Anthropic Revenue Breakdown: API, Subscriptions, and S-1 Data

Anthropic’s October 26, 2026 S-1 does not split revenue the way outside observers usually do. It books one operating segment and two pricing lines—token consumption versus subscriptions—then overlays a cloud-marketplace channel. API (direct, Bedrock, Vertex, Azure/Foundry), enterprise minimums, Claude Code, and Pro/Max/Team all sit inside those two lines and are not disclosed as products.

What the S-1 actually reports

The filing’s economic unit is the token, not the product. Consumption is “metered in tokens and priced according to the model selected.” Subscriptions are prepaid monthly or annually, usage-capped, and recognized ratably. Enterprise customers often sign annual or multi-year minimums, but revenue is still recognized when tokens are used, so a “contract” is not a separate GAAP bucket. Cloud service providers (CSPs) bill and collect, then remit net of fees, while Anthropic keeps the end-customer contract and books the sale gross.

For the six months ended June 30, 2026, revenue was $16.286 billion, up from $1.163 billion a year earlier. Consumption was $13.064 billion (80%) and subscriptions $3.222 billion (20%). Full-year 2025 was $4.585 billion, of which consumption was about $3.80 billion (83%) and subscriptions $789 million (17%). [1] [2] [3]

S-1 line	FY2024	FY2025	H1 2025	H1 2026	Share of H1 2026	Confidence
Total revenue	$386M	$4,585M	$1,163M	$16,286M	100%	High — filing, multiple outlets
Consumption-based	$288M (75%)	~$3.80B (83%)	$983M	$13,064M	80%	High
Subscription-based	$98M (25%)	$789M (17%)	$180M	$3,222M	20%	High
Through CSP marketplaces	~32%	~47% ($2.16B via Amazon + Google)	~44%	~48% (~$7.8B)	48%	High on share; medium on the $7.8B (48% × $16.286B)
Direct / non-marketplace (residual)	—	—	—	~52% (~$8.5B)	52%	Medium — arithmetic, not a labeled line
US billing address	—	—	$820M	$11,763M	72%	High
International billing address	—	—	$343M	$4,523M	28%	High

Sources: Form S-1 filed October 26, 2026, as reported by FourWeekMBA, Kepler, Reuters (draft), Blockspace, and TokenPost. [1] [4] [5] [6] [7]

Quarterly recognized revenue, compiled from the same filing coverage: Q1 2025 $376 million, Q2 $787 million, Q3 $1.382 billion, Q4 $2.040 billion, Q1 2026 $4.731 billion, Q2 2026 $11.555 billion. [8] [4] Q3 2026 was disclosed to investors at about $16.8 billion (Bloomberg, Financial Times, Reuters), up roughly 45% sequentially. The S-1 instead states Q3 was 86 times Q4 2024; one reconstruction using a $197 million Q4 2024 base gets about $16.9 billion. Treat $16.8–16.9 billion as the same quarter reported two ways, not two different numbers. [9] [10]

What this means if you are mapping channels: Bedrock, Vertex, and Microsoft Foundry/Azure are a distribution overlay on consumption, not a product segment. Direct API, Claude Code token burn, and enterprise overages are also consumption. Pro, Max, Team, and at least some Enterprise seat fees are the subscription line. You cannot add “API + enterprise + Claude Code + consumer” and expect 100% without double-counting.

Where the requested categories diverge from the S-1

The prospectus has one segment, coordinated by the co-founders as CODM. It does not report API versus applications, direct versus Bedrock versus Vertex versus Azure, or Pro versus Max versus Team versus Enterprise. [3] [11]

The mismatches that matter:

