IMPLICATOR
Andrew Megalaa’s subscription tests needed a book. For the SemiAnalysis study bearing his byline, analysts fed chunks of War and Peace into paid Claude and ChatGPT accounts, watching usage meters tick from zero toward 100%. The novel supplied filler the models would accept. Large blocks of gibberish could trigger refusals.
The October 5 study found roughly five times the API-equivalent dollars on same-price Claude and OpenAI plans across tiers, comparing Opus 5.5 with GPT-6.1 Sol, the mid-tier models both companies market as daily drivers.
The result arrives after OpenAI halved its $200 plan’s allowance. The figure converts usage into dollars at list prices; it does not prove every subscriber finishes five times as many tasks.
The Breakdown
- SemiAnalysis's October 5 meter tests found same-price Claude plans give roughly five times the API-equivalent value of OpenAI plans when comparing Opus 5.5 with GPT-6.1 Sol.
- OpenAI halved the API-dollar value of its $200 Pro plan on September 29; existing subscribers keep the old limits through October 29.
- On flagship models, Fable 5.1 and GPT-6 Astra, SemiAnalysis calls the limits quite similar, so the fivefold gap is a mid-tier result.
- SemiAnalysis models subscriptions at about 10% of Anthropic's revenue but over 40% of its inference compute.
AI-generated summary, reviewed by an editor. More on our AI guidelines.
How the meter was read
“API-equivalent value” means what the tokens allowed by a subscription would cost at the lab’s public per-token prices. The study uses SemiAnalysis’s own September usage mix for agent workloads. Neither lab publishes absolute allowances in tokens or dollars; OpenAI supplies relative multipliers.
SemiAnalysis sells these measurements. Its Subscriptions Dashboard is available only to paid Tokenomics Model subscribers, and its comparisons involving Cursor and Devin sit behind the article’s paywall.
Random tags isolate fresh input; fixed tags test cached text, and technical essays force long outputs because models refuse requests to “repeat SemiAnalysis 100,000 times.” Analysts count tokens between meter movements, discard incomplete steps and subtract incidental costs until the range narrows to plus or minus 5%. After 500 million cache-read tokens produce no meter movement, they assume those reads are free.
The measured rates produce token allowances for each window. The five-hour window resets repeatedly within a week, so the weekly caps determine monthly totals. Multiplying those totals by blended API prices produces the dollar comparison.
FREE WEEKDAY MORNING BRIEFING
Track what your AI subscription actually buys.
The Implicator Morning Briefing filters the AI news cycle to the stories worth your attention and explains their consequences. From San Francisco, every weekday at 4:45 a.m. Pacific, 7:45 a.m. Eastern.
About five minutes. No hype. No spam.
Sol costs much less per token than Opus, which affects that conversion. Yet the October study also compares token volumes and finds a large gap in Claude’s favor. That cross-check removes API prices from the comparison.
The tests detected variation within a provider: one of three identical subscriptions had roughly 20% lower limits. The unnamed provider confirmed an “extremely tiny” A/B test, saying it “didn’t just decrease limits wholesale” and was testing “how to better balance when people hit limits.” SemiAnalysis says this shows a provider can silently change limits.
What OpenAI changed
At 2:41 a.m. ET on September 29, Thibault “Tibo” Sottiaux, who leads Codex at OpenAI, posted the warning: “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.”
SemiAnalysis’s meter tracking confirmed the halving. Eligible existing Pro $200 subscribers keep their old limits through October 29. The new $500 monthly plan supplies only 21% more GPT-6 Astra than the old $200 plan. Over the summer, SemiAnalysis had estimated that the fully consumed old $200 Pro plan, with 20 times the Plus allowance, could supply up to $14,000 a month at API prices, against roughly $8,000 for Claude Max.
A separate price change reduced the dollar comparison further: cheaper tokens shrink API-dollar value unless allowances rise. GPT-6.1 Sol’s cache-read price fell 50%, while its token limits on the $200 plan stayed unchanged. That lowered its API-equivalent value roughly 30%. Anthropic increased Opus allowances with the 5.5 release by about 20% on Max and 50% on Pro, insufficient to offset its API price cut.
