Anthropic: Models Under 6 Months Old Made 80%+ of Q3 Revenue

Anthropic’s S-1, filed with the SEC on 26 October 2026, says “model versions released in the six months ended September 30, 2026 accounted for over 80% of our revenue for the three months ended September 30, 2026.” The filing calls the “continuous and overlapping cadence of new model version releases” inherent to staying at the frontier.

The same filing shows where the compute went. Cost of revenue, which the S-1 says is substantially all customer-serving compute, was $2,380 million in the first half of 2026, against $8,945 million of compute and infrastructure expense: at most 27% of the total, down from at most 44% a year earlier, by our arithmetic. A chart by MoE Capital, passed to us on 27 October, shows the same figures, which we reproduced from the filing.

Video explainer · Anthropic’s S-1

A 16:9 video explainer on Anthropic’s S-1 and its model cycle, narrated from the published text of this piece and four related FourWeekMBA pieces.

The Filing on Its Model Cycle

The S-1 says revenue is shaped by “the frequency and capability of new model version releases”, and that “As we have released successive versions of Claude models, we have continued to see revenue growth as new capabilities unlock additional use cases”.

It describes compute as a loop: “We operate an iterative cycle in which compute deployed to model development yields more capable and efficient models. Those efficiency gains free compute to serve additional customer workloads.” The over-80% figure is the filing’s own example of that cadence; it does not name the model versions.

Where the Compute Went

Compute and infrastructure expense was $8,945 million in the six months to 30 June 2026, up from $2,334 million a year earlier, the S-1 says. A note to the financial statements regroups the same expenses by function: cost of revenue was $2,380 million, against $1,019 million a year earlier.

“Substantially all of cost of revenue consists of compute and infrastructure costs associated with customer-serving workloads”, the note says, and cost of revenue “includes certain personnel-related costs and allocated overhead costs”. So $2,380 million divided by $8,945 million, 26.6%, is an upper bound on the customer-serving share, by our arithmetic; a year earlier it was 43.7%.

MoE Capital’s chart labels the remaining 73% as research and training. In the S-1, research and development, $9,741 million in the half, “primarily consist of compute and infrastructure costs associated with model development, research, and internal usage workloads”, and sales and marketing include compute for “providing free access to the Company’s intelligence service”. The rest of the compute bill is spread across those lines.

The Margin in the Chart

Revenue was $16,286 million in the half, the S-1 says. Less $2,380 million of cost of revenue, that is 85.4% of revenue, against 12.4% a year earlier on $1,163 million of revenue, by our arithmetic. MoE Capital’s chart shows 85% and notes it is before cloud-marketplace fees.

The S-1 itself no longer reports a gross margin. It says that during 2025 Anthropic changed its presentation and that “cost of revenue, gross profit (loss), research and development, sales and marketing, and general and administrative are no longer presented as separate line items in the consolidated statements of operations.”

The chart compares the 85% with 40% to 50% for inference clouds that host open models, which MoE Capital bases on reported figures for Cerebras and its own estimates for Fireworks and Together. Those comparison figures are MoE Capital’s; we did not check them.

Connecting the Coverage

On 26 October we reported that Anthropic puts $2.4B of $8.9B compute in cost of revenue, from the same note, and that Anthropic runs compute as one pool, with the compute bill at 48% of revenue. Earlier that day we reported that Anthropic’s compute bill fell to 48% of revenue from 251% across the quarters in the filing.

Today we reported that Anthropic’s compute leases not yet started reached $321.4 billion at 30 June, the capacity the next models would run on.

The Structural Read

As we read it, the over-80% sentence is the filing’s own measure of how fast revenue moves to new model versions: within one quarter, models released in the prior six months carried most of it.

As we read it, the 27% is a ceiling, not an estimate: cost of revenue also includes personnel and overhead, and free-tier serving and internal use of Claude sit in other expense lines, so the S-1 gives no direct serving-versus-training split.

As we read it, the 85.4% in MoE Capital’s chart is revenue less cost of revenue, a measure the S-1 stopped presenting in 2025; it leaves out the compute booked in research and development.

Three Implications

FOR INVESTORS READING THE S-1 The filing gives cost of revenue only in a note to the financial statements, and it no longer presents gross profit, so any gross margin, including MoE Capital’s 85%, is a calculation from that note.

FOR MODEL BUYERS The S-1 says new model versions carried over 80% of third-quarter revenue, so the mix of what customers pay for changed within months, by the filing’s own example.

FOR COMPUTE PROVIDERS The S-1 says Anthropic plans compute on a total basis rather than by individual workload, so serving and model development draw on the same committed capacity.

The Business Engineer Lens

This story maps onto the Business Engineer framework The State of The Inference Economy.

The framework sets out its premise: “Training was a cost center: episodic, concentrated among roughly twenty frontier model labs, amortizable across trillions of tokens.”

As we read it, the S-1 adds a frontier lab’s own figures to that question: over 80% of third-quarter revenue came from model versions released in the prior six months, the filing says. The framework was published in March 2026 and does not discuss this filing.

What Is Not Established

The S-1 says Anthropic plans compute “on a total basis rather than by individual workload”, so in the passages we read it gives no direct split between serving and training; the 27% and 44% are upper bounds from the functional note. The filing does not name the model versions behind the over-80% figure. We could not find MoE Capital’s original post of the chart.

The Bottom Line

Anthropic’s S-1 says model versions released in the six months to 30 September made up over 80% of third-quarter revenue, while cost of revenue was at most 27% of first-half compute, down from 44%, by our arithmetic, reproducing MoE Capital’s chart.