Anthropic and OpenAI's $200 AI Plans Buy an Editable Amount of Work

Your $200 AI subscription doesn't buy $200 of usage. It buys whatever the vendor's internal credit table says that week, and the table is editable.

That is the mechanic behind SemiAnalysis's claim that Anthropic subscriptions offer "5x+ more value" than OpenAI's. The number is the output of the analyst's own model, and that model is the very product SemiAnalysis sells through its Tokenomics Model and Subscriptions Dashboard. Strip the marketing claim and what remains is more useful: a way to read any AI plan, not just these two.

How credit systems diverge from public prices

Both Anthropic and OpenAI sell subscriptions as monthly access to a bundle of model usage, but the bundle is metered by an internal credit whose cost is set per model, per workload. Public API rates and credit cost ratios can move in opposite directions, because the credit table can be tweaked independently of the price card.

That is the (plan, model, workload) tuple the source leans on. The same $200 sticker can mean a heavy research run on Claude Sonnet, a thrifty batch job on Haiku, or a Codex-style coding agent (OpenAI's autonomous coding assistant) that eats credits fast. The dollar price on the Claude pricing page, Pro at $17 per month on annual terms ($200 paid up front) or $20 per month month-to-month, with the higher-tier Max plan starting at $100, is not a usage budget. It is a starting meter reading.

What the operating numbers show

SemiAnalysis's model puts Anthropic subscriptions at roughly 10% of overall revenue, while consuming over 40% of inference compute (the GPU power needed to actually run model queries) and lowering revenue per megawatt, or dollars earned per unit of inference power and the lab's core efficiency measure, by around $36 million. In plain terms: the plans are a worse deal for Anthropic per unit of power than its API business, and a better deal for the customer per dollar than the API card alone would suggest.

The same dynamic is more material at OpenAI, where subscriptions are a larger share of total revenue. A credit-cost nerf at OpenAI hits the income line harder, which is one reason OpenAI has been more deliberate about how it walks the line.

The reset/nerf cycle

Credit tables are not stable. Labs change limits through promos, new model releases, and what SemiAnalysis calls silent credit-cost tweaks. OpenAI's recent generous resets have been linked to a surge in Codex adoption, and they have forced Anthropic to repeatedly walk back planned subscription nerfs to avoid backlash.

Independent reporting on the next reset window backs that up. An X post from commentator thsottiaux describes GPT-6 Sol and Luna launching with a 50% API price cut and a banked reset loaded into Plus, Pro, and Business accounts. That is not an OpenAI official statement; it is a customer read of release notes. But it shows the shape of the cycle: a new model arrives, the credit math shifts, and existing subscribers either get a free top-up or watch their effective usage fall.

Where the "5x" comes from, and what to do with it

SemiAnalysis's headline number is the output of a model that compares credit cost ratios across plans, not a public benchmark. The firm also sells a Subscriptions Dashboard covering OpenAI, Anthropic, Meta, SpaceXAI, Cursor, Cognition, Z.ai, MiniMax, and Moonshot, with new providers added on release. The dashboard is useful. The number that headlines it is also the number that sells it.

The 5x claim is best treated as a directional signal: same dollar, different workload, depending on plan, model, and week. It is not a verdict, and it is not a price.

A reader framework

The audit is straightforward. Find the model you actually run on your plan and read its current credit cost; the Claude pricing page shows the plan price, not the workload price, and credit costs change with each model release. Cross-check against the release calendar: OpenAI and Anthropic both use model launches to reset the math, and the SemiAnalysis newsletter tracks the cadence in detail. Then ask whether you are paying for usage you cannot use, or using more than you are paying for. The third-party ChatGPT pricing reference is a snapshot, not the live number.

The next reset is the watch item. If thsottiaux's read of the GPT-6 launch holds, OpenAI Plus, Pro, and Business accounts will see a banked credit top-up alongside the API price cut. Anthropic's response, on the same cycle, has historically been to walk back its own planned nerfs within a week or two.

The credit table is editable. The audit is the same: read the credit cost for the model you actually run, on the day you actually run it.