SemiAnalysis Tests: Claude's $200 Subscription Buys $11K in API Value

The Measured Value

What SemiAnalysis did is dumb and honest: open a bunch of paid accounts on both subscriptions, then hammer them with specially designed prompts, burning one token type at a time. Input, cache writes, cache reads, output — cycling through, watching how many tokens each tick of the official usage meter consumes. Only counts once error converged within ±5%. Then convert to dollars at official API list prices.

The metric is called “API equivalent.” If a $200/month subscription’s quota would cost $11,726 at API pay-as-you-go rates, its API equivalent is $11,726. It’s not what you paid, nor the company’s cost — just a ruler for “what is this subscription actually worth.”

Results split into two layers. Flagship models are close: GPT-6 Astra on ChatGPT Pro 200 has an API equivalent of $2,897; Claude Fable 5.1 on Max 20x is $2,485. But Fable can only consume 50% of the plan’s quota — after burning through that $2,485, Claude still has half its quota for other models, while ChatGPT hits empty.

The gap opens on mid-tier models — the tier both sell as the “daily driver”: Claude Max 20x running Opus 5.5 has an API equivalent of $11,726; ChatGPT Pro 200 running GPT-6.1 Sol manages only $2,084. A 5.6x difference. The lower tiers look the same: Max 5x at $5,725 versus Pro 100 at $1,055; Claude Pro at $1,178 versus ChatGPT Plus at $211.

Some will call it unfair: Sol’s API list price is cheaper than Opus’s to begin with. SemiAnalysis also ran it by raw token count: 28.6B Opus tokens a month versus 10.2B for Sol — still a 2.8x gap.

The workload used was agentic — AI doing multi-step tasks on its own. SemiAnalysis used its own real September usage mix: 96.6% cached input. Before reusing this conclusion, check whether your token mix looks like that.

OpenAI Cut Quotas Last Week

SemiAnalysis also confirmed something: on September 29, OpenAI cut the $200 plan’s quota by nearly half, matching what Codex lead Thibault Sottiaux said on X.

Astra’s API equivalent fell from $5,734 to $2,897 — a 49% cut. Sol fell harder, from $5,904 to $2,084, down 65%. Sol’s token quota dropped 48%; even though GPT-6.1 Sol’s cached input got cheaper, the quota wasn’t topped back up.

Legacy users who bought before September keep old quotas until October 29, then move to the new ones. The new $500 plan looks pricey — its Astra allowance is only 21% more than the old $200 plan. Across Pro 100, 200, and 500, tokens per dollar are now identical.

Anthropic’s approach was subtler: after the Opus 5.5 price cut (20% off input/output, 60% off cache reads), Max-tier Opus quotas rose ~20% and Pro-tier ~50%. Neither company passed the full API price cut to subscribers; Anthropic just passed more.

Subscriptions Are the Most Subsidized Tier

SemiAnalysis did rough math. Anthropic’s subscriptions are ~10% of revenue but eat 41.7% of inference compute. Per megawatt of revenue, subscriptions contribute roughly $16M versus $102.9M for API. Subscriptions drag Anthropic’s overall per-MW revenue down by about $36M.

Margins look worse. Assuming 92% API gross margins, a Max 20x user maxing out Opus 5.5 quota runs at negative 369% margin; maxing Fable 5.1, it’s 1%. At 20% average utilization, those recover to 6% and 80%. All SemiAnalysis estimates — not Anthropic’s disclosed financials.

The pattern is simple: the pricier the model, the worse it looks inside a subscription. Max 20x running Fable 5.1 has an API equivalent of 12.4x the monthly fee; running Opus 5.5, it’s 58.6x. Anthropic collects the subsidy through implicit tier design; OpenAI just cut directly last week. SemiAnalysis’s read: for model tiers above Fable, subscription quotas will only shrink — possibly excluded from subscriptions entirely.

One aside: Chinese vendors still subsidize subscriptions. By API-equivalent per dollar, they average slightly below OpenAI’s ~12x level.

Three Caveats on the Conclusion

First, the 5x is a snapshot. SemiAnalysis says it themselves: quotas can change anytime. The September 29 cut is the proof. Any “what it’s worth” conclusion has a short shelf life.

Second, API equivalents distort. If a model’s list price is set high, the equivalent inflates; set low, it shrinks. Look at token counts alongside dollars.

Third, don’t mistake equivalent for workload. $11,726 is quota at list prices — not $11,726 of actual work done. Whether your workload is agentic, how much is cached input, and how much quota you actually burn all change the answer.