Premiums Take More Work

If You Only Read One Thing

Better products can leave the bill surprisingly unchanged. Moonshot’s Kimi K3.1 Pro cuts token prices while doing more work per task; American Express collects higher card fees while spending more to retain customers. The independent Kimi evaluation exposes the common question: how much of an improvement reaches the customer, and how much must the supplier spend to keep earning a premium?

Kimi Makes the Premium Harder to Defend

Moonshot has made proximity to the AI frontier cheaper to purchase. It has not established that every completed job will cost less. That distinction matters for the companies financing expensive proprietary models: competition can force more capability into the same customer budget before it forces the budget down.

Kimi K3.1 Pro launched on October 23. Artificial Analysis’s independent evaluation places it fourth on its Intelligence Index, at 55 versus 44 for Kimi K3. The leader scores 58. These are results on a particular basket of tests, not proof that the models are interchangeable across every business.

The economic detail is more revealing than the ranking. Artificial Analysis reports first-party prices of $2 per million input tokens and $10 per million output tokens, a third below K3’s rates. Yet its launch evaluation puts cost per task at roughly the predecessor’s level. K3.1 Pro generates more output during those tests. A cheaper unit can support a more ambitious process without producing a cheaper invoice.

That gives Moonshot two routes to pressure rivals. Customers can seek similar results at a lower unit price, or spend the same amount for more capable work. Either route makes a premium harder to justify through model reputation alone. The relevant comparison becomes the result purchased by the entire bill, including checking and correcting the answer.

Availability still limits the competitive claim. Moonshot’s announcement makes the model available through its products and API, but promises the full downloadable weights only by November 4. Independent hosting is a prospective market change, not a service already established by this launch. The terms and practical cost of running those weights will matter.

My medium-confidence investment read over the next 12–24 months favors software businesses that can substitute model suppliers while retaining valuable customer workflows. Model vendors depending on a persistent capability premium face more pressure. This extends the distribution argument in October 16’s licensing analysis, with new evidence that a rival can approach the frontier on independent tests.

The strongest objection is straightforward: reliability, support and specialized performance may justify a premium even when aggregate scores converge. The test is sustained paid adoption outside Moonshot’s own products after weights arrive. A high ranking without recurring outside deployment would weaken the bargaining-power thesis.

Amex Pays to Keep the Premium

American Express’s advantage is that customers pay to belong as well as to transact. Its latest results show that the membership proposition needs continuing investment. Higher fees are evidence of pricing power; the cost of keeping the proposition attractive determines how much of that power becomes profit.

The October 23 earnings release reports $20.1 billion in quarterly revenue, up 9%, and $4.68 in adjusted earnings per share, up 13%. Reported earnings grew faster, helped by a business-sale gain. The adjusted measure also removes the associated restructuring effect, making it a cleaner starting point for comparison.

The more useful split sits in the quarterly filing. Net card fees rose 16%, while merchant discount revenue, the fees collected on card spending, rose 8%. Amex attributes card-fee growth to higher average fees on premium products and expansion of those portfolios. It can deepen an existing relationship without relying entirely on another shopping trip.

Keeping that relationship valuable is expensive. Card Member services expense rose 36%, with new Platinum benefits and greater benefit usage among the drivers. Those expenses and membership fees are not a self-contained profit segment: benefits also encourage spending, and other costs sit elsewhere. Subtracting one line from the other would manufacture a margin the company does not report.

Amex also plans roughly 3,500 job cuts, a spokesperson told Banking Dive. The question is whether it can buy loyalty more efficiently than rivals. A benefit that draws spending onto an Amex card can support both renewal and merchant revenue. A benefit added merely to match a competitor can become a recurring cost with little differentiation. The same customer activity can therefore strengthen the network while making the premium harder to maintain.

My medium-confidence view over the next year favors Amex’s ability to defend customer revenue over an assumption of easy margin expansion. Its established membership and merchant relationships offer more ways to recover the cost of benefits than a standalone perk provider has. This applies the familiar customer-relationship pattern outside AI; it does not establish that any current share valuation is attractive.

