Five X Marks the Spot: SemiAnalysis Measures a 5x Value Gap Between Claude and ChatGPT Subscriptions

Two hundred dollars buys you a subscription at OpenAI, and two hundred dollars buys you a subscription at Anthropic. According to a new limit-testing study published by SemiAnalysis this week, those two identical price tags hide a gap of roughly five times in what you actually get back — at least on the mid-tier models both labs market as the daily driver for most users.

The research firm, which has built a reputation for adversarially measuring the parts of the AI economy that companies decline to publish, tested subscriptions from Anthropic, OpenAI, Meta, SpaceXAI, MiniMax, Moonshot, Z.ai, Cursor and Cognition. Its conclusion is stark: on agentic workloads running each lab’s workhorse model — Claude Opus 5.5 versus GPT-6.1 Sol — Anthropic’s plans deliver about 5x more API-equivalent value per month than OpenAI’s. The article is by Andrew Megalaa, Max Kan and Dylan Patel; the full dashboard sits behind the subscriber paywall, but the public portion carries the headline numbers.

The headline numbers

SemiAnalysis defines “API-equivalent value” as a plan’s full monthly usage limit priced at first-party list API rates — in other words, what the tokens included in your subscription would cost if you bought them through the API instead. On that measure, at the mid-tier:

Plan fee	ChatGPT (GPT-6.1 Sol)	Claude (Opus 5.5)	Claude advantage
$200/month	$2,084 (10.4x) on Pro 200	$11,726 (58.6x) on Max 20x	~5.6x
$100/month	$1,055 (10.6x) on Pro 100	$5,725 (57.3x) on Max 5x	~5.4x
$20/month	$211 (10.6x) on Plus	$1,178 (58.9x) on Pro	~5.6x

The pattern is remarkably consistent: OpenAI’s plans multiply your fee by roughly 10x in API value, while Anthropic’s multiply it by roughly 58x. SemiAnalysis calls Anthropic “an overwhelmingly better deal” at this tier — and it anticipates the obvious objection. GPT-6.1 Sol is much cheaper per token than Opus 5.5, so dollar-denominated value should flatter the expensive model. The firm says the gap remains large even measured in raw tokens per dollar, not just dollars of API value.

At the flagship tier, the picture flips to near parity. A $200 ChatGPT Pro plan is worth about $2,897 of GPT-6 Astra usage, while a $200 Claude Max 20x plan is worth about $2,485 of Fable 5.1. But there is a footnote that matters: Fable 5.1 can consume only up to half of a plan’s usage limit. After burning $2,485 of Fable, your Anthropic plan “would still have 50% left,” while the OpenAI plan “would be fully exhausted after $2,897 of Astra.” That reserve is exactly what lets a Claude subscriber drop down to Opus and keep working.

How you measure an unpublished limit

The methodology is the quiet star of the report. Subscriptions don’t publish token allowances; they show a usage meter from 0 to 100 percent over 5-hour and 7-day windows. To turn that meter into numbers, SemiAnalysis isolated one token type per experiment: input tests add a random tag to every call so nothing gets cached; cache-write tests mark prompts for caching with a new tag each time; cache-read tests reuse a fixed tag; output tests use a technical essay to force long generations. The firm then counted usage in “steps” — the tokens consumed between two movements of the meter — dropping partial first and last steps to avoid bias, and kept adding steps until the measured range narrowed to within plus or minus 5 percent.

The team also subtracted the fixed overhead that no request can avoid, treated very cheap cache reads as free when the meter failed to move after 500 million tokens, and converted the result into per-window and per-month limits priced at blended API list rates. Two workload shapes anchor the analysis: an agentic mix (0.4% fresh input, 96.6% cached input, 2.6% cache writes, 0.3% output) drawn from SemiAnalysis’s own September usage, and a chat mix (2% input, 75% cached input, 13% cache writes, 10% output).