• API (direct and via clouds) is not a line. Marketplace sales were about 11% of revenue in 2023, 32% in 2024, 47% in 2025, and 48% in H1 2026. The 2025 $2.16 billion figure Reuters saw in the confidential draft was Amazon and Google only. The public H1 2026 48% includes Amazon, Google, and Microsoft, with no provider split. [1] [6] [7] Direct access is the residual ~52%, but that residual mixes the first-party API, claude.ai, Claude Code, Cowork, and direct enterprise—not “direct API” alone.
• Enterprise contracts overlap consumption. Minimum-spend deals convert usage into a commitment, then revenue is still recognized on tokens. The closest disclosed proxy is customer size, not contract type: customers with $100,000 or more of trailing-twelve-month revenue produced $12.0 billion, about 74% of H1 2026 revenue (more than 6,000 such customers at June 30; more than 11,900 by September 30). [2] [12]
• Consumer subscriptions are inside the 20%, not equal to it. Subscription revenue includes Claude Pro and Max plus Team and Enterprise packages. The filing does not isolate individuals. Net revenue retention of about 900% as of September 30, 2026 explicitly excludes Pro and Max, because fixed fees do not expand the way usage does. Consumer is therefore less than 20% of total revenue; how much less is not disclosed. Secondary write-ups disagree on whether “much” of the $3.2 billion is Pro/Max or whether individuals are a small slice of that 20%. [1] [13] [14]
• Claude Code is a usage multiplier, not a segment. The S-1 calls it the product that “most concretely defined what autonomous AI work looks like,” ties ~900% net revenue retention to agentic products like Claude Code, and gives a spend gap rather than a dollar total: in the quarter ended September 30, 2026, Claude Enterprise users who did not use Claude Code spent less than $75 a month on average, while the top 5% of Claude Enterprise users who did use it spent about 20 times that. Over half of $100,000-plus customers used Claude Platform plus at least one first-party app (Claude Code or Cowork) in Q3 2026, up from about 30% in Q4 2025. [1] [2] [15]

Gross accounting also inflates the top line relative to a net reseller model. Platforms bill the end customer and remit net of “platform and other fees.” Reuters’ analysis of the 2025 draft put distribution fees above $350 million on $2.16 billion of Amazon and Google marketplace sales, roughly 16 cents per dollar—well above published marketplace listing rates of 1.5–3%, because the fee bucket is broader than a store commission. [6] [5]

Estimated split into the categories people actually want

These shares are reconstructions. Anthropic has not published them. They do not add cleanly to the S-1 lines, because Claude Code and enterprise usage are mostly consumption, and Bedrock/Vertex/Azure are a channel through that consumption.

Category asked for	Best available figure	Period	What it actually measures	Confidence
API + enterprise usage (combined)	~80% of revenue; $13.1B consumption in H1 2026	H1 2026	S-1 consumption line. Includes direct API, cloud-marketplace API, Claude Code tokens, and enterprise usage. Not “API only.”	High as a pricing split; low as a product split
Cloud marketplaces (Bedrock + Vertex + Azure/Foundry, combined)	~48%, ~$7.8B	H1 2026	Channel, mostly consumption. No AWS vs Google vs Microsoft split.	High on combined share
Amazon + Google marketplaces only	$2.16B, 47%	FY2025	Confidential draft, Reuters. Microsoft not in this pair.	High for 2025; not comparable 1:1 to the 2026 three-cloud figure
Direct / non-marketplace	~52%, ~$8.5B	H1 2026	Residual. Includes direct API and first-party apps.	Medium
Subscriptions (Pro, Max, Team, and enterprise seats)	$3.222B, 20%	H1 2026	Upper bound on consumer + Team. Enterprise seats are inside this number.	High on the total; low on the consumer slice
Consumer Pro/Max alone	Not disclosed. Must be <20% of total. One synthesized estimate puts Pro/Max at ~8% and Team at ~2%.	2026	Value Add VC, explicitly not an audited segment. An earlier chart from the same outlet had consumer nearer 22%—do not use that.	Low
Enterprise as large accounts, not “contracts”	$12.0B, ~74%	H1 2026	Customers with ≥$100k TTM revenue. Overlaps API, Claude Code, and seats.	High as a size cut; not a contract cut
Claude Code	Company: >$0.5B run-rate (Sept 2025), $1B (Dec 2025), >$2.5B (Feb 12, 2026), enterprise > half of Claude Code revenue	Company disclosures	Run-rate, not GAAP revenue. Feb 2026 figure was ~18% of the company’s then ~$14B run rate.	High on company dates through Feb 2026
Claude Code	~$8B annualized	May 2026	Sacra / FutureSearch, alongside company ARR of $47B (~17%). Not in the S-1.	Medium
Claude Code	$15.12B tracked ARR; ~21.9% of tracked company ARR	Week ending Aug 10, 2026	TickerTrends alt-data. Anthropic has not confirmed. Growth had slowed to ~5% over four weeks.	Low–medium
Synthesized channel mix	API 45%, enterprise contracts 35%, Claude Code 10%, Pro/Max 8%, Team 2%; clouds “included in API”	2026 estimate, updated Oct 26	Value Add VC. Conflicts with the $8–15B Claude Code run-rate estimates above, which imply a higher Code share of ARR than 10% of recognized revenue.	Low

