Claude subscriptions deliver up to 5 times more value than OpenAI's, according to SemiAnalysis

SemiAnalysis has published a detailed analysis of how AI subscription limits — ChatGPT, Claude and others — really work, and concludes that, despite accounting for a relatively small share of the major labs' total revenue, these plans consume a disproportionate amount of compute and erode their margins. According to the firm's figures for Anthropic, subscriptions account for just 10% of the company's revenue but absorb more than 40% of its inference compute, which would cut average revenue per megawatt by some $36 million. For OpenAI, subscriptions carry an even greater weight in total revenue.

The article explains the internal logic of these plans: each monthly payment grants an amount of 'credits' that are consumed differently depending on the model and the type of token (input, cache write, cache read and output), and that internal cost ratio need not match the price ratio those same models have in the public API. That is why, the authors argue, it makes no sense to say that a plan 'is worth $X' in the abstract: its real value depends on the specific combination of plan, model and type of task used, and it also changes over time because providers are constantly adjusting limits — sometimes publicly, with promotions or new launches, and sometimes silently.

To measure this systematically, SemiAnalysis has built a 'Subscriptions Dashboard', a panel that tracks on a daily basis the limits of every OpenAI and Anthropic plan, as well as Meta, SpaceXAI, Cursor, Cognition, Z.ai, MiniMax and Moonshot, with the intention of adding new providers, plans and models as they appear. The methodology consists of running series of calls designed to maximize a single token type at a time (using, for example, passages from 'War and Peace' for the input- and cache-token experiments, and long technical essays to force extended outputs) and observing how far the provider's usage meter moves at each step, in order to infer how many tokens of each type actually fit within each plan's 5-hour limit and weekly limit, and from there derive an 'API-equivalent value' in dollars.

One curious side finding from the calibration process: the researchers detected that one of three identical accounts with the same provider showed limits 20% lower than the other two. After contacting the provider, they confirmed this was not different treatment based on account age but an 'extremely small' A/B test on how to handle the moment users hit their limits; the provider itself insisted it was not an across-the-board cut. The episode serves as proof that subscription limits can be changed silently at any time, and that the measurement method used is sensitive enough to detect it.

In the article's central comparison, Anthropic comes out well ahead of OpenAI. On the top-end models — Fable 5.1 versus GPT-6 Astra — the limits are similar, but with one relevant difference: on Claude plans, Fable usage can consume only up to 50% of the total limit, so a $200-a-month plan retains half its quota after burning through $2,485 of equivalent value in Fable, while the comparable OpenAI plan is completely exhausted after $2,897 of Astra. The gap widens sharply at the model tier both companies present as the everyday option for most users: Opus 5.5 versus GPT 6.1 Sol. There, according to the analysis, Anthropic systematically offers around five times more API-equivalent value than OpenAI, a gap that holds even when compared by raw token volume rather than by dollar value.

The contrast has sharpened with a recent change by OpenAI: the company halved the equivalent value of the $200-a-month plan and, at the same time, introduced a new $500 tier. Those who had already subscribed to the $200 plan before the cut will keep the old, more generous limits until October 29, but new sign-ups start out with the reduced limits. The new $500 plan, for its part, offers only 21% more Astra than the old $200 plan, and the equivalent value for the Sol family of models even fell, in part because OpenAI also cut the price of cached input tokens for GPT 6.1 Sol. The big selling point of the $500 plan is access to a speed of 300 tokens per second ('Ultrafast'), whose limits SemiAnalysis says it is still testing.

The analysis also notes that, before this cut, OpenAI's $200 plan was far more heavily subsidized than its other plans: normalized per dollar spent, the $100 Pro plan offered twice the value per dollar of the Plus plan on Astra, and the $200 plan doubled that figure again. After the cut, the three Pro plans ($100, $200 and $500) now offer the same value per dollar across all models. Anthropic, by contrast, already offered uniform value per dollar across all its subscription tiers before this move, which, combined with the above, leads the authors to conclude that OpenAI's reputation among independent developers for generous limits — against a stingier Anthropic — no longer holds: today Opus 5.5 on any Claude plan offers far more value than any OpenAI offering. The only point in OpenAI's favor is that none of its Pro plans has a 5-hour limit, which makes it easier to use up a larger share of the total monthly quota — something that, according to the article, does not make up for the nearly 4x value difference in favor of Opus 5.5.

Another section of the article examines whether the labs adjust subscription limits when API prices change. Several recent price cuts are mentioned: Fable 5.1 made cache reads 75% cheaper than Fable 5; Opus 5.5 cut input and output token prices by 20% versus Opus 5, and cache reads by 60%; and GPT 6.1 Sol cut cache reads by 50% versus GPT 6 Sol, which was itself already 60% to 67% cheaper than 5.6 Sol. Despite these API price reductions, OpenAI did not change Sol's token limits on the $200 plan when it launched version 6.1, which caused the plan's equivalent value to fall by around 30%. Anthropic likewise did not raise Fable's limits when it launched version 5.1, although it did raise Opus limits alongside the 5.5 price cut — by around 20% on the Max plans and 50% on the Pro plan — an increase that, according to the article, did not fully offset the API price cuts.

The piece interprets these moves as two different strategies for trimming the subscription subsidy. Anthropic gradually reduces the equivalent value as more powerful models are used within the same plan: the drop is small between Sonnet 5.5 and Opus 5.5, but becomes very steep with Fable 5.1. Assuming 100% utilization and API gross margins of 92%, burning through the quota with Opus 5.5 would imply a gross margin of -369% for Anthropic, versus barely 1% if all consumption were on Fable 5.1; at a more realistic 20% utilization, those margins rise to 6% and 80% respectively. The authors note that Anthropic would probably already have software margins on its subscriptions if everyone used only Fable, and that Opus-class models should get cheaper over time as the company manages to reduce their serving size. OpenAI, by contrast, opted for what the article calls the 'nuclear option': cutting the limits of all its tiers at once down to Fable's level. The expected result would have been a public backlash, but according to the analysis OpenAI has largely managed to dodge it, possibly thanks to the combination of positive coverage from its developer event and the fact that it kept the old limits for existing plans for an extra month.

The article also devotes a section to subscriptions from Chinese labs which, despite having less compute available, continue to offer subsidized plans, though with very different levels of subsidy depending on the model; broadly speaking, value per dollar increases the higher the subscription tier, and on average it still sits somewhat below what OpenAI offers.

Finally, the article introduces one last comparison — between use of OpenAI and Anthropic models on their own first-party plans versus their use through third-party applications such as Cursor or Cognition (Devin) — but the available content cuts off right at that point, with a notice that the rest of the article is exclusive to paying SemiAnalysis subscribers. It is therefore not possible to offer details here about that final comparison, since the material provided does not include that portion of the text.