SemiAnalysis tests same-price US$200 AI subscriptions: at API list prices, Claude's mid-tier allowance is worth about five times ChatGPT's — one firm's value conversion, not five times the work

SemiAnalysis, a firm that analyzes the economics of AI compute, said in an October 5 article that it sent repeated requests from paid accounts, measured how much each step of the usage meter consumed, and then converted that at API list prices into "how much API usage the plan is worth each month." Comparing the mid-tier model each vendor offers as its everyday workhorse — Opus 5.5 on Claude, GPT-6.1 Sol on ChatGPT — the Claude plan came out at roughly five times the value. "Mid-tier" here means the everyday-workhorse tier; it doesn't mean the two models are equally capable. For the top-tier models, the two plans convert to about the same value, because on the Claude plan, Claude's top-tier model, Fable 5.1, can use at most half of the plan's allowance on its own (SemiAnalysis, 2026-10-05). So the five-times figure applies only to teams whose main model is mid-tier, not to teams that mostly run the top-tier model.

Separately, the tech news site WinBuzzer, relaying OpenAI's plan notes, reports that new subscribers to the US$200 plan from September 29 get half the previous usage, while existing subscribers keep their old allowance until October 29 (WinBuzzer, 2026-09-30). Whether the five-times comparison used ChatGPT's allowance before or after the cut changes the multiple, and we haven't checked which.

Verification: the five-times figure and the half-allowance cap come from SemiAnalysis alone, measured on a small number of accounts; we read the public sections, not what sits behind the paywall. The OpenAI allowance cut is relayed by WinBuzzer; we didn't read OpenAI's own plan page. ⚠️ The five-times figure prices usage at list rates; it does not mean five times as much work gets done: different models burn different amounts of usage on the same task, and list prices themselves distort the conversion, as the authors acknowledge; the article is also selling the firm's own paid dashboard. ⚠️ Our analysis was produced with help from Anthropic's models, and Anthropic is a party to this item.

Judgment update: our August 8 issue, in an item about a self-declared AI supporter who said the current price can't hold, argued that if heavy users really are being subsidized, the correction would come as vendors tightening their plans, not as companies cutting budgets. OpenAI's allowance cut is the first case pointing that way, though WinBuzzer's account doesn't say why OpenAI made the cut. That issue also didn't anticipate that vendors would tighten in different ways: Anthropic limits only its top-tier model, and on September 22 its chief product officer, Mike Krieger, announced higher five-hour limits for subscribers (Mike Krieger, 2026-09-22). That is a counterexample, but to the reading that "every vendor will tighten across the board." Our current read: tightening will happen, but it lands in different places — OpenAI cut the whole plan for new subscribers; Anthropic capped only its top-tier model, and with Krieger's higher five-hour limits for subscribers, mid-tier use actually got looser. We're keeping both readings on the books and will settle them against the condition below.

What would prove this wrong: Anthropic raises mid-tier model allowances across the board before the end of 2026 without any tightening; verdict date December 31, 2026.

OpenAI announced on October 5 that, to meet the EU AI Act's requirement that generated text be identifiable in a machine-readable way, it is moving in three steps. API customers can opt in to watermarking on some models starting now, with the default still off. In the coming weeks, text output from ChatGPT and from Codex, its coding tool, will carry an invisible watermark for eligible users in the EU; this is not a global default. And the detection tool is open at first only to approved researchers and professional organizations that apply (OpenAI, 2026-10-05). The technique, called textGrain, adds an invisible statistical preference as the model picks each word, and the detector looks for that preference.

A token is a chunk of text a few characters long; models read and write text, and bill for it, in tokens. OpenAI's own published limits: with the false-positive rate (the share of human-written text wrongly flagged as machine-written) set at 1%, for text like psychology writing, about 80% of 200-token passages are detected, and about 95% of 400-token passages. Text with few word choices, such as math, scores noticeably lower. In a separate word-swap test, replacing 10% of the words in a 400-token passage with synonyms drops the detection rate from about 92% to 66%; replacing 25% leaves just 17%.

Verification: we read OpenAI's announcement itself, not the technical report. Every number is OpenAI's own measurement with no third-party check; its claim that performance "matched or exceeded" SynthID, Google's text watermark, is also self-assessed.

Judgment update: This doesn't change any existing judgment, but it adds one reading. By OpenAI's own numbers, "no OpenAI watermark detected" in the EU can't be taken as evidence that a person wrote the text, because changing a quarter of the words mostly defeats detection. Two things to watch: whether a third party checks these detection rates, and whether the detection tool opens to anyone beyond researchers.

Our October 5 issue covered the setup before the hearing; what's new today is the hearing itself. It took up "preconsidered" bills: drafts that haven't been formally introduced. The heaviest provision: an AI model that hasn't passed third-party validation, or that a human operator can't shut down, could not be sold or deployed in New York City; as the previous issue noted from the Council's records, the formal introduction date is October 8.

CBS New York's on-the-scene report from the evening of October 5: all four companies testified by video — Logan Graham for Anthropic, Morgan Dwyer for OpenAI, Alice Friend for Google and Shane Cahill for Meta. SpaceX was subpoenaed and wrote back that it would cooperate but not appear; the city is taking legal action. There was no vote (CBS New York, 2026-10-05). The previous issue, following amNewYork, said the subpoena went to "Elon Musk's AI company"; CBS names SpaceX. In this issue we haven't checked how the two descriptions relate. Speaker Julie Menin asked at the hearing: "How would a kill switch work? Would that be effective?" She also said the companies' inability to quantify how large the risk is was "troubling at best."

Verification: For CBS, we read the key excerpts on the page, not the full report. Of CNBC's same-day report we read only the headline; a search summary mentioned an Anthropic whistleblower testifying at the hearing, but we didn't read the original and don't use it. ⚠️ In the passages we read, no company stated support for or opposition to the third-party validation bill above.

Judgment update: the previous issue logged the working read that "who must sign off before a model goes to market may first be written into law by a city or state." Our confidence in it is unchanged. Before a rule like this reaches the statute book, we'd expect a signal first: a vote, or a company taking a position. Today's hearing had neither: no vote, and in the passages we read, no company position on the bill.