Company Claude Code milestones: general availability in May 2025; over $500 million run-rate by the September 2, 2025 Series F announcement, with usage up more than 10x in three months; $1 billion by December 3, 2025; over $2.5 billion on February 12, 2026, more than doubled since January 1, with weekly active users doubled and business subscriptions up 4x, and enterprise more than half of Claude Code revenue. [16] [17] May and August figures are third-party. [18] [19] [20]

The mechanism is straightforward and is why a seat count understates the business. A chat user on a capped plan produces a fixed fee. An agent that reads a repo, edits files, runs tests, and retries burns orders of magnitude more output tokens—the expensive side of Anthropic’s price card. That is why the S-1 can show Enterprise users who skip Claude Code at under $75 a month while the heaviest Code users inside Enterprise spend about 20 times that, and why net revenue retention, excluding Pro and Max, was about 900%. [1] [21]

Company-level run rate, for scale: about $9 billion at the end of 2025, $14 billion in February 2026, over $47 billion at the late-May Series H, $65 billion at the end of July (figures shared with investors; CNBC confirmed the July figure), with SemiAnalysis around $74 billion by mid-October and Sacra around $70 billion in October after Q3 growth slowed. Recognized H1 revenue of $16.3 billion is not the same object as a 28-day consumption run rate. [18] [17] [10]

Customer concentration: Cursor, Copilot, and the top five

Concentration was real in 2025 and has diluted in share even as the largest accounts kept setting revenue records.

Disclosed, unnamed:

• Top two customers were an aggregate of about 24% of 2025 revenue. Reuters, which saw the confidential draft, reported each at about 12%. On $4.585 billion, that is roughly $550 million each, about $1.1 billion combined. [6] [22]
• For the nine months ended September 30, 2026, no single customer was more than 10%, and the top five were about 25%. Outside the top five: about 64% of Q1 2026 revenue and about 79% of Q3, so the top five fell from about 36% of the quarter to about 21%, even though revenue from each of the top five hit a record in Q3. [23] [9]
• A draft circulated before the public filing, reported by the Wall Street Journal, put about one-fifth of H1 2026 revenue in two customers. The public nine-month language is the one to use for the later period. [24] [25]
• Many large customers are not on long-term locks and can cut usage immediately. That is the risk factor the filing emphasizes, not a named logo. [23]

Named only by inference, not by the S-1:

• In August 2025, VentureBeat reported that Cursor and GitHub Copilot drove about $1.2 billion of a roughly $4–5 billion run rate, close to a quarter. That lines up in magnitude with the later 24% / two-customer 2025 disclosure, but a mid-year run rate is not full-year GAAP, and Anthropic has never confirmed the names. [26] [27]
• Motley Fool and others have speculated that the H1 2026 pair is SpaceX (which acquired Cursor for about $60 billion in August 2026) and Microsoft via GitHub Copilot. That is industry inference, not a filing fact. A claim that the official prospectus identifies the two customers as Amazon and Google confuses the marketplace channel (47–48% of sales, with Anthropic still contracting the end customer) with the 10% customer test. Reuters treated those as separate disclosures. Do not treat Amazon and Google as the named 12% customers. [28] [6] [11]
• Sacra’s post-filing note says top customers including Meta and Microsoft cut internal Claude Code usage in Q3 (about 50% and 35%, respectively) as part of a broader slowdown. That is Sacra’s reporting, not an S-1 table. [18]

Other concentration the filing does quantify: government contracts are about 1% of revenue. About 5–7% of revenue in the first nine months of 2026 came from usage that violated policy, including model distillation, and Anthropic warned that cracking down could reduce revenue. [29] [30] [31]

The base under the top five has widened fast. $100,000-plus customers went from about 20 at the end of 2023 (about 90% of revenue) to about 200 in 2024, about 1,500 in 2025 (about 76% of revenue, $3.5 billion), more than 6,000 at June 30, 2026 ($12.0 billion, 74% of H1), and more than 11,900 at September 30, including more than 1,900 at $1 million-plus and about 250 at $10 million-plus. The long tail is about 40 million customers each under $100,000. The $100,000-plus group includes every Fortune 10 company and more than half the Fortune 500, but Fortune 500 spend in that tier averaged less than 0.1% of those companies’ total expenses in Q3—penetration, not saturation. [32] [2] [14]

Claude Code’s growth, and why it is not a clean revenue line

Claude Code is the clearest product-level ramp Anthropic has disclosed, and then stopped disclosing in dollars.

