“LLMs will become swappable commodities”: our call—only below the top tier, so far — resolves 2027-07-25

The debate: many people say the big AI models will become so alike that they can only compete on price, turning into a commodity like electricity.

Why it matters: if that is true, buyers should treat a model as a swappable part and avoid lock-in; if it is not, the cost of lock-in and the risk of price increases need to be recalculated.

Who is on each side: the commodity camp uses chip foundries as its analogy (anyone can make it, so only price matters); after checking three watchpoints, we accept only a layered version, not across-the-board commoditization.

Our call: not yet at the top tier. The strong version of the claim, that every large model becomes an interchangeable good sold on price, does not hold so far; a layered version does. The most expensive tier still has pricing power while the tiers below it are getting cheaper, the same shape as chip foundries, where the leading edge is a near-monopoly that can raise prices and only the trailing edge competes on price. Commoditization is a line climbing the ladder, not a switch that flips for a whole layer. The card resolves on 2027-07-25; readings as of October 5 follow.

Correction and update, October 5, 2026: this section and the headline used to say the thesis “fails” against all three of its own watchpoints, which had all moved the opposite way. That said more than the evidence: only the enterprise-share reading has a before-and-after comparison, the 11% switching figure has no baseline, and the margin reading gave no number. The September 30 interim check has also produced readings (last three paragraphs). Both have been rewritten; the direction of the call is unchanged, but its support is weaker than first written.

Switching: in a mid-2025 survey by the venture firm Menlo Ventures, only 11% of teams had changed model vendors in the past year, and 66% had upgraded within their existing vendor. Money moves slowly, but with no earlier figure to compare against, this alone does not show strong lock-in.

Margins: the research firm SemiAnalysis estimates that Anthropic's inference gross margin rose from 38% in 2025 to the mid-60s in the first quarter of 2026. That is an outside estimate; the company has not published audited figures.

Enterprise share: Menlo measured Anthropic's share of enterprise LLM spending rising from 32% in mid-2025 to 40% in December 2025, with the top three at 88% combined. This is the only one of the three with a before-and-after comparison, and it points to concentration.

September 30 interim check: Menlo has not published a new report. As of October 5 the latest is still the December 2025 one, so there is no new share reading. The Sonnet item does have a reading: Anthropic's official pricing page states that Claude Sonnet 5's introductory price, $2 per million input tokens and $10 per million output tokens, is now the standard price, and the $3/$15 standard price scheduled for September 1 will not take effect.

Under item (3) of the receipt, that means the call has to be revisited. The result: “flagship prices are rising” now holds only for the most expensive tier. OpenAI's new flagship GPT-6 Astra lists at $10/$50, above July's GPT-5.5 at $5/$30; Anthropic's top model Fable 5.1 stays at $10/$50, but the tier below, Opus 5.5 (released September 22), costs 20% less than Opus 5.

The flip condition on prices needs both a real cut in flagship pricing and the Sonnet promotion not reverting. The second half has happened. In our reading the first has not: the top tier's list price has not fallen, and while Anthropic says Opus 5.5 performs at the level of Fable 5.1 on most work, that is the vendor's own claim. This is the closest the call has come to flipping; commoditization has climbed one rung.

BG2 Pod / Altimeter (guest host), interview with Ali Ghodsi (Databricks CEO) and Arvind Jain (Glean CEO)
“the the LLMs is a commodity as Arvind said. You…”

Stratechery(Ben Thompson)

Stratechery(Ben Thompson)

Anthropic (official newsroom)
“Today, Claude Code's run-rate revenue has grown to…”

Checked through 2027-07-25. There are five things to watch, in two grades; the numbers match "What would overturn it" below.

Three would overturn the whole call (any one of these three flips it):

(1) Menlo Ventures' next enterprise report, expected around 2026-09, shows enterprise spending share becoming less concentrated.

(2) Anthropic's audited gross margin, once disclosed in a listing filing, comes in well below the mid-60s.

(3) Flagship prices are actually cut AND Claude Sonnet's end-August promotion does not revert to standard pricing. Both have to happen; one alone is a warning sign, not a flip. The Sonnet half was confirmed on 2026-09-30 (the promotional price became the standard price); flagship prices have not been cut.

Two would only downgrade one part of the reading, not overturn it:

(4) Within 12 months a lab fails to bring a second adjacent application in-house while a third-party app holds its ground on that turf: the reading that a model maker that also builds apps has a ceiling gets downgraded.

(5) Once the compute shortage clears, the data and application layers grow faster than the labs: the case that value is moving downstream gets stronger.

What would overturn it

Any one of these three flips the call, revives the commodity thesis and voids the "near-monopoly at the top" reading (numbers match the first three above):

(1) Menlo's next report shows enterprise spending share becoming less concentrated.

(2) Anthropic's first disclosed gross margin comes in well below the mid-60s.

(3) Flagship pricing is actually cut AND the Sonnet promotion does not revert to standard pricing (both together).

Read the full deep dive →

Jul 25, 2027

2026-09-30 (interim check: Menlo Ventures' mid-year report, and whether the Sonnet price reverts after August 31)