SemiAnalysis: Claude Plans Can Take Over 40% of Inference

SemiAnalysis estimates that subscriptions make up about 10% of Anthropic’s overall revenue but can take up over 40% of its inference compute. The research firm set out the figures in an article published on 5 October 2026, writing: “Despite being just 10% of overall revenue, subscriptions can take up over 40% of inference compute and lower blended revenue per MW by ~$36M.”

The article, titled “Anthropic Subscriptions Offer 5x+ More Value Than OpenAI”, compares what the same monthly plan buys across labs. SemiAnalysis calls the Anthropic figures “rough numbers”. They are its estimates, not figures Anthropic has published.

A Tenth of the Revenue, Over 40% of the Compute

SemiAnalysis describes subscription plans as heavily subsidized, and argues that because of this they “can also have a large impact on margins and revenue per MW despite only being a small portion of total revenue”.

It writes that subscriptions are even more important for OpenAI, because they make up a larger portion of its total revenue. Its conclusion for anyone building a model of the labs: “if you want to accurately model AI lab financials, you need to understand their subscription limits.”

How SemiAnalysis Measures a Plan

Subscriptions do not show a price per token, so SemiAnalysis measures them. It runs repeated calls that isolate one token type at a time, input, cache writes, cache reads or output, and records how far the provider’s usage meter moves. It keeps adding steps until the range is within ±5%, then converts the share of a limit used by a million tokens into an API-equivalent dollar value.

For the agentic workload it applies its own usage ratios from September, as reported in its paid Tokenomics Model. In one test, one of three identical accounts at a provider had about 20% lower limits; the provider told SemiAnalysis the account was in an “extremely tiny” A/B test. SemiAnalysis says this shows “providers can silently change subscription limits at any time”.

Claude Plans Against ChatGPT Plans

On SemiAnalysis’s measurements, limits are similar for GPT-6 Astra and Fable 5.1, but Fable can only use 50% of a Claude plan’s limit. It writes that a $200 Anthropic plan would still have 50% left after consuming $2,485 worth of Fable 5.1, while the equivalent OpenAI plan would be exhausted after $2,897 of Astra.

At the mid tier, Opus 5.5 against GPT 6.1 Sol, SemiAnalysis says Anthropic offers about 5x the API-equivalent value across the board. It notes that OpenAI’s Pro plans have no 5-hour limit, but does not think this offsets the ~4x higher API-equivalent value of Opus 5.5 on Claude plans.

Its tracking also confirmed the 50% value cut for OpenAI’s $200 plan. Plans bought before the cut keep the old limits until 29 October, and the new $500 plan offers 21% more Astra than the old $200 plan, by SemiAnalysis’s count.

Price Cuts and Limits

SemiAnalysis lists recent price cuts: Fable 5.1 cut cache reads by 75% against Fable 5, and Opus 5.5 cut input and output pricing by 20% against Opus 5 and cache reads by 60%.

It writes that Anthropic did not raise Fable token limits with 5.1 but did raise Opus limits with the 5.5 price cut. The increase was about 20% more tokens on Max plans and about 50% on the Pro plan, which in its estimate was not enough to fully offset the API price cut.

What Fable Does to the Margin

SemiAnalysis says subscription gross margins are “way lower than API”, and that Anthropic’s way of reducing the subsidy is to give less API-equivalent value on more premium models. The drop is small between Sonnet 5.5 and Opus 5.5 and becomes large with Fable 5.1, it writes.

Assuming 100% utilization and 92% API gross margins, it calculates that a user maxing out Opus 5.5 corresponds to a -369% gross margin on the plan and one maxing out Fable 5.1 to 1%. At what it calls a more realistic average utilization of 20%, those become 6% and 80%.

Its reading: “Anthropic would likely already have software like margins on their subscription plans if everyone only used Fable”. OpenAI, by contrast, cut straight to Fable-level limits across the board, SemiAnalysis writes.

SemiAnalysis’s Earlier Margin Calls

In an August 2026 article, SemiAnalysis wrote that it told Tokenomics Model subscribers in January “why inference gross margins are north of 60%”, and that a June deep dive “showed how Opus 4.8 in particular had 85%+ margins”. It added: “This has since become the default number everyone cites when analyzing Anthropic.”

The January call went to Tokenomics Model subscribers. We did not read the January or June reports, only the free articles that describe them.

The Structural Read

SemiAnalysis’s method treats a subscription as a bundle of credits that each model and token type spends at a different rate. That is why it says the same plan has no single value: it depends on the model and the workload.

On its measurements, Anthropic gives less API-equivalent value on its premium models, while OpenAI cut limits across the board. SemiAnalysis reads both as ways of reducing the subsidy on subscriptions.

Its margin figures rest on assumptions it states, a 92% API gross margin and a utilization level. As we read them, they show the direction of the effect more than a reported margin.

Three Implications

PLAN BUYERS SemiAnalysis measures about 5x the API-equivalent value on a Claude plan against ChatGPT at the Opus 5.5 versus GPT 6.1 Sol tier, and says providers can change limits without notice.

LAB FINANCES In its rough numbers, subscriptions are about 10% of Anthropic’s revenue but take over 40% of inference compute and lower blended revenue per MW by about $36M.

ANALYSTS SemiAnalysis says modelling a lab means tracking its subscription limits, which it now measures through a dashboard for its paying Tokenomics Model clients.

The Business Engineer Lens

This story maps onto the Business Engineer framework The State of The Inference Economy.

The framework puts it this way: “Inference is a revenue engine: continuous, distributed across millions of applications and agents, incurring marginal cost with every query.”

As we read it, a flat monthly subscription sits awkwardly against that marginal cost, and SemiAnalysis’s estimates show where the gap appears: over 40% of compute for about 10% of revenue, in its rough numbers.

What Is Not Established

A summary published by Crypto Briefing on 11 October attributes to SemiAnalysis an estimate that Anthropic’s inference business could reach margins of up to 88%. That figure does not appear in the two SemiAnalysis texts we read, the 5 October article and the August article; the closest figures there are the 85%+ for Opus 4.8 from June and the 92% API gross margin used as an assumption on 5 October. SemiAnalysis’s paid Tokenomics Model and institutional reports were not read.

Every figure here is a SemiAnalysis estimate or measurement. The 10% and over 40% are what it calls rough numbers, and the margin figures depend on the utilization and API margin it assumes. Anthropic has not confirmed them in any text we read.

The Bottom Line

SemiAnalysis’s 5 October 2026 article estimates that subscriptions bring in about 10% of Anthropic’s revenue while taking over 40% of its inference compute, and that a Claude plan gives about 5x the API-equivalent value of OpenAI’s at the Opus 5.5 versus GPT 6.1 Sol tier. On its assumptions, a plan used only for Fable 5.1 would carry an 80% gross margin at 20% utilization, and one used only for Opus 5.5 a 6% margin.