One side finding deserves attention on its own: of three identical test accounts at one provider, one showed roughly 20 percent lower limits. The provider confirmed it was part of an “extremely tiny” A/B test on limits. The lesson SemiAnalysis draws is double-edged — providers can silently change your limits at any time, and a careful enough methodology can catch them doing it.

Why the labs behave differently

The report argues that subscriptions matter far more to lab economics than their revenue share suggests. For Anthropic, SemiAnalysis estimates subscriptions represent about 10 percent of revenue but over 40 percent of inference compute — a subsidy that lowers blended revenue per megawatt by roughly $36 million (modeled figures from the firm’s Tokenomics Model, not company disclosures). Each lab has chosen a different way to shrink that subsidy.

Anthropic lowers value as models get stronger. Within a single Max 20x plan, Sonnet 5.5 delivers about $12,529 (62.6x the fee) on agentic work, Opus 5.5 about $11,726 (58.6x), and Fable 5.1 about $2,485 (12.4x). At 100 percent utilization and assumed 92 percent API gross margins, the firm models Opus 5.5 at roughly negative 369 percent gross margin and Fable 5.1 at about 1 percent; at a realistic 20 percent average utilization those become roughly 6 percent and 80 percent. In SemiAnalysis’s view, Anthropic “would have software-like margins if everyone used only Fable” — the ladder is a steering mechanism, not an accident.

OpenAI cut limits across the board. The firm describes it as the “nuclear option”: moving straight to Fable-class limits rather than degrading gracefully by model tier. SemiAnalysis confirmed the roughly 50 percent value reduction on the $200 plan — OpenAI halved tokens per model tier and also cut cached-input pricing for Sol, which together pushed Sol-class value down by more than half. Plans bought before the cut keep old limits until October 29. The new $500 Pro plan offers only about 21 percent more Astra than the old $200 plan; its headline feature is Ultrafast speed, quoted at 300 tokens per second. Where OpenAI’s ladder once rewarded scaling up — Pro 100 gave about 2x Plus per dollar, Pro 200 another 2x — Pro 100, 200 and 500 now offer the same tokens per dollar across every model.

The report notes one genuine counter-argument in OpenAI’s favor: none of its Pro plans carry a 5-hour limit, which makes it easier to actually consume a high share of the monthly allowance. SemiAnalysis does not think that offsets a ~4x value gap against Opus 5.5.

What the numbers don’t prove

To its credit, the study is explicit about its own limits. API list price is not what heavy users would actually pay — the figure is a comparison index, not realized savings. The workload is SemiAnalysis’s own; a mix that is 96.6 percent cached input rewards models with cheap cache reads, and your usage may look nothing like it. Token efficiency is unmeasured — if one model finishes a task in fewer tokens, a smaller plan goes further. Quality is not in the chart: more dollars of value does not mean better results per task. And the whole thing is a snapshot of limits that labs can change without notice, which is why the firm suggests rechecking combinations daily.

The other finding worth flagging: Chinese labs still offer heavily subsidized plans, with per-dollar API-equivalent value averaging a little under the ~12x OpenAI gives — far below Anthropic’s ~58x — while third-party wrappers like Cursor and Devin deliver worse value than first-party plans for the same underlying models.

The practical read

For anyone coding with an agent daily, the guidance is straightforward. On Claude, Sonnet and Opus deliver by far the most value — keep Fable for tasks that genuinely need it. On ChatGPT, Sol stretches your allowance further than Astra per task if the quality is sufficient. Don’t buy up for value alone: OpenAI’s higher tiers now price headroom and speed, not better rates. And if you hold a $200 ChatGPT plan purchased before the cut, October 29 is the date your old limits die.

Five-x is a big number, and it will not last forever — Anthropic’s subsidy and OpenAI’s rationing are both transitional states on the road to margins that work. But as a snapshot of where the two most-watched AI labs stand today, the message is unusually crisp: same price, very different plans.