The counterargument is that much of the spending builds benefits whose appeal lasts beyond the launch year. If renewal stays strong while service-cost growth normalizes, the investment can pay back. The next quarterly comparison between card-fee growth and services-expense growth will test that operating argument, alongside retention commentary.

The Contrarian Take

Everyone says: Strong spending means consumers are comfortable, so gloomy surveys must be missing the real economy.

Here’s why that’s incomplete: A premium-card portfolio and a national household survey measure different populations and different things. The University of Michigan’s final October survey puts sentiment at 48.0, almost unchanged from September, while year-ahead inflation expectations rose to 4.7%. Amex’s results cannot establish that the median household shares its customers’ spending capacity. Nor does a larger nominal bill establish that a customer bought more: higher prices can increase recorded spending while leaving satisfaction worse. The useful conclusion is narrower than either optimism or recession alarm. A business attached to relatively resilient customers can keep growing in an economy where many households dislike the terms of participation.

Under the Radar

Amex’s credit improvement contains a modeling decision. The filing says the quarterly card-balance reserve release primarily reflected enhancements to expected-loss models. A reserve release reduces the amount charged against profit for anticipated losses; it is different from collecting cash previously thought lost. Credit performance can remain healthy, but this earnings contribution should not be treated as an indefinitely repeatable improvement in customer behavior. The reserve explanation supplies that distinction.
Kimi’s enterprise launch preserves an internal gate. Organization administrators must enable K3.1 Pro; access is off by default. That launch condition separates a model’s public availability from an employee’s permission to use it. Better capability can arrive faster than enterprise approval. For providers competing for organizational budgets, migration effort and administrative confidence can therefore preserve an incumbent’s position even after a rival improves its test scores.

Quick Takes

Slower activity is not yet inflation relief. October’s US composite purchasing managers’ index, a survey measure of business activity, slipped to 57.6 from 58.4; readings above 50 indicate expansion. Selling prices rose at the fastest pace since August 2022. Europe’s corresponding activity index fell to 51.9, but price pressure remained elevated. The shared problem is that cooling growth alone does not establish cheaper inputs or room for monetary easing. (Source)
CRH now owns the infrastructure exposure. CRH completed its Arcosa acquisition on October 23, following June’s agreement at approximately $8.5 billion enterprise value. The assets combine construction aggregates with energy-transmission products. Ownership gives CRH a broader route into infrastructure spending; it also puts integration and capital-allocation performance inside its own results. Completion removes deal uncertainty, not the need to earn a return on the purchase price. (Source)
Long-run inflation expectations moved too. Michigan’s October reading rose to 3.5% from 3.4%, the second consecutive monthly increase. Expectations are beliefs, not an official inflation forecast or proof of a wage-price spiral. Their persistence nevertheless raises the cost of assuming households will quickly dismiss expensive fuel as temporary. Businesses seeking another price increase may face customers who increasingly expect their purchasing power to stay under pressure. (Source)

The Thread

The cost of defending a premium deserves its own place in the analysis. Kimi makes more capable work available at lower unit prices, but the evaluated task bill does not fall proportionately. Amex collects more membership revenue while funding a richer service proposition. In each case, the improvement has to be assessed alongside the resources consumed to deliver it.

My inference is that competition can tax an incumbent through required reinvestment before it visibly reduces prices. An AI supplier may have to provide more reasoning for the same budget. A card issuer may have to supply more benefits for the same renewal probability. Revenue stability can conceal an increasingly expensive defense.

That is a different mechanism from yesterday’s retreat into narrower product portfolios. The test is whether suppliers preserve customer retention while reducing the incremental cost of each improvement. If they do, stronger products can expand margins. If each competitive advance demands another equally costly response, the customer relationship remains valuable but becomes less profitable to own.

Predictions

I predict: Moonshot will publish downloadable Kimi K3.1 Pro model weights by November 4, 2026, meeting the deadline in its launch announcement. (Confidence: high; Check by: 2026-11-05)

2026-10-24 · 03:23 AM ET