“Across our token-based datasets, we’ve observed a ~50% Y/Y decline in token prices,” the data firm YipitData wrote in its Anthropic Pre-IPO Overview, published on 22 October 2026 and updated on 23 October. The note is written for investors preparing to evaluate Anthropic as a public company, and it lists what it says has contributed to the drop.
A token is the small chunk of text an AI model reads or writes, and AI companies charge per million of them. YipitData’s figures are estimates from its own panels, not company figures, and the note does not publish the prices behind the ~50%.
Business Pill · COST PER TASK
A one-minute explainer of cost per task: pricing the job rather than the word. It teaches the general idea only and says nothing about any company in this story.
The key insight: As we read it, the number that matters is the direction, not the level: YipitData says token prices in its datasets fell about 50% in a year and lists Anthropic’s price cuts and caching among the contributors, while the same note says adoption growth is “beginning to flatten out.”
YipitData names four contributors: “Price cuts from Opus 4.5, Sonnet 5’s introductory pricing that became permanent, greater cache hit rates, and Fable 5.1 cache hit pricing reductions have contributed.” Two of the four are list-price changes. The other two concern caching, where a model reuses text it has already processed and the customer pays a lower rate for those tokens.
The same note says that across its token and pricing datasets “we observed an uptick in spend driven by Opus 5.5 (9/22 launch).” Among the questions it suggests investors put to Anthropic: “Could pricing stabilize if you introduced more access to premium tier models like Mythos?”
YipitData’s estimate of a ~50% year-on-year decline in token prices across its token-based datasets, shown as an index with the year-earlier level set to 100. The index is ours, for illustration; YipitData does not publish the underlying prices. Source: YipitData Anthropic Pre-IPO Overview, 22-23 Oct 2026.YipitData’s ~50% year-on-year decline in token prices across its token-based datasets, shown as an index with the year-earlier level at 100. Index ours, for illustration. Source: YipitData Anthropic Pre-IPO Overview, 22-23 Oct 2026.
What Was Posted on X
GPU rental prices are still rising while September AI token dollar spend was basically flat (+0.2% MoM).
Thatโs NOT weak demand. Token volume kept rising even as unit prices fell. As per @vercel usage (token demand) is soaring, spend is +1,180% since December as the average price of token has fallen ~47% due to cheaper open-weight models (these tokens require almost the same compute as the frontier tokens for comparable models).
Dollar spend is a noisy signal right now. Token volume is the cleaner read on actual inference demand.
Rising GPU rates = supply still tight (packaging, HBM, power). Volume growth = the demand that ultimately drives more silicon, interconnect, and capacity.
Different metrics for different parts of the stack:
โข Token volume โ better structural signal for the semi/infra trade โข GPU rental prices/utilisation โ more direct for capacity owners (neoclouds)
Falling token prices + sticky-high GPU rates is NOT the end of the AI buildout. Itโs Jevons playing out in real time.
Puru Saxena, founder of AlphaTarget, reads token prices the other way round from a spending slowdown: he says September AI token dollar spend was basically flat (+0.2% month on month) while GPU rental prices kept rising, and, citing Vercel usage, that spend is up 1,180% since December as the average token price fell about 47%. His conclusion is that token volume, not dollar spend, is the cleaner read on inference demand. The figures and the view are his, citing Vercel; they sit alongside YipitData’s ~50% decline above.
The Prices on the Page Today
YipitData gives the size of the decline but not the prices. The list prices below come from our earlier reporting. They show where Anthropic’s models sit now against rivals, not what they cost a year ago.
MiniMax’s listing, in our MiniMax piece, the same as DeepSeek V4.1 Flash’s off-peak rate
List prices as reported in our pieces of 16 to 23 October 2026. Different models with different capabilities; prices only, not a quality ranking.
Two points stand out in our earlier reporting. Moonshot priced Kimi K3.1 Pro at half of the $4 and $20 Anthropic lists for Claude Opus 5.5, by our arithmetic, and at the same list price Anthropic gives Claude Sonnet 5.5. Anthropic also sells the same Opus 5.5 at two speeds, with fast mode at double the standard price for “up to 2.5x higher output tokens per second,” according to its docs.
The Competition YipitData Flags
YipitData writes that “Within OpenRouter, China-based providers like DeepSeek and Z.ai have taken significant share of usage, in an environment where it’s easy to switch models.” OpenRouter is a service that routes developers’ requests to many models through one account. Our MiniMax piece reported that OpenRouter’s rankings put Space Bunny Alpha, which MiniMax describes as an earlier version of M3.1 Flash, second by tokens processed over the 30 days to 18 October, at 56.7T.
Cheaper Tokens, Flatter Adoption
The same YipitData note says Anthropic’s business adoption growth is “beginning to flatten out, meaning more growth must come from existing customers.” Ramp’s letter of 14 October, covered in our Ramp piece, records “two straight months of virtually no growth in token spend” among the firms on its cards. The two datasets measure different things, so they are not combined here.