On September 10, Sottiaux had called Astra demand “really unprecedented” and said Pro $200 users “put the most strain on our systems.”
After Sottiaux’s announcement, @MKantautas wrote on X: “Ok guys, back to claude. RIP openAI.”
Know someone who'd find this useful? ✉️ Email it to a friend in one click, or they can subscribe free here.
The case for OpenAI
Sottiaux’s argument concerns completed work. Cheaper, more efficient models should let developers “still get more work done than if you were on the Pro $200 subscription one month ago.” He said OpenAI would keep cutting API prices rather than inflate list prices to make subscriptions look generous.
OpenAI Pro plans also have no five-hour cap, allowing subscribers to consume their weekly allowance in long sessions. SemiAnalysis calls this the only counterargument and says it does not offset the Opus advantage.
On flagship models, SemiAnalysis calls the limits “quite similar”, so the fivefold gap is a mid-tier result. Anthropic restricts Fable to half a plan’s allowance. In the October measurements, a $200 Claude plan retained half its allowance after $2,485 of Fable 5.1 usage at API prices. OpenAI’s equivalent plan was exhausted after $2,897 of Astra. Sottiaux’s comments came before the study and do not address it; the finding rests on SemiAnalysis’s own measurements.
How each lab trims the subsidy
The October Tokenomics Model estimates that subscriptions generate about 10% of Anthropic’s revenue while potentially consuming over 40% of inference compute, costing roughly $36 million in blended revenue per megawatt, in SemiAnalysis’s model.
Anthropic assigns lower API-equivalent allowances to premium models. The study’s margin calculations assume 92% API gross margins. At full subscription use, consuming the maximum Opus 5.5 allowance yields a modeled gross margin of minus 369%, compared with 1% for Fable 5.1. At an assumed average consumption of 20% of the allowance, those margins become 6% and 80%, respectively.
These are modeled subscription margins, dependent on consumption and the assumed cost of serving API traffic. The API-dollar allowances themselves also change whenever either lab cuts its public token prices. OpenAI’s approach cuts allowances across model tiers, a move SemiAnalysis calls “the nuclear option.”
SemiAnalysis writes: “OpenAI used to be lauded by indie developers for their generous subscription limits compared to Anthropic, but this is simply no longer true today.”
Frequently Asked Questions
What does API-equivalent value mean?
It is what the tokens a subscription allows would cost at the lab's public per-token API prices. SemiAnalysis converts measured allowances into dollars using its own September usage mix for agent workloads.
How did SemiAnalysis measure the limits?
Analysts fed chunks of War and Peace and technical essays into paid accounts, counted tokens between usage-meter movements and repeated runs until the range narrowed to plus or minus 5%.
Does the 5x gap apply to every model?
No. The fivefold figure compares Opus 5.5 with GPT-6.1 Sol. For the flagship Fable 5.1 and GPT-6 Astra, SemiAnalysis calls the limits quite similar, though Anthropic caps Fable at half a plan's allowance.
What did OpenAI change on its $200 plan?
OpenAI said on September 29 that the reopened Pro $200 plan nets out at half the API-dollar value of the old one. Existing subscribers keep their old limits through October 29, and a new $500 plan offers 21% more Astra than the old $200 plan.
What is OpenAI's counterargument?
Thibault Sottiaux says cheaper models let developers get more work done than a month ago, and OpenAI Pro plans have no five-hour cap. SemiAnalysis says the missing cap does not offset the Opus advantage.
AI-generated summary, reviewed by an editor. More on our AI guidelines.
Related stories
Free AI briefing · Weekdays
The AI stories that matter, sourced and explained.
Join the Morning Briefing. It goes out every weekday at 4:45 a.m. Pacific, with later sends for the East Coast, Berlin and Tokyo.
Free when you sign up: The Paperclip Compendium, our tested guide to running AI agents.
Free. Unsubscribe in one click.
Marcus Schuler
Editor-in-Chief and founder of Implicator.ai. Former ARD correspondent and senior broadcast journalist with 10+ years covering tech. Writes daily briefings on policy and market developments. Based in San Francisco. E-mail: editor@implicator.ai
