← Back to blog
AiAbout 6 min read

Anthropic's Filing Is the First Honest Look at What AI Inference Costs

Published Nov 14, 2026
Anthropic's Filing Is the First Honest Look at What AI Inference Costs

Every argument about AI image and video pricing eventually runs into the same missing number. Nobody outside the labs knows what a generation costs to serve. Vendors publish per-second and per-image rates, competitors undercut each other by 30% in a week, and the entire debate happens without a single audited cost figure underneath it.

Anthropic just published one, indirectly. Its amended S-1 filed with the SEC on November 10, and the roadshow that began on November 12, put a real income statement next to a real revenue curve for the first time. The company sells text models, not image models. But the shape of its economics is the shape every inference-heavy product inherits, and image and video generation is the most inference-heavy category in the market.

The terms, briefly

Anthropic is offering 825 million Class A shares at $70 to $78. About 801 million come from the company, roughly 23.6 million from existing holders. Ticker ANT, Nasdaq Global Select Market. On the 13.09 billion shares the filing says will be outstanding afterwards, that range implies somewhere between $916 billion and $1.02 trillion on a basic count; on a fully diluted basis the top end reaches about $1.4 trillion. Pricing is expected November 18, with trading on November 19, subject to regulatory clearance. Up to 20% of the offering is earmarked for retail. Twenty-two banks are on the cover, Goldman Sachs and Morgan Stanley among the leads.

The gross figure is around $64 billion. Add a reported $26 billion private placement with chip and cloud partners and the total lands near $90 billion. It is one of the largest listings ever attempted, and it is happening while AI equities are being sold off.

The revenue is real and so is the burn

The numbers that matter are on the inside pages.

Annualised revenue run-rate hit $81 billion in October, up $10 billion from the month before. First-half 2026 revenue was $16.29 billion against $1.16 billion a year earlier. Third-quarter revenue came in at $17.0 billion, which on its own is more than the entire first half.

Against that, the first half carried a pre-tax loss of $80.75 billion, up from $12.9 billion. That figure is dominated by non-cash items, share-based compensation above all, which is why it looks absurd next to the revenue line. The more informative number is the quarterly operating result: a $567 million operating loss on $17 billion of revenue in Q3.

Read that again. Seventeen billion in quarterly revenue, five hundred and sixty-seven million in operating loss. That is roughly a 3% operating margin deficit. For a company that was burning cash at a spectacular rate eighteen months ago, it is close to the line.

The filing also discloses more than 11,900 customers paying over $100,000 a year. That is the number that explains the revenue curve. This is not a consumer subscription business with a long tail; it is a few thousand large contracts, which makes the revenue predictable and the concentration risk obvious.

That customer profile also shapes what the company can charge. Large accounts negotiate, and they negotiate on volume. A pricing page is an opening position for them, not a rate card, which is one reason the headline numbers for enterprise AI have stayed firm while the consumer-facing prices for images and video have fallen every month.

Why this matters for a two-cent video

Here is the uncomfortable arithmetic that the image and video market has been avoiding.

Anthropic's near-breakeven operating result comes from selling text tokens at prices that have held roughly steady for a year, to customers who renew. Image and video generation does not have that profile. A single 4K image or a fifteen-second clip consumes orders of magnitude more compute per request than a chat turn, the price per request is a hundred times smaller, and the customer is far more likely to be a one-off.

When ByteDance prices 720p video at ¥0.99 per second, or Google cuts 720p to $0.07 per second, those numbers come from no cost model anyone outside has seen. They are competitive postures, taken by companies with other revenue to lean on. ByteDance has advertising. Google has search. Both can run a video model at cost or below it for as long as the strategic point is worth making.

Anthropic does not have that option, which is precisely why its filing is useful. It shows what happens when an AI company has to stand on its own numbers: revenue has to grow faster than compute, gross margin has to climb, and the operating line has to approach zero. It took Anthropic roughly six quarters of hypergrowth to get within 3% of it.

The capex question nobody has answered

The roadshow is happening against a specific backdrop. AI-linked equities have been under pressure, with trend-following hedge funds posting one of their worst weeks in two years. The proximate cause is a fight over data centre construction: candidates promising tighter controls on new builds won broadly in the recent US elections, and the market is repricing the assumption that compute can be added indefinitely at a known price.

Goldman's read, as reported, is that political and regulatory friction on data centre construction may constrain supply and raise the scarcity value of existing infrastructure. That is a reasonable argument. It is also an argument that compute gets more expensive, not less, which is bad news for any product whose price is denominated in seconds of GPU time.

Anthropic's $112.4 billion in cash plus $19.3 billion in short-term investments, as of June 30, is the buffer that lets it argue it can build through the cycle. A video startup with $40 million in the bank cannot make that argument, and will keep pricing at whatever the market leader sets.

What to watch on November 19

Three things will tell you more than the first-day pop.

First, whether the offering prices at the top of the range. A $78 print implies investors accepted the diluted valuation; a $70 print with a weak book says the 20% retail allocation did the work.

Second, the customer concentration disclosure. Eleven thousand nine hundred large accounts is a healthy number until you learn how much of the $81 billion comes from the top fifty.

Third, and least discussed, what Anthropic says about inference cost per unit of revenue in the quarters after listing. If gross margin expands while prices hold, the whole industry's assumption that inference gets cheaper with scale is confirmed. If it does not, then the two-cent video is a marketing expense, and it will be repriced the moment the marketing budget moves.

None of this is a prediction about Anthropic's stock. It is a note that the image and video market has been pricing its products against a cost curve it never had to publish. Somebody finally did, and the curve is flatter than the price lists suggest.

Related articles

Google replaced the model behind Search, and the number that moved was a shell benchmark

Gemini 4 Hydrogen became the default on 12 November, taking over from Gemini 3.8 Flash across Search, the Gemini app and the developer API. It runs about 2.5 times faster, scores 57% on Terminal-Bench 4.1 where its predecessor scored 18.2%, and costs $1.50 in and $7.50 out per million tokens.

Three ways a coding agent wastes money, measured on 30 repositories

TokenEconomizer watches what an agent did in a repository you work in often, then stops it from rediscovering the same build command, loading tools it never uses and looping on one error. Across five agent and model combinations it cut mean cost by 18.4% to 30.9%, cheaper in all 30 repository cells, with Pass@1 up in four of five.

earshot gives coding agents 922 sounds, and every one is rendered from code

An agent can write a game, an interface or a film and never hear any of it. earshot answers requests like "three variations of a footstep on gravel" from a catalogue of 922 designs built from public-domain recordings, with no audio files in the repository at all. It runs on CPU, as a CLI, a Python package, an MCP server and a Claude Code plugin.

The agents stealing API keys were not sophisticated. The gateways were open.

Amazon's decoy servers spent October posing as inference APIs and logged agents built on ARTEX pulling provider keys out of LiteLLM, Dify, Langflow and Open WebUI. The way in was a default master key and a half-finished install. ARTEX itself ships with budgets, scope gates and evidence for every finding.