On the revenue side, a draft prospectus viewed by The Wall Street Journal showed Anthropic generated around $16.3 billion in revenue in the six months ended 30 June 2026 and spent nearly $9 billion on computing capacity and infrastructure, as covered in our compute piece.
Connecting the Coverage
Our overview of the YipitData note led on its finding that the top 1% of customers make up nearly 50% of Anthropic spend, with B2B retention of 87% after month 12 and 95% after month 24. This piece takes one line from that note, the price decline, and sets it against the prices we have reported.
Our Kimi piece covered Moonshot’s $2.00 and $10.00 price for a model its blog places just behind Opus 5.5 on Terminal-Bench 4.0, and our fast-mode piece covered Anthropic’s $8 and $40 premium tier, out of research preview from 15 October 2026.
The Structural Read
As we read it, two of the four contributors YipitData lists are Anthropic’s own pricing decisions: the Opus 4.5 price cuts and Sonnet 5’s introductory pricing becoming permanent. YipitData does not say how much each contributed.
As we read it, the other two are about caching: higher cache hit rates and cheaper Fable 5.1 cache hits lower what a customer pays per token without a change in the headline list price.
As we read it, lower-priced rivals are visible in our reporting: Kimi K3.1 Pro at $2.00 and $10.00, the same as Sonnet 5.5 and half of Opus 5.5, and MiniMax M3.1 Flash at $0.15 and $0.60, while YipitData says China-based providers have taken significant share of usage on OpenRouter.
YipitData, Anthropic Pre-IPO Overview, 22 October 2026
“Across our token-based datasets, weโve observed a ~50% Y/Y decline in token prices.”
Three Implications
FOR INVESTORS READING THE PROSPECTUS YipitData suggests asking Anthropic: “Could pricing stabilize if you introduced more access to premium tier models like Mythos?” Anthropic already sells Opus 5.5 fast mode at $8 and $40, double its standard price, per its docs.
FOR BUYERS OF AI The list prices in our reporting run from $0.15 per million input tokens for MiniMax M3.1 Flash to $8 for Opus 5.5 fast mode, and YipitData says switching models on OpenRouter is easy.
FOR READERS OF REVENUE FIGURES As we read it, with lower prices per token, revenuegrowth leans on volume: YipitData says “more growth must come from existing customers,” and Ramp recorded “two straight months of virtually no growth in token spend.”
The framework describes the token as “the unit of cognition the model produces, the unit of compute the data center serves, the unit of price the lab charges, and the unit of value the enterprise extracts.”
As we read it, YipitData’s figure measures the third of those roles, the unit of price the lab charges, falling by about half in a year. The framework was published in May 2026 and does not discuss YipitData’s data.
What Is Not Established
The prices behind the ~50%. YipitData does not publish the year-earlier or current prices it measured, or the mix of models.
Whether the decline is in list prices or in what customers actually paid. YipitData includes higher cache hit rates among the contributors, which lower the effective price without a list-price change.
Anthropic’s view. The note poses questions for the company; it does not carry answers, and we found no Anthropic response in it.
How representative the panels are. YipitData says its analysis combines data from 1,500 SMB businesses, 200 large enterprise companies and millions of consumers.
Business Engineer Framework
The CFO’s Guide to the Token Economy
A Business Engineer guide to why the cost of intelligence falls while the cost of deploying it rises.
YipitData estimates that token prices in its datasets fell about 50% in a year, and lists Anthropic’s price cuts, greater cache hit rates and cheaper cache hits among the contributors. In the same note it reports flattening adoption and China-based providers taking significant share of usage on OpenRouter, while Anthropic’s list prices still run from $2 to $8 per million input tokens across the models in our table.
95,000+ executives read Business Engineer for the AI strategy frameworks cited by ChatGPT, Claude, and Perplexity.
A note on sourcing. The ~50% decline, its contributors and the competition and adoption lines are YipitData’s, from its Anthropic Pre-IPO Overview (published 22 October 2026, updated 23 October), which says its analysis combines data from 1,500 SMB businesses, 200 large enterprise companies and millions of consumers, accounting for more than $4B in annual AI spend. They are YipitData’s estimates from its panels, not company figures. List prices come from our earlier pieces, linked in the table. The index in the chart is ours, for illustration.
Gennaro Cuofano is a CRO and tech executive who has worked in AI since late 2015, starting with bringing NLP, natural-language generation, voice and chatbot products to the marketing industry. His background is in law and finance: he holds a Master's degree in Law and an International MBA with an emphasis on corporate finance (LUISS Business School and the University of San Diego, 2012), and worked as a financial analyst at a real-estate investment firm in San Diego and as an assistant controller. His work focuses on business model strategy, business engineering and, more broadly, structural analysis: how companies make money, read from their own filings. He created FourWeekMBA and leads research there and at The Business Engineer, his newsletter on AI and business strategy, with over 95,000 subscribers and more than 1,000 published analyses. His writing has also appeared on Entrepreneur (for example, How to Digitize a Family Business in Europe) and HackerNoon. Find him on LinkedIn and Substack. See how we source.
Scroll to Top
Discover more from FourWeekMBA
Subscribe now to keep reading and get access to the full archive.
