What $200 a Month of AI Buys, and What Changes It Most

How you call a model API moves what $200 buys by roughly an order of magnitude, with the model held constant. The conversion table below turns each vendor-published rate into input tokens at a fixed $200: Claude Opus 5 alone spans 40M tokens at the list rate to a derived 800M through the batch-plus-cache lane, a 20x range with the model held constant.

Two definitions carry the table. Cached input is not a discount you negotiate; it is a different price for a different operation — re-reading a prompt prefix the provider has already processed, billed at 0.1x the base input rate wherever the vendors below publish the multiplier. Batch is not a cheaper model; it is an asynchronous lane — the same weights answering within a service window instead of now, at 50% off both input and output where it exists.

Every cell below is the same arithmetic: $200 divided by a vendor-published price per million input tokens. The rates themselves live in our maintained per-million-token price index; this page inverts them into the question a budget line actually asks — how many tokens does a fixed $200 buy, and which lever changes that most. Rows marked published cell are combined prices the vendor prints itself; the one row marked derived is our multiplication of two published multipliers, never a vendor figure.

* Derived, not vendor-published: 0.5 (batch) × 0.1 (cache read) × $5.00 base input = $0.25 per million tokens. Anthropic’s FAQ confirms the two discounts combine, but the company never prints this cell the way OpenAI does one row up. † Alibaba documents cache and batch as two separate rules and nowhere confirms they stack, so no combined Qwen figure appears in this table — multiplying them would invent a number the vendor has not stated. The list lanes for GPT-5.6 Sol, Gemini 3.7 Flash and GLM-5.3-Flash are promotional, with vendor-announced end dates of November 21, 2026, December 31, 2026 and September 9, 2026 respectively.

This page publishes one derivation and one borrowed axis, and it is deliberate about which numbers are whose. The rates belong to the vendors, the quality scores belong to Artificial Analysis, and the only thing that is ours is the inversion — tokens per fixed budget — plus the matrix of which levers each vendor publishes.

The conversion table only works because each vendor publishes a different set of levers, and the differences are structural, not cosmetic. Anthropic and Alibaba publish nearly identical percentages — cache reads at one tenth of base input, batch at half price — yet only Anthropic confirms in writing that the two combine. OpenAI goes further and prints the compounded cell in its own batch table. Google publishes both levers but states that batch context caching is billed at the same standard cache rate, so there is no extra stack to compute. DeepSeek has no batch lane at all and instead publishes the survey’s deepest lever: a full cache-by-time-of-day price matrix.

The mechanics behind the two big levers are covered in depth elsewhere: our prompt caching engineering guide explains what a cache read physically is, and the half-price batch lane maps who offers batch and on what terms. This page takes both as given and asks only what they do to a fixed budget.

Hold the model constant and walk the lanes. Claude Opus 5 at $200 buys 40M input tokens at list, 80M through batch, 400M through cache reads, and a derived 800M with the two combined — the same weights answering the same prompts across a 20x range. Now compare across models at the list rate only: Opus 5’s 40M against Qwen3.8-Max’s 100M is a 2.5x gap, and against Gemini 3.7 Flash’s promotional 267M the gap is under 7x. Step further out and the between-model range is wider still: GLM-5.3-Flash’s promotional list rate returns 2,667M for the same $200. Both axes move the answer by more than most buyers assume, and only one of them appears on a pricing page.

DeepSeek makes the same point with printed cells rather than derivation. V4-Pro’s peak cache-miss rate is $1.32 per million tokens; its off-peak cache-hit rate is $0.022. Both are published table cells, and the ratio between them is 60x, entirely inside one model. At $200 that is the difference between 151.5M tokens and roughly 9.1B. The off-peak window is generous, too: DeepSeek’s definition leaves every hour outside two weekday UTC bands, and all of the weekend, at the lower rate.

The practical reading order for a $200 budget follows directly. First ask which of your traffic is re-read prefix — system prompts, tool definitions, shared context — because the 0.1x cache-read multiplier is the single largest lever most vendors publish. Second, ask what fraction of your work can wait for an asynchronous window, because batch halves both directions where it exists. Only third ask which model, because by then the same model’s price has already moved by an order of magnitude under your first two answers.

Tokens per dollar mean nothing without a quality axis, and the quality axis is not ours. The scores below are Artificial Analysis’s Intelligence Index, retrieved for this dataset from its public leaderboard, at the highest reasoning effort AA lists for each model. The ranking is AA’s judgment, not Digital Applied’s; we pair it with the cheapest lane each vendor publishes and draw one observation from the pairing.

The observation the pairing earns: the quality gap between the top and bottom of AA’s table is fifteen index points, and the budget gap between lanes routinely dwarfs what that quality gap costs. Claude Opus 5’s derived batch-plus-cache lane ($0.25 per million) is cheaper than Gemini 3.1 Pro Preview’s list rate ($2.00 per million) — eight times cheaper, for a model AA scores fifteen points higher. On these numbers, learning to use a vendor’s published levers buys more capability per dollar than shopping down the quality table does.

There is a second way to spend $200 a month on frontier AI, and it deliberately resists this table. Claude Max at $200 is described in Anthropic’s help center only as “20 times more usage per session than the Pro plan” — no token count, no message count. ChatGPT Pro at $200 is described in OpenAI’s help center the same way: “Pro $100 unlocks 5x higher usage than Plus, while Pro $200 unlocks 20x usage than Plus.” OpenAI adds that “There is no setting to increase or bypass a model’s usage allowance.”

This is a finding, not a research gap. Both help-center articles were checked specifically for a token or message figure, and neither publishes one — the multiplier framing is the vendors’ own choice. The $200-a-month plans therefore cannot be merged into the $200-of-API table above, and any post that does merge them is inventing the conversion the vendors declined to publish. How those multipliers have shifted over the year is tracked in our coding-plan limit change ledger, and the broader included-versus-metered divide is the subject of our subscriptions-versus-API comparison — background here, because the conversion table has already made this page’s point without it.

One more absence worth stating precisely: no pricing page collected for this dataset publishes a prepaid volume-discount rate. Anthropic’s page says volume discounts “are negotiated on a case-by-case basis,” and its usage tiers govern rate limits, not price. If a fixed $200 is your whole lever budget, the published machinery — cache, batch, and DeepSeek’s clock — is the machinery there is. Mapping which of your workloads can actually ride those lanes is the kind of question our AI transformation engagements open with.