Semiconductor and AI research firm SemiAnalysis published a subscription-quota benchmark on October 5, and the conclusion is blunt: at the $200-a-month tier, running agentic workloads on mid-range models, Anthropic's subscription converts to an API-equivalent value of more than 5 times OpenAI's. The two models compared are Claude Opus 5.5 and GPT-6.1 Sol.
The report is credited to Andrew Megalaa, Max Kan and Dylan Patel. Testing covered Anthropic, OpenAI, Meta, SpaceXAI, MiniMax, Moonshot and Z.ai, plus two wrapper-product subscriptions, Cursor and Cognition.
How the quotas were measured
Neither company discloses how many tokens a subscription quota is worth — both only show a usage-percentage meter, plus two rate-limit windows, five hours and seven days. SemiAnalysis worked backward: it isolated input, cache-write, cache-read and output tokens separately, measured how much each type moves the usage meter, converted the two windows' readings into a monthly token budget, then priced that budget at official API rates.
According to Implicator, the team fed War and Peace into paid accounts in chunks — random tags to force fresh input, fixed tags to test cache hits, and long technical text to force long outputs — repeating the test until the margin of error narrowed to within plus or minus 5%.
The two companies are adjusting quotas in opposite directions
The test landed right as both companies were mid-adjustment.
OpenAI reopened its $200 Pro subscription on September 30 and switched to pricing quotas by API spend. Codex lead Thibault Sottiaux put it this way at the time:
"If you do the math, it will net out at half the dollar in API spend compared to the old Pro $200 plan."
SemiAnalysis cites a detail: this tier's quota in Codex and Work dropped from 20x Plus to 10x, and a new $500 Pro tier was added above it. The report describes OpenAI's approach as cutting limits across the board.
Anthropic went the other way: API input/output pricing for Opus 5.5 dropped 20%, cache-read pricing dropped 60%, while the Opus quota on the Max tier rose about 20% and on the Pro tier about 50%. Per the report's analysis, Anthropic is tightening margins by raising how much quota the flagship Fable model consumes relative to Opus, rather than cutting quotas across the board.
That's why the "5x" mostly shows up at the mid-tier model level. At the flagship tier, SemiAnalysis found little difference between Fable 5.1 and GPT-6 Astra.
A rough tally of subscription cost
The report offers a more telling ratio: subscriptions make up only about 10% of Anthropic's revenue, yet may consume more than 40% of its inference compute.
Doing the rough math on those two ratios: every unit of compute a subscription uses brings in about 0.25 units of revenue; every unit of compute other business, like API, uses brings in about 1.5 units. That's roughly a 6x gap. This is only a simplified estimate based on the report's ratios — it doesn't account for cache hits or peak/off-peak scheduling, and Anthropic has never published its own figures.
That gap explains two things: why OpenAI is willing to risk user churn to cut quotas, and why Anthropic raised Opus's quota while also dialing up Fable's consumption rate. For developers who rely heavily on Claude Code, subscriptions are currently far more economical than pay-as-you-go — how long that lasts depends on how long Anthropic is willing to keep subsidizing it.
Behind the paywall
The report's public section stops right at "third-party wrapper subscriptions don't match official ones" before hitting a paywall — no numbers are given in the public version for exactly how much worse products like Cursor and Devin perform. Reports vary on what subscriptions from domestic Chinese vendors like MiniMax, Moonshot and Z.ai measured at, so those aren't cited here.
Worth noting: this kind of benchmark depends on the precision of the usage meter's display, and vendors can adjust how they measure usage on the backend at any time. Neither Anthropic nor OpenAI has publicly responded to the report's conclusions.
Sources: SemiAnalysis subscription-quota benchmark, Techmeme, CocoLoop, AI Weekly, Implicator; the "5x" figure is based on the API-equivalent value of Opus 5.5 versus GPT-6.1 Sol at the $200 tier, and compute-to-revenue ratios are SemiAnalysis estimates.