What your AI vendor's revenue actually tells you

A vendor’s revenue headline is not information you can act on. Anthropic’s annualized revenue run rate surpassed $65 billion at the end of July 2026, up from $47 billion in May and $9 billion at the end of 2025, and investors expect the year to close between $100 billion and $120 billion [8]. That is a real number about a real business. It tells you almost nothing about whether the model you pinned in March will still answer requests in December, or what you will pay for it. The facts that decide those things sit on five boring pages: the pricing page, the deprecation table, the commercial terms, the consumer terms, and the privacy setting on your account.

This guide is for the solo operator or small team paying for Claude, ChatGPT or Gemini out of the business account and building something on top of it. It is not for procurement teams working from a negotiated master agreement, who have their own notice periods and their own lawyers, and it is not advice about buying shares in anyone. The question here is narrower and more useful. Your vendor is growing fast and reorganising itself around that growth. What follows is which of its commitments to you are actually written down, how much warning you get when they change, and what to do about the parts nobody has committed to.

Run rate is a projection, not a bank balance

Annualized run rate is a recent month or quarter multiplied out to twelve months. It is a forecast dressed as a fact, and it moves in big steps because the underlying period is short. Anthropic added about $18 billion of run rate between May and the end of July 2026 [8]. OpenAI doubled its revenue to $40 billion over a longer stretch, from $20 billion at the end of 2025 [8]. Both figures are revenue. Neither is profit, and neither company has to publish one while it is private.

The reason to notice the numbers at all is the stage they signal. Anthropic was last valued at $965 billion in late May 2026, when it raised a $65 billion round, and both it and OpenAI have filed confidential IPO paperwork, with Anthropic expected to reach the public markets ahead of OpenAI, possibly as soon as this autumn, seeking a valuation of $2 trillion or more [8]. A company in that position starts optimising for measures that will appear in quarterly filings: retention, gross margin, revenue per customer. Those pressures do not usually reach you as an outage. They reach you as a repackaged plan, a changed limit, a model you liked being retired in favour of one with better unit economics. None of that is sinister. It is what the incentive structure produces, and it is the part you can plan for.

The only promises that bind a vendor are on the legal pages

Roadmap talk is not a commitment. Notice periods are. Anthropic’s commercial terms say the company “may update the published rates, to be effective the earlier of 30 days after the updates are posted by Anthropic or Customer otherwise receives Notice” [3]. The same terms let either party terminate for convenience and require Anthropic to give 30 days’ prior notice [3]. So on the API, your worst case on price is roughly a month of warning, and your worst case on the relationship itself is the same.

The consumer side runs on its own document, and it splits in an important place. Anthropic’s consumer terms say fees will not change during your current initial or renewal term, and that for an increase “we will inform you at least 30 days in advance of the change”, with cancelling before the next renewal as your way out [9]. The same document reserves the right to “modify, suspend, or discontinue the Services or your access to the Services, in whole or in part, at any time without notice to you”, tempered by a promise to strive for reasonable advance notice when a service stops [9]. Read those two clauses together. The price you pay carries a clock. What you get for the money does not.

Model retirement runs on a third clock. Anthropic says it gives “at least 60 days’ notice before model retirement for publicly released models” [1]. OpenAI’s published minimums are longer and more graded: at least 6 months for generally available models, at least 3 months for specialised chat, Codex and deep research variants, and as little as 2 weeks for preview models [6]. Google takes a third approach on the Gemini API, publishing shutdown dates that “indicate the earliest possible dates on which a model might be retired” and promising to communicate the exact date with advance notice [7]. Three vendors, three different promises, none of them the same as the others.

Write those numbers down once, per vendor, on a single line each: days of notice on price, days of notice on model retirement, days of notice on termination. That line is your actual planning horizon. Everything else is atmosphere.

Model retirement is the deadline that lands on your calendar

Retirement is the one vendor decision that will definitely break something you built, on a date you can already look up. Anthropic separates two states. A deprecated model is “still functional but no longer recommended” and carries an assigned retirement date, while a retired model “is no longer available for use” and requests to it “will fail” [1]. Nothing degrades gracefully. It stops.

The table is specific. claude-3-7-sonnet-20250219 was retired on 19 February 2026, a year to the day after the date stamp in its name, and claude-3-5-haiku-20241022 went the same day [1]. claude-sonnet-4-20250514 and claude-opus-4-20250514 were retired on 15 June 2026, claude-opus-4-1-20250805 on 5 August 2026, and claude-3-haiku-20240307 on 20 April 2026 [1]. Across those entries the shortest life a dated snapshot got was twelve months from its own date stamp and the longest was just over two years. Nobody promises twelve months, so treat it as the floor you plan against, then check the table rather than trusting the floor.

