Video summary
Anthropic's IPO Filing Shows What Your Claude Plan Really Costs
Main summary
Key takeaways
Key Takeaway
Anthropic’s filing suggests that serving paying customers is currently profitable on average, while total company spending slightly exceeds revenue because of model development and other costs. The presenter’s unusually heavy Claude usage likely costs Anthropic more than his $100 monthly subscription, though the estimates are highly uncertain. The likely pressure point for heavy users is tighter usage limits, not necessarily higher plan prices.
What the Filing Says
- Anthropic reported a $104 billion net loss for the first nine months of the year. About $94 billion was non-cash, mainly accounting losses as convertible notes and similar investor arrangements increased in value with Anthropic’s valuation.
- Excluding those paper charges, the company had an operating loss of about $1.25 billion. It generated $6.6 billion more cash than it used over the period.
- Quarterly results varied: Anthropic reported an operating profit of nearly $1 billion in Q2, followed by an operating loss of about $500 million in Q3. The filing attributed the Q3 result in part to a $2 billion non-cash charge and $5 billion of short-term compute, which cost more than long-term contracts.
- Cost of revenue, mostly compute used to serve paying customers, was $4.8 billion against $33 billion of revenue over nine months. That is roughly 14.5 cents per revenue dollar, or an 85.5% gross margin. Q3 margin was about 86%, despite full-price SpaceX capacity and expensive short-term compute.
- Serving costs have fallen sharply: nearly $2 per dollar of revenue in 2024, about $0.60 last year, and under $0.15 currently.
- Research and development cost $19 billion—about four times serving costs—and sales and marketing cost $6.6 billion. The latter includes compute for free users. Overall, Anthropic spent about $1.04 for each dollar of revenue and had raised $141 billion.
- Subscriptions accounted for about one-fifth of revenue, or roughly $7 billion, including Team and Enterprise as well as Pro and Max. The filing does not separate Pro and Max. The presenter’s rough inference is that subscriptions are profitable in aggregate: even if all $4.8 billion of serving costs were assigned to subscribers and API serving cost nothing, subscription revenue would still exceed those costs.
Presenter’s Estimate of His Claude Plan Economics
The presenter pays $100 per month for Max and says his last 30 days of usage would cost about $3,550 at API list prices—roughly 35 times the subscription fee.
His estimated cost to Anthropic for that month:
- Low case: about $175
- Middle case: about $400, with a stated range of roughly $300–$500
- High case: about $1,200
These are the presenter’s estimates, not Anthropic’s disclosed per-user costs. The high case reflects his long sessions: on average, Claude revisited about 200,000 tokens of context for each token it generated, compared with about 5,000 in the DeepSeek figures he cites. Anthropic’s actual cost impact from long contexts is unknown.
Cost Methodology and Inference
- The presenter found that 96% of his tokens were cache reads, billed at one-twentieth of the normal input price.
- A prompt’s initial processing is prefill; generating a response token by token is decode. Decode is slower and helps explain why output tokens are priced at five times input tokens. Cache reads reuse previously processed context, avoiding that prefill work.
- His cost scenarios use different assumptions: an efficient open model running on newer NVIDIA hardware for the low case; Anthropic’s average serving economics and an analyst estimate for the middle case; and a higher cost allowance for long-context sessions in the high case.
- As a comparison—not a direct estimate of Anthropic’s costs—the presenter cites DeepSeek’s published serving figures: at an assumed $2 per GPU-hour, GPU rental cost about $87,000 for a day. Billing that day’s tokens at DeepSeek R1 prices would have produced about $560,000, implying compute costs of roughly 15.5% of the billed amount.
- SemiAnalysis reportedly estimated that fully using a $100 plan could deliver about $5,700 of API value. According to the video, Anthropic said its 2025 weekly limits would affect less than 5% of subscribers.
Plans, Risks, and Recommendations
The presenter says the typical subscriber uses far less than he does, helping offset heavy users. His conclusion is that plans can be profitable on average, even if some users receive more value than their fee.
He describes limits changing over time while headline plan prices remained stable: Pro at $20, and Max at $100 or $200. Examples include weekly limits added in 2025, reported faster limit exhaustion during weekday peak hours, a brief Pro-plan change affecting about 2% of new sign-ups, limits that doubled in May after the SpaceX deal, and a later temporary boost ending with weekly limits 17% below the boosted level but 25% above the earlier level.
The filing reportedly warns that Anthropic may change usage limits again and that margins can be squeezed as new compute capacity comes online. The presenter also notes that the company was short on compute and buying costly short-term capacity.
- Typical users: The presenter says plans are likely a fair deal and does not expect an imminent change that matters to users who rarely approach their limits.
- Light users: Compare the subscription with paying for API usage directly; API billing may cost less.
- Heavy users: Expect the gap between usage and subscription price to narrow. Avoid building workflows that depend on today’s caps.
The presenter cautions against buying a $20,000 local machine on the assumption that cloud-plan subsidies are about to disappear. His rationale is that inference costs keep falling: Anthropic’s Opus cache-read price reportedly fell from $1.50 to $0.50 to $0.20 per million tokens, while industry prices for comparable quality have been falling roughly tenfold per year.
His expectation is that tightening will show up first in usage limits, rather than plan prices.
Companies, Assets, and Instruments Mentioned
- Anthropic / Claude: Pro and Max subscriptions, API, Team and Enterprise plans; no stock ticker is given.
- Convertible notes and similar investor arrangements.
- Amazon and Google: Identified as investors holding such arrangements.
- SpaceX: Compute capacity discussed as temporarily offered at a reduced fee during ramp-up; the amount of the discount was not disclosed in the video.
- DeepSeek / R1, NVIDIA GPUs, and AI-compute infrastructure.
- GitHub Copilot and Cursor: Cited as examples of other services moving toward usage-based billing.
Presenter, Sources, and Disclosure
Presenter: Vasko
Sources mentioned: Anthropic’s IPO filing; Anthropic API pricing; DeepSeek’s published inference-system figures; SemiAnalysis plan-value testing.
The presenter says the video is about subscription-plan economics, not the stock, and gives no recommendation to buy or sell shares.
Rate this summary
Your feedback will help improve summaries.
Improve this summary
Reprocess with a stronger model when the summary feels incomplete or inaccurate.
Translate summary in another language
Ask questions to this video
Chat for follow-up questions, clarifications, and source-backed answers.