated, or will they commoditize?

Their pricing tells you which future each one is building for.

The Pricing Landscape ¶

Before getting into the bets, here’s where things actually stand.

API pricing (per 1M tokens, input/output):

*Gemini 3 Pro is still in preview; stable pricing may settle closer to $1.50/$10.

Consumer subscriptions:

Every provider uses the same playbook underneath—model tiers as screening mechanisms, output tokens priced 3-5x more than input, batch discounts at 50%, subscriptions that are massively cheaper than equivalent API usage. The mechanics are identical. The positioning is where they diverge.

You could argue the price differences simply reflect cost structures—Google’s custom TPU advantage, Anthropic’s smaller scale. But cost-plus pricing doesn’t explain the strategic choices around bundling, tier design, and what products they don’t offer. Anthropic could launch a budget tier and doesn’t. Google could unbundle Gemini and doesn’t. These are choices, not accounting.

Google’s Bet: Models Commoditize ¶

Google prices as if models are heading toward utility status—and if you’re the lowest-cost producer of a utility, you win.

Flash-Lite at $0.10/$0.40 per million tokens. Generous free tiers. An $8/month AI Plus plan. Google doesn’t need Gemini to be a profit center because Gemini isn’t really their product—it’s a feature inside a larger ecosystem. Workspace, Android, Chrome, Search, Cloud. The LLM is the loss leader. The ecosystem is the business.

This is clearest in the bundling. Google AI Ultra at $249.99/month—the most expensive consumer AI plan from any provider—includes YouTube Premium, 30TB storage, Google Home Premium Advanced, and top-tier Gemini. The bundle justifies a price that “$250 for AI” alone never would. Different customers value different parts: some want the AI, some want YouTube ad-free, some want the storage. Bundling captures value across all of them.

The strategic logic is simple: if models eventually converge on quality, the company that already has distribution and ecosystem lock-in wins by default. You don’t beat that by building a better model. You beat it by having built a better everything-else. Google is betting that intelligence becomes a commodity—and positioning to be the one who commoditizes it.

This is an existential threat to every standalone AI company. Google can subsidize Gemini indefinitely and recoup the cost across the rest of its business. For Anthropic and OpenAI, AI is the business. Every dollar of API revenue matters. You cannot win a price war against a company that doesn’t need AI to be profitable.

Anthropic’s Bet: Differentiation Persists ¶

Anthropic prices as if the models are not converging—and charges accordingly.

Opus 4.6 at $5/$25. Haiku 4.5 at $1/$5—their budget tier costs 10x Google’s. No ad-supported consumer plan. No $8 option. Anthropic looked at Google’s bundling strategy and said—we’re not playing that game.

This only works if the differentiation is real. And so far, it is. Dario Amodei made the point in a recent interview with Dwarkesh Patel: “Everyone knows Claude is good at different things than GPT is good at, than Gemini is good at. It’s not just that Claude’s good at coding, GPT is good at math and reasoning. It’s more subtle than that.” The tokens for “restart your Mac” are worth cents. The tokens telling a pharmaceutical company to move an aromatic ring from one end of a molecule to the other could be worth tens of millions. If you’re selling the aromatic-ring tokens, you can charge a premium.

Anthropic’s model tiers reinforce this. Haiku/Sonnet/Opus isn’t “small/medium/large”—it’s a screening mechanism designed to make customers reveal their price sensitivity. Budget customers choose Haiku. Enterprises choose Opus. The quality gap between tiers is deliberately calibrated to prevent enterprises from trading down. If Sonnet were 95% as good as Opus at hard tasks, no one would pay the 67% premium. The gap has to be real and visible—and maintaining that gap is how Anthropic justifies premium pricing.

Anthropic is betting on two kinds of differentiation: model quality (Claude is measurably better at coding, reasoning, nuanced tasks) and ecosystem trust (compliance, safety, enterprise relationships). The first can erode quickly. The second is stickier—and probably the real moat.

Being the most expensive might actually be why Anthropic survives. Premium pricing only works with genuine differentiation, which forces them to stay differentiated. If they slashed prices to match Google, they’d be fighting a two-front war against Google’s subsidies and OpenAI’s distribution—and losing both. The high price isn’t a weakness. It’s a strategic constraint that keeps them focused on the only thing that can save them: being measurably better at the tasks enterprises pay for.

The risk is obvious: what if differentiation erodes? What if open-source models close the gap? Then Anthropic is charging a 3-5x premium for a product that’s 5% better—and that math doesn’t hold.

OpenAI’s Bet: Hedge Everything ¶

OpenAI’s pricing looks incoherent until you realize they’re hedging both sides.

GPT-5 at $1.25/$10 undercuts Anthropic while staying above Google’s budget tier. It’s the middle. In a commoditizing market, the middle is usually where you get squeezed—too expensive to win on price, not differentiated enough to justify a premium.

But OpenAI isn’t just an API company, and this is what the pricing table misses. ChatGPT’s consumer distribution is arguably the strongest moat any of the three have. The $8 ad-supported Go tier (launched January 2026) is designed to capture the hundreds of millions of free-tier users who want more but won’t pay $20—OpenAI calls it their fastest-growing plan. The $200 Pro plan captures power users. That’s a 25x price gap on the same underlying technology—and the cheap version is deliberately degraded (ads, restricted models) to protect the premium tier. This is textbook versioning: create a worse version not because it costs less to produce, but because selling a good cheap version would cannibalize the expensive one.

The Microsoft partnership is the other hedge. Copilot embedded across Office, GitHub, and Azure effectively gives OpenAI its own bundling defense—the same playbook Google is running, just through a partner. Microsoft is already subsidizing OpenAI’s models the way Google subsidizes Gemini.

