Same-price Claude plans buy five times OpenAI's mid-tier API usage in SemiAnalysis meter tests

OpenAI's cut is one of several changes behind the fivefold figure ^[10][11]. Over the summer, SemiAnalysis estimated that the $200 Pro plan, which carried 20 times the Plus allowance, was worth as much as $14,000 a month in API spend when fully used, compared with about $8,000 for Claude Max ^[1]. Thibault Sottiaux, who leads Codex at OpenAI, posted the new terms at 2:41 a.m. ET on September 29. "In effect, if you do the math, it will net out at half the dollar in API spend compared to the old Pro $200 plan," he wrote ^[7]. Halving the summer estimate leaves about $7,000, roughly $1,000 below Claude Max ^[21]. The report does not say which model the summer estimate assumed.

Cheaper tokens shrink a dollar comparison if allowances do not go up. Sol's cache-read price on GPT-6.1 was cut in half, but its token limits on the $200 plan did not change, and that lowered Sol's API-equivalent value by roughly 30% ^[10].

API-equivalent value is what a plan's permitted tokens would cost at the lab's public per-token prices, using SemiAnalysis's own September usage mix for agent workloads ^[12]. To read the meters, analysts fed chunks of War and Peace into paid accounts, counted tokens between meter movements and discarded incomplete steps until the range narrowed to plus or minus 5% ^[14]. Because the five-hour window resets many times a week, the weekly caps set the monthly totals ^[20]. After 500 million cache-read tokens produced no meter movement, the analysts treated those reads as free ^[15].

Implicator, which reported the study, names the obvious objection: Sol costs much less per token than Opus, and that affects the conversion ^[2]. The study also compares raw token volumes, a check that leaves API prices out, and still finds a large gap in Claude's favor ^[2]. Implicator adds that the dollar figure does not prove every subscriber finishes five times as many tasks ^[3].

The figures come from a single firm with a product built on them. SemiAnalysis's Subscriptions Dashboard is open only to paid subscribers of its Tokenomics Model, and its comparisons involving Cursor and Devin sit behind the article's paywall ^[18].

Renewal is this quarter's decision; whether the gap lasts is next quarter's. In my view the second question turns more on Anthropic's costs than on OpenAI's pricing. SemiAnalysis models subscriptions at about 10% of Anthropic's revenue but more than 40% of its inference compute ^[19]. On that model, subscribers consume four times the share of compute that their share of revenue would suggest ^[23].