Loading today’s stories
Invest1 distinct publisher2 min readUpdated
SemiAnalysis says a $200 plan can yield $8,000 of Anthropic API-equivalent usage and $14,000 of OpenAI's. Even the best published caching fix leaves the gap several times wide.
The Investor · Invest desk
Compiled by The InvestorSomething wrong?How this is made
Forty times and seventy times: those are the ratios once the API-equivalent figures are divided by the $200 price [1][2]. Held for a year, the ceiling user on the Anthropic plan draws about $96,000 of list-rate inference for $2,400 in fees, and on the OpenAI plan about $168,000 [3][4].
Most subscribers will never get near that [7]. The researchers got there by running one long workload continuously against the weekly caps rather than holding conversations [3], and agentic coding is precisely the workload that behaves that way, consuming far more tokens than chat or a one-shot request [14]. A team leaving agents running overnight is not an unusual user. It is the test subject.
Prompt caching is the one mitigation with a published number attached: an independent study cited in the research cut effective API cost by 88.6%, to roughly $0.57 per million tokens [10]. Apply that discount to the ceiling and the Anthropic case still represents about $912 a month of wholesale inference on a $200 plan, the OpenAI case about $1,596 [5][6]. Measured against its own best fix, the subsidy is still 4.6x to 8x [7]. Caching narrows the gap. It does not close it.
Price is not the only dial available. The $200 tiers are sold as 20x the entry plan's usage, with $100 tiers at 5x [6], and that is a quota, not a dollar amount. It can be redefined without anyone announcing a price rise, and it does the same work on margin.
There is also evidence buyers will not volunteer to cover the difference. Ramp's August AI Index put Anthropic ahead of OpenAI in US business adoption, 43.5% of companies to 39.7% [11], but Ramp lead economist Ara Kharazian has noted that Anthropic's most expensive model took just 6% of the tokens customers bought in its first month [12]. Adoption is broad; appetite for the top of the menu is not. That is the demand curve the labs have to reprice against, and it argues for quieter limits ahead of visible list prices.
SemiAnalysis frames its finding as a margin question, whether flat-rate plans hold without tighter caps or higher prices [15]. The buyer's version is duller and more useful: the durable number is not $200 a seat, it is the multiple sitting behind it, and a headcount plan built on today's multiple contains an input nobody has priced. The 30x spread in token use across identical tasks [9] means that input cannot be estimated from your own usage history either.
Follow any of these and your For You feed starts watching them — no settings page required.
Ranked by verification strength, evidence, and original report placement.
SemiAnalysis found that heavy users can extract roughly $8,000 of API-equivalent usage from Anthropic's $200 Claude plan in a month, based on standard API rates.
SemiAnalysis found that heavy users can extract about $14,000 of API-equivalent usage from OpenAI's $200 plan.
SemiAnalysis purchased subscriptions from both companies and performed extensive coding work until the maximum amount of work allowed in the week was reached, using an extremely long continuous workload rather than regular conversations.
Initial findings from the SemiAnalysis study were released in June, with additional findings released on August 23.
Industry insiders had assumed a $200 subscription would yield about $2,000 worth of tokens in a month; users who fully utilise all weekly limits can reap three to seven times more than that.
Anthropic's Max plans start at $100 monthly with either 5x or 20x the usage of Claude Pro; OpenAI's $100 plan provides 5x Plus usage and its $200 plan 20x.
Evidence-backed comparisons of source perspectives and observed adoption signals. Read the methodology
Which Builder, Operator, and Investor concerns the observed source mix emphasized—not a truth score.
Evidence, demonstrated adoption, hype gap, incentives, and confidence are assessed independently, each on its own current evidence. How these are measured.
Thin: one secondary outlet relaying unlinked third-party research
Every number in the cluster reaches the reader through a single crypto-media publisher summarizing outside work. The primary SemiAnalysis report is neither linked nor dated by year, the eight-model SWE-bench Verified token study and the prompt-caching study are both unattributed, and two inputs (Gartner, the WSJ revenue figures) are credited to the outlet's own prior coverage. The methodology that is described — buy both plans, run one continuous coding workload to the weekly cap — is coherent and internally consistent, and the arithmetic multiples follow directly, which keeps this above the floor.
Vendor adoption quantified; ceiling-user prevalence not
There is real third-party usage data in the cluster: Ramp's August AI Index puts Anthropic at 43.5% and OpenAI at 39.7% of U.S. companies, and Anthropic's priciest model drew just 6% of purchased tokens in month one — evidence that paid agentic access is widely bought but that top-tier consumption concentrates narrowly. What is absent is the distribution that the story's thesis needs: how many subscribers actually sit at the weekly ceiling. The article concedes most probably never get there without sizing them.
Ceiling case presented as the seat economics
The framing generalizes a deliberately extreme measurement. The 40x-70x multiples come from running one continuous workload until weekly caps are exhausted, and the same coverage admits most subscribers never approach that; list API rates are also treated as the vendors' cost of serving, which they are not — the cluster's own caching datapoint cuts effective cost by 88.6%, leaving roughly $912-$1,596 rather than $8,000-$14,000. The gap is positive rather than extreme because the residual after maximal caching still exceeds the $200 price several times over, so the direction of the argument survives its own discount.
Commercially motivated originators, self-citing relay
None of the numbers originate with a disinterested party. SemiAnalysis sells research and benefits from headline-grabbing findings about frontier-lab economics; Ramp's AI Index is a marketing artifact of its corporate-spend data and its own economist supplies the interpretation; Gartner's five-fold inference forecast is advisory product. The relaying publisher cites its own prior articles as sourcing and appends an investment disclaimer, and neither OpenAI nor Anthropic was given a chance to contest figures about their own margins.
Low-moderate: direction plausible, magnitudes unverified
Confidence is limited by single-publisher sourcing with no primary documents in the cluster and by two internal soft spots — the unnamed studies behind the token-burn numbers and the use of list API pricing as a cost proxy. The qualitative core (agentic coding makes flat-rate seats structurally hard to price, and heavy users capture multiples of what they pay) is coherent and survives the cluster's own caching discount; the specific $8,000/$14,000 and 40x/70x figures should be treated as one firm's unaudited ceiling measurement.
invest
The AI trade's weak link is the buyer: Anthropic's best model took 6% of its tokens1 distinct publisher
product
Card data, not vendor decks: the enterprise AI split is unsettled, not settled1 distinct publisher
invest
The first AI IPO writes the comp sheet, and Anthropic is holding the pen1 distinct publisher
leadership
The AI bill nobody reconciles: cost per finished task, not per million tokens1 distinct publisher
Distinct publishers with included, body-backed reporting in this cluster.
1 article · August 23, 2026