What an AI subscription is really worth: SemiAnalysis measured the meter, not the bill

What SemiAnalysis measured about the AI subscription, and what it did not#

On 5 October the analyst firm SemiAnalysis published a limit test of the major AI subscriptions under a headline claiming Anthropic's plans offer more than five times the value of OpenAI's. The post then went round the timeline faster than a fox through a hedgerow, and the version most people repeated is the tidy one: a $200 Claude plan is worth five times a $200 ChatGPT plan, as if an AI subscription had one price tag and one worth.

The tidy version is not quite what the firm said. Its method turns each plan's usage meter into dollars at API prices, which it calls API-equivalent value. It does that by measuring how far the meter moves per million tokens, then pricing those tokens using its own September agentic usage mix. The headline figure is roughly five times on the pairing of Claude Opus 5.5 against OpenAI's GPT-6.1 Sol.

That is the authors correcting the headline in their own thread, and it is the most useful sentence in the whole episode. The firm also says, in a second post in the thread, that on a different model tier the limits are similar across the equivalent Anthropic and OpenAI plans. So the five times figure belongs to one model pairing, on one workload mix, on one measuring day. It is a real measurement. It is not a law.

The caveats the authors print themselves#

SemiAnalysis is candid about the limits of its instrument. In the article it says reading from the cache costs almost nothing on some plans and may not move the meter at all, which means a heavily cached workload can look cheaper or dearer depending on how a provider meters it. We could read only the unpaywalled part of the article, so figures that appear only in the paid section are left out here rather than guessed.

What OpenAI changed on 29 September, and why OpenAI usage limits are the live issue#

The five times figure has a context, and the context has a date. At its DevDay on 29 September OpenAI launched a $500 Pro plan and said Pro 200 usage in ChatGPT Work and Codex falls from 20 times to 10 times the Plus allowance, with the change effective from 30 October. The same report says existing Pro 200 subscribers keep their current limits through 29 October and receive a one-time grant of usage credits worth $2,500, expiring at the end of the year.

Follow the trail of dates and the buyer's calendar turns out to be short. A team on Pro 200 today has about three weeks before the allowance changes, and the grant that cushions it has an expiry date of its own. Engadget notes that OpenAI was the first lab to offer a $200 plan at the end of 2024, and that nearly every other provider, Anthropic included, followed with a $200 tier of its own. So a change at the top of OpenAI's range is a change to the benchmark everyone else is priced against.

On the new plan, the same report says the $500 tier is the only Pro option with the Ultrafast speed tier, up to 300 tokens a second in Codex. SemiAnalysis, according to its own article, found the new plan only offers 21 per cent more Astra than the old $200 plan. Read those together and, through the thicket of plan tiers, the shape is plain: the price of the ceiling rose faster than the ceiling.

The Anthropic side, dated#

Anthropic's side moved too, and with less precision. Its Opus 5.5 launch page of 22 September says it is increasing five-hour usage limits on Pro, Max, Team and seat-based Enterprise plans and offering a rate-limit reset that subscribers can save. The launch post, as we read it, gives no figure for the size of the increase, so a buyer cannot budget against it. On price, Claude's pricing page lists Pro at $20 a month and Max from $100 a month, with Max giving five or twenty times Pro's usage.

A fox reading both pages learns two different lessons. OpenAI has announced its change, with a date, a ratio and a goodwill credit. Anthropic has announced a direction, without a number. A buyer cannot budget an AI subscription against a direction.

Claude vs ChatGPT is the wrong question, because value is not a bill#

API-equivalent value answers one question about an AI subscription: how many dollars of list-price tokens does the meter let you burn? It does not answer the question a budget holder cares about, which is how many dollars of work you get done. Those can point in opposite directions, and on the independent evidence they do.

Artificial Analysis publishes a cost per task for the models on its Intelligence Index. For Claude Opus 5.5 it records a score of 58 and a cost of $5.98 per task, after generating about 260 million output tokens to run the index. For GPT-6.1 Sol it records 52, a cost of $0.72 per task and about 67 million output tokens. Both pages are for the maximum-effort versions of the models, and Artificial Analysis lists the list prices behind them: $4 in and $20 out per million tokens for Opus 5.5, $2 in and $10 out for Sol with a 95 per cent cache discount.

Put the two sets of numbers side by side and the picture splits. On one meter, a Claude plan hands you roughly five times the API dollars. On another, the same model costs more than eight times as much per task to run on the API. Both can be true, because the first measures what a plan lets you consume and the second measures what the model consumes to finish a job. A headline token price is not a running cost, and a plan's "value" is not a saving.

Notice what this does to the question in the search box. Claude vs ChatGPT has no stable answer, because the answer depends on which model, at which effort setting, on which tasks, on which plan, on which day. The most honest sentence available is the dull one: measure your own work.

ChatGPT Pro vs Claude Max: what the agent spend data says about seats#

If you want a sense of how far consumption can swing around a seat, look at the one dataset that shows real agent usage at scale. On 5 October Epoch AI published an analysis of OpenAI's own figures for coding-agent use by its researchers. At API list prices, the median researcher's daily agent use reached $601 by mid-August, and the 90th percentile exceeded $7,000 a day. The median was growing about 1.8 times a month, a doubling every 34 days.

Epoch states the limits as clearly as the figures. The values are at list prices rather than internal cost, they are self-reported by OpenAI and cannot be independently verified, the two statistics describe different people at one moment rather than the same people over time, and the firm calls such rapid growth probably unsustainable. Take it as a shape, not a forecast.

The shape is the useful part. A heavy user can consume more than eleven times what a middling one does, and usage is climbing. A flat seat price cannot hold that spread for long, which is why providers keep rewriting allowances, and why a seat comparison such as ChatGPT Pro vs Claude Max is really a comparison of two meters that their owners can change. SemiAnalysis found evidence of exactly that: in its article, providers can silently change limits.

That is the real risk of an AI subscription to a business, and it is not about which brand wins. It is that the unit you budget in is controlled by someone else and can be revised on a Tuesday. Our AI consultancy work treats this as a procurement problem first and a model problem second, and our earlier note on Sign in with ChatGPT and GPT-6.1 Sol makes the same point from the developer side.

A Monday plan for an AI subscription budget#

You can turn all of this into a short AI subscription routine that any team lead can own. It asks for an afternoon, a spreadsheet and a refusal to trust a single number.

Here is an export of our AI usage by person and model for the last 30 days.
For each person, state: tasks completed, the model used, and the share of sessions that hit a limit.
Then list the three people most likely to be on the wrong plan, with one reason each.
Do not estimate dollar value from list prices. Flag any figure you cannot derive from the export.

The prompt asks the assistant to refuse the very shortcut that caused this week's confusion. It will not always manage it, so read the output the way you would read any analyst's note: with a pencil in hand.

Do all of that and the fox is no longer reading someone else's scoreboard. The quarry is a budget it can defend. It is reading its own tracks. Every provider will keep moving the ground, and the fox keeps up its patient prowl. The AI subscription that looks best on a Tuesday headline may not be the one that finishes your Friday work.

Frequently asked questions#