Currently active Claude models carry forward-looking earliest dates instead of retirement dates: claude-haiku-4-5-20251001 is listed as retiring not sooner than 15 October 2026, claude-sonnet-5 not sooner than 30 June 2027, claude-opus-5 not sooner than 24 July 2027, and claude-fable-5-1 not sooner than 1 September 2027 [1]. OpenAI publishes firm shutdown dates instead, including gpt-3.5-turbo-0125, gpt-4-0613, gpt-4-turbo, o1 and o3-mini on 23 October 2026, and gpt-5-2025-08-07 and o3-2025-04-16 on 11 December 2026 [6]. Google’s dates mean a third thing: gemini-2.0-flash is listed with a shutdown date of 1 June 2026 and gemini-3.6-flash as its replacement, but the page is explicit that such dates are the earliest a model might go, not the day it will [7].

The work is small and nobody does it. Anthropic tells you how to find your own exposure: open the Usage page in the Claude Console, click Export, and read the downloaded CSV to see usage broken down by API key and model [1]. Put each retirement date in the calendar with a reminder 90 days ahead, which is more than the 60 days Anthropic guarantees and less than the 6 months OpenAI guarantees on generally available models [1][6]. Then you get to choose your migration week instead of having it chosen for you at 2am.

Falling token prices are not falling bills

The direction of headline prices has been down, not up, which surprises people who expect a growing vendor to squeeze. Claude Sonnet 5 is $2 per million input tokens and $10 per million output, against $3 and $15 for Sonnet 4.6 [2]. Claude Opus 5 is $5 and $25, where Opus 4.1 was $15 and $75 [2]. Claude Haiku 4.5 is $1 and $5, and Claude Fable 5.1 sits at the top of the range at $10 and $50 [2]. Batch processing takes 50% off both sides [2]. Prompt caching is cheaper still: a 5-minute cache write costs 1.25x base input and a 1-hour write 2x, while cache reads cost 0.1x base input, or 0.025x on Fable 5.1 and Mythos 5.1 [2]. There are modifiers in the other direction too, including a 1.1x multiplier on Claude 4.6 and later models when you pin inference to the US, and $10 per 1,000 server-side web searches [2].

Per-token prices are still not what you pay. You pay the rate times the number of tokens the job actually consumes, so a model with a lower rate that writes more tokens to reach the same answer can cost you more than the one it replaced. The same gap exists on subscriptions. Claude Pro is $20 a month billed monthly or $17 billed annually, Max starts at $100 a month at 5x or 20x Pro’s usage, and Team seats are $25 a month billed monthly with premium seats at $125, all governed by limits that reset on a rolling five-hour session window plus weekly caps on paid tiers, rather than by a fixed message count [4]. So the honest unit of measurement is cost per completed piece of work per month, recorded somewhere you will look again. Record it before a price change, not after, or you have no baseline to compare against.

Concentration is the part you control

You cannot negotiate a solo operator’s way into better notice periods. You can decide how much of your business sits behind one vendor’s login. The cheapest form of insurance is keeping the assets that took you months to build outside the product that generates them: prompts, evaluation cases and expected outputs in your own repository, in plain files, not saved inside a vendor’s prompt library where they are one account change away from being awkward to retrieve.

The second piece is a rehearsal, not a plan. Once a quarter, take one real job you ran last month and run it on a second vendor’s current model. Record the price and where the output got worse. That gives you a real switching estimate instead of a hopeful one, and it is the only thing that turns “we could move to Gemini” into a sentence with a number behind it. The third piece is the settings nobody checks. On consumer Claude accounts, your chats are used to improve the models only if you choose to allow it, and Incognito chats are not used even when that setting is on [5]. Thumbs up or down feedback works differently: the entire related conversation is stored for up to 5 years, de-linked from your user ID before Anthropic uses it [5]. Commercial products, including Claude for Work and the API, are covered by separate documentation [5]. Confirm which regime each of your accounts is under, because the answer differs between the browser tab you use for drafting and the key your code calls.

What still goes wrong

Notice periods only cover the changes a vendor is willing to name, and the commitments above name three things: rates, termination, and retirement dates. None of them is a promise that a model’s behaviour holds steady between the day you tested it and the day it retires, and that failure mode is the quiet one. A replacement with a similar name follows your prompts slightly differently, your outputs get slightly worse, and no email arrives. Anthropic has committed to long-term preservation of model weights [1], which is a commitment about keeping the past available, not about what the endpoint you call today does next month. Saved test cases are the only thing that catches the difference, and running them is a chore you will skip.

The commitments are also thinner than they read. Thirty days’ notice on a price change is a notice requirement, not a price cap [3]. On the consumer side the 30-day clock attaches to fees, while usage limits live inside a service the same terms let Anthropic modify, suspend or discontinue at any time without notice [9], which is why the five-hour window and the Max multipliers are the parts most likely to move under you without warning. And the revenue reporting that starts these conversations is selective by construction. Run-rate figures are revenue rather than profit, private companies publish them at moments of their choosing, and a number disclosed weeks before a listing is a number chosen to be disclosed [8].

Finally, the fallback plan costs hours you may not have. Testing a second vendor every quarter, keeping evaluation cases current and re-pricing your workload is worth budgeting half a day for, each time. If your entire AI spend is $40 a month, that is not worth it, and the correct response to all of the above is to pin nothing, use the current default model in the browser, and accept that you will occasionally be surprised. The work in this guide starts paying somewhere north of a few hundred dollars a month, or the first time a customer-facing feature depends on a model ID.