Date	Claude Code figure	Source	Confidence
May 2025	General availability	Anthropic	High
Sept 2, 2025	>$500M run-rate; usage >10x in three months	Anthropic	High
Dec 3, 2025	$1B run-rate, ~6 months after GA	Anthropic	High
Feb 12, 2026	>$2.5B run-rate; WAUs doubled since Jan 1; business subs 4x; enterprise >50% of Code revenue	Anthropic	High
May 2026	~$8B annualized	Sacra, FutureSearch	Medium
Aug 10, 2026	$15.12B tracked ARR; ~5% growth vs four weeks earlier; ~21.9% of tracked company ARR	TickerTrends	Low–medium
May 2027	FutureSearch p50 $18.5B (p10 $9.8B, p90 $33B)	Forecast, not a disclosure	Low

The strategic point is that Code competes with Anthropic’s own largest historical customers. Cursor and Copilot made Claude the default model inside someone else’s product and, on the 2025 evidence, a quarter of the company. Claude Code takes the same workload direct: the developer (or the enterprise that standardizes on it) pays Anthropic for the agent, not only for tokens resold through a third-party IDE. That is why Code can both add revenue and pressure the Cursor/Copilot share at the same time. SpaceX’s August 2026 acquisition of Cursor, and Anthropic’s public commitment to keep serving Claude inside Cursor after OpenAI moved to cut Cursor off, shows the indirect channel is still strategically important even if its share of a much larger base has fallen below the 10% naming threshold. [33] [28]

Code is also the wedge into the rest of the suite. Cowork, the knowledge-work agent previewed in January 2026, had by September 30 outpaced Claude Code’s adoption at the same point after Code’s February 2025 preview, and Anthropic began merging Cowork into Claude.ai in September 2026. Multi-product attach among $100,000-plus customers rose from about 30% to over 50% in three quarters. The S-1’s growth story is not “more chatbot seats.” It is “a coding agent creates a heavy user, who then pulls platform spend and a second app.” [1] [15]

How to use this if you are competing or underwriting

Do not build a model that treats API, enterprise, and Claude Code as additive S-1 segments. Start with 80/20 consumption versus subscription, haircut marketplace revenue for fees if you care about net dollars (2025 implied about 16% of Amazon/Google marketplace sales, and that rate is not disclosed for 2026), and treat Claude Code as a driver inside consumption whose dollar share is somewhere between the company’s last official print (>$2.5 billion run-rate in February, then about 18% of company run rate) and alt-data around 20% of tracked ARR by August—not a precise 10% or 17%.

The concentration risk has changed shape. In 2025, two customers were 24%, and the leading external hypothesis is still Cursor and GitHub Copilot. By the first nine months of 2026, no one customer was above 10% and the top five were 25%, with their quarterly share falling from about 36% to about 21%. That is diversification by explosion of the denominator, not proof those logos got smaller in dollars—the filing says each of the top five set a Q3 revenue record. The binding risk is usage-based contracts with no lock-in against multi-year compute commitments, not a single IDE vendor at 12%.

For a competitor, the opening is not the consumer plan. Pro and Max are excluded from the retention metric precisely because they do not expand, and they are a subset of a 20% subscription line. The opening is either (1) being the default model inside someone else’s agent, which is how Cursor and Copilot became outsized in 2025, or (2) owning the agent surface so token burn accrues directly, which is what Claude Code did from a standing start in May 2025 to a multi-billion run rate by early 2026. Cloud marketplaces are a distribution tax and a distribution advantage at once: nearly half of H1 2026 revenue, no disclosed split among AWS, Google, and Microsoft, and the same three companies are also compute suppliers.

