Anthropic is growing faster than almost anyone else in the industry: its annualized revenue run rate exceeded 65 billion dollars by the end of July, versus roughly 9 billion dollars at the end of 2025. At the same time, the company’s most expensive model sold poorly, while customers are increasingly shifting routine work to simpler models. We examine how the pricing works and why the price list no longer answers the question, “How much will the work cost?”
Price list
| Model | Input, per million tokens | Output, per million tokens |
|---|---|---|
| Claude Opus 5 | 5 $ | 25 $ |
| Claude Sonnet 5 | 2 $ (3 $ as of September 1) | 10 $ (15 $ as of September 1) |
| Claude Fable 5 | 10 $ | 50 $ |
| GPT-5.6 Sol | 5 $ (4 promotional price) | 30 $ (20 promotional price) |
| GPT-5.6 Terra | 2 $ | 12 $ |
| Grok 4.6 | 2 $ | 6 $ |
| Gemini 3.7 Flash | 0,75 $ | 3,75 $ |
Some prices have an expiration date. The price of Sonnet 5 stayed at 2 and 10 dollars through the end of August, then rose to 3 and 15. The promotional price for GPT-5.6 Sol is valid for about three months from 21 August. The rate for Gemini 3.7 Flash is fixed through 31 December 2026, after which it doubles. Grok 4.6’s price doubles automatically once a request exceeds 200 thousand context tokens.
Why the price list says little
Models use different numbers of tokens for the same task, and the difference erases the difference in price. On the same code review, Opus 5 read roughly 50 percent more and wrote 65 percent more than the production system running on GPT-5.6. The final bill grows twice over: because of the rate and because of usage.
The reverse is also true. An expensive model pays for itself if it solves a complex task in fewer attempts—where a cheaper one would require three tries and manual polishing.
The flagship nobody buys
The clearest illustration is the fate of Fable 5. According to payment service Ramp, which analyzes the spending of 70 thousand companies, two months after launch the flagship model accounted for about 11 percent of customers’ dollar spending on Anthropic tools, but only 6 percent of the tokens purchased. The gap between the shares is explained by the price: few tokens, lots of money.
Opus 5, released at a lower price at the end of July, overtook Fable 5 in corporate spending. The risk dynamic did not change: an expensive flagship can become nonessential, while an affordable model can easily be replaced by a competitor’s model.
But the money keeps growing
The company’s revenue looks solid. Its annualized run rate passed 30 billion dollars in April, 47 billion in May, and exceeded 65 billion by the end of July—roughly a sevenfold increase in a year. Preliminary second-quarter revenue exceeded 11,5 billion dollars, compared with 787 million a year earlier. Investors expect 100 to 120 billion for the year, and the company is preparing to go public.
Financial indicators reflect the situation with a delay, and that is the main risk. Developers shift workflows gradually: first routine work moves to cheaper models, then some complex tasks, and only afterward does this become visible in revenue.
Where they are looking for growth
In parallel, both major labs are assembling teams in Asia, targeting not researchers but executives with local connections. OpenAI hired Kiran Mani as its Asia-Pacific director, brought in Prabhjit Singh from Uber to lead India, and recruited Sandhya Devanathan from Meta to oversee Southeast Asia and Australia. Anthropic put Irina Ghose, previously head of Microsoft India, in charge of its Indian operations.
The logic is simple: in the United States, the competition is for engineers, while on Asian markets, what matters is the ability to negotiate with major corporate clients and regulators. Few people combine experience at American technology companies with an understanding of local markets, so companies are fighting over them.
How to calculate your costs
Calculate the cost of a completed task, not a thousand tokens: measure the full usage across a dozen typical scenarios, including retries and revisions.
Check how long the price you are using to build your budget remains in effect. Half the entries in the table above are temporary rates with a known increase date.
Separate routes by task type. A cheap model for routine work and an expensive one for complex tasks saves more than searching for a single “best-value” model.
Compare models before you start
The service sets its plans, limits and model catalog. If they differ from this article, contact us so we can update it and record a new review date.
Browse modelsAffiliate link: your price stays the same and the project earns a commission.