SemiAnalysis: Anthropic Claude plans deliver 5x+ the API-equivalent value of OpenAI subscriptions

Anthropic plans come out far ahead

Subscriptions remain the dominant way consumers and small businesses pay for AI, and a new SemiAnalysis analysis argues they are heavily subsidized marketing tools rather than profit centers. The report's headline conclusion: Anthropic's Claude subscriptions offer more than five times the API-equivalent value of OpenAI's plans. The piece circulated on the front page of Hacker News after publication.

Subsidies as strategy

SemiAnalysis has previously argued that these plans lose money directly but earn it back as customer acquisition. The goodwill generated by OpenAI's generous limit resets, the firm says, was partly responsible for the recent surge in Codex adoption, and it has pushed Anthropic to repeatedly reverse planned cuts to its own subscription limits to avoid a public backlash.

The subsidies are financially significant. For Anthropic, subscriptions account for roughly 10% of revenue yet consume over 40% of inference compute, and they lower blended revenue per megawatt by about $36 million, according to the analysis. Subscriptions make up an even larger share of OpenAI's total revenue, SemiAnalysis says, which means anyone modeling AI lab financials needs to understand subscription limits, not just API pricing.

Credits make plan value a moving target

The core mechanic is credits. A monthly fee buys a pool of credits, and each combination of model and token type — input, cache write, cache read, output — consumes a different amount. Because those credit costs can diverge sharply from published API prices, the same plan is worth different amounts depending on what you run: SemiAnalysis shows the API-equivalent value of a $200-per-month Claude plan changing substantially across models and workloads.

The value also shifts over time. Labs adjust limits publicly around promotions and model launches, but they can change limits silently by tweaking credit costs. OpenAI's pricing page previously advertised its $100 plan as offering five times the usage of the $20 Plus tier, and its $200 plan at twenty times, before the company halved the $200 plan's usage and removed the comparisons entirely.

How the numbers were measured

Measuring any of this is awkward because providers expose only a coarse 0–100% usage meter across 5-hour and 7-day windows, sometimes with a separate meter for individual models. SemiAnalysis says it runs experiments that isolate one token type at a time — passages of a well-known novel for input and cache tests, a technical essay to force long outputs — and records how far the meter moves per million tokens.

Because meters tick in fixed steps, the team tallies the tokens used between each step, discards partial steps at the start and end of a run, and repeats until the rate is pinned within plus or minus 5%. If 500 million cache-read tokens fail to move a meter at all, cache reads are counted as free. Weekly limits cap monthly totals, since the 5-hour windows reset many times within a week. Multiplying usable token volume by blended API prices yields each plan's API-equivalent value.

SemiAnalysis now tracks these rates daily through a dashboard offered to subscribers of its Tokenomics Model, covering every OpenAI and Anthropic plan plus offerings from Meta, Cursor, Cognition, Z.ai, MiniMax and Moonshot, with new providers and models added as they launch.

Why it matters

If you pay for AI tools, the practical takeaway is that the same $20 or $200 buys very different amounts of compute depending on the vendor, the model and your workload — and the ratio can change without any announcement. On SemiAnalysis's numbers, Anthropic currently offers by far the richer deal, which should weigh on anyone choosing between Claude and ChatGPT subscriptions.

The second audience is analysts and investors. Because subsidized subscriptions consume a large share of inference capacity while contributing comparatively little revenue, they materially affect margins and revenue per megawatt. Per SemiAnalysis, ignoring subscription limits means misunderstanding the unit economics of the labs themselves.