---

Recent Findings Supplement (October 2026)

Anthropic filed its S-1 on October 26, 2026, providing the first detailed public financials (all data below from or directly derived from this filing and contemporaneous analyses unless noted). H1 2026 revenue reached $16.286 billion (up ~14x from $1.163 billion in H1 2025), with Q2 alone at $11.555 billion. The filing segments primarily by consumption-based (usage/API tokens, 80% or $13.064 billion) versus subscription-based (20% or $3.222 billion), plus channel (cloud marketplaces ~48%). It does not break out the exact requested categories (direct API vs. Bedrock/Vertex, specific consumer plans, or Claude Code standalone).[1][2]

Claude Code (Anthropic’s agentic coding product) is a major growth and expansion driver within consumption revenue. Enterprise users of Claude Code spend ~20x more on average than non-users (<$75/month vs. ~$1,500/month for top 5% of Claude Enterprise users). This usage fuels the ~900% net revenue retention (NRR) as of September 30, 2026 (excludes flat-fee Pro/Max subscribers). Third-party estimates place Claude Code run-rate revenue at >$2.5 billion by February 2026 and ~$8 billion annualized by May 2026.[1][3]

Customer concentration is notable but diversifying, with heavy reliance on large usage-based accounts including coding platforms. More than 11,900 customers generated ≥$100k TTM revenue as of September 30, 2026 (up from >950 a year earlier); these accounted for ~74% of H1 2026 revenue ($12 billion). Top 5 customers represented ~25% of revenue over the nine months to September 30, 2026 (no single customer >10%). Reports indicate two customers (widely speculated as Cursor and GitHub Copilot) accounted for ~20% ($3.26 billion) of H1 2026 revenue. Cursor reached ~$4 billion annualized revenue by June 2026; GitHub Copilot had ~4.7 million paid users. Over 50% of $100k+ customers now use both the Claude Platform and at least one first-party app (up from ~30% in Q4 2025).[4][5]

Cloud marketplaces (AWS Bedrock, Google Vertex AI, Microsoft Azure) handled ~48% of H1 2026 revenue (~$7.8 billion), up from ~11% in 2023 and ~47% in 2025. Anthropic recognizes this gross (as principal, with direct customer contracts) while CSPs handle billing/collection and take fees. Direct sales are the remaining ~52%. International billing was 28% of H1 revenue.[6][1]

S-1 segments differ materially from requested categories. The filing uses only consumption (primarily API/usage, including via Claude Code and other agents) versus subscription, plus the marketplace channel. It provides no product-level splits (e.g., exact Claude Code contribution, Pro/Max/Team/Enterprise within subscriptions, or direct API vs. Bedrock/Vertex within consumption or marketplaces). Consumer subscriptions form an unspecified portion of the 20% subscription bucket.[7]

Table of estimates (H1 2026 or latest available; all post-April 2026 data prioritized):

• API (direct + via AWS Bedrock/Google Vertex/Microsoft Azure): ~$13.1B consumption-based (80% of total); of which marketplaces ~48% of total revenue (~$7.8B). Direct ~52%. Source: S-1; confidence high (explicit aggregates).[1]
• Enterprise contracts: Mix of consumption (majority) and subscriptions; >11,900 $100k+ customers drove ~74% of H1 revenue. Top 5 ~25%. Source: S-1; confidence high.[4]
• Claude consumer subscriptions (Pro, Max, Team): Unspecified portion of $3.2B subscription-based (20%). Pro/Max excluded from NRR metric. Source: S-1 inferences; confidence medium-low (no split).[1]
• Coding products (Claude Code): Key driver of consumption/expansion; users spend ~20x more; third-party est. ~$8B ARR (May 2026). Source: S-1 context + Sacra analysis; confidence medium (growth/impact explicit, dollar est. third-party).[3]
• Customer concentration (e.g., Cursor/GitHub Copilot reliance): Top 2 plausibly ~20% ($3.26B) of H1; Cursor ~$4B ARR (June 2026 est.); Copilot ~4.7M paid users. Top 5 ~25%. Source: S-1 + reports (Stahel/Motley Fool); confidence medium (speculation on exact names, but concentration confirmed).[5][8]

Overall run-rate context (third-party, post-S-1 analyses): Annualized revenue reached ~$65B by July 2026, slowing to ~$70B by October amid Q3 growth moderation (top customers reportedly reduced Claude Code usage). Claude Code and agents drove much of the prior hyper-growth and efficiency gains (compute fell to 48% of revenue; first GAAP operating profit in Q2).[3]

These figures reflect explosive but usage-dependent growth with channel and customer risks (e.g., marketplace dependence, variable consumption). No finer public breakdowns exist beyond the S-1 aggregates as of late October 2026.