So OpenAI is betting on differentiation through the API (premium models, reasoning capabilities) while simultaneously building consumer distribution and bundling partnerships in case differentiation fails. The risk isn’t existential; it’s bifurcation. Their consumer business is a fortress. Their API business is the part that gets squeezed if models converge.

                    High Price
                        │
                        │         ANTHROPIC
                        │         (model quality)
                        │
Weak ───────────────────┼──────────────────── Strong
Competitive Moat        │                     Competitive Moat
                        │  OPENAI API
                        │  (developer base)  OPENAI CONSUMER
                        │                    (distribution +
                        │                     Microsoft)
                        │                  GOOGLE
                        │                  (ecosystem)
                        │
                    Low Price

The Subscription Trick ¶

One thing all three agree on: subscriptions are a better business than API. But the reason is more interesting than it looks.

Run the numbers on Anthropic’s tiers. According to Anthropic’s own data, the average Claude Code developer spends ~$6/day on API tokens, and 90% spend under $12/day. That puts typical heavy usage at $200-360/month.

Usage Level          API Cost/Month    Subscription    Savings
─────────────────────────────────────────────────────────────
Light (few queries)       $2-5            $20         -$15 (overpay)
Medium (daily use)      $100-200          $20          80-90%
Heavy (90th %ile)       $200-360          $200        break-even to 45%

The light user overpays—the $20 subscription is a gym membership they’ll never fully use. The medium user gets an 80-90% discount. And even the heavy user roughly breaks even or saves meaningfully on the $200 Max plan.

But the interesting question isn’t the savings math—it’s why Anthropic would offer a plan that at best captures the same revenue as API billing. Because the subscription isn’t a pricing decision—it’s a customer acquisition cost disguised as a product. Once Claude Code is embedded in a developer’s daily workflow at $200/month, the switching cost is enormous. Anthropic is trading margin for lock-in.

This is where the bets converge. Whether you think models commoditize or stay differentiated, lock-in at the application layer matters. Claude Code isn’t just an API wrapper—it’s a product that generates switching costs. Anthropic isn’t selling tokens; they’re selling a workflow that happens to consume tokens. Google is doing the same thing through Workspace integration. OpenAI is doing it through ChatGPT habits and Microsoft Copilot.

Every company is racing to convert model advantage into application-layer lock-in before the window closes. They just disagree on how much time they have.

Who’s Right? ¶

The honest answer: the evidence cuts both ways.

The case for commoditization is straightforward. Budget models are already at $0.05-$0.10/MTok and falling toward utility pricing. The mid-tier ($1-3/MTok) is converging on benchmarks—the jump from budget to mid-tier is transformative for most tasks, while mid-tier to premium is marginal for 80% of use cases. Open-source models keep closing the gap. DeepSeek prices at hardware cost. Meta’s Llama is free and increasingly competitive—arguably a bigger threat to the differentiation thesis than DeepSeek, because Meta has the resources to keep iterating indefinitely without needing to monetize the model itself. If this trend continues, the premium erodes and the bundler wins.

The case for persistent differentiation is subtler but real. Three years in, the models still genuinely differ—not just on benchmarks but on tone, reasoning style, reliability on edge cases. Enterprise customers pay for trust, compliance, and consistency, none of which show up in benchmark comparisons. Dario’s cloud analogy is instructive: “There are three, maybe four, players within cloud. I think that’s the same for AI.” But unlike cloud—where an EC2 instance is an EC2 instance—AI models are not interchangeable. That means more pricing power for anyone who maintains a real quality edge.

And then there’s the hedge. Consumer distribution is valuable regardless of which future arrives. If models commoditize, ChatGPT is the default interface. If they differentiate, OpenAI has the brand and the Microsoft channel. The hedge works—but it also means OpenAI doesn’t dominate either scenario. They survive. They don’t necessarily win.

The equilibrium probably looks like airlines. Economy (budget models) is fully commoditized—you won’t think about the cost, the same way you don’t think about the cost of a Google search. Business class (mid-tier) is a temporary battleground heading toward compression. First class (premium) survives for specialized, high-stakes use cases where trust and capability justify the price.

                ┌─────────────────────────────────┐
    Premium     │  Specialized, high-stakes use   │  Stable margins
    ($5-25/MTok)│  Compliance, trust, capability  │  (if differentiated)
                └─────────────────────────────────┘
                              │
                        (convergence pressure)
                              ↓
                ┌─────────────────────────────────┐
    Mid-tier    │  Current battleground           │  Temporary margins
    ($1-3/MTok) │  Heading toward commodity       │  → compression
                └─────────────────────────────────┘
                              │
                        (convergence pressure)
                              ↓
                ┌─────────────────────────────────┐
    Budget      │  Already commodity              │  Near-zero margins
    ($0.05-0.10)│  Utility pricing                │  (volume game)
                └─────────────────────────────────┘

The pricing pages tell you what each company believes the answer is. Google believes models converge and the ecosystem wins. Anthropic believes differentiation holds and quality wins. OpenAI believes the answer is uncertain and distribution wins while you figure it out.

The tempting conclusion is that someone has to be wrong. But if the market segments like airlines—economy commoditizes, first class holds, business class compresses in between—then all three could be right, just for different customers. Google wins budget. Anthropic wins premium. OpenAI captures the messy middle through sheer distribution.

The question that should keep all of them up at night isn’t which bet wins. It’s how big each tier turns out to be—and whether the tier you’re betting on is large enough to build a business around.