Thirteen Rows, One Ratio: Why Every anthropic models Row Prices Output at Exactly Five Times Input

Divide the output rate by the input rate on every row of the OrcaRouter anthropic models page and the same number comes back thirteen times.

$5.00 and $25.00. $10.00 and $50.00. $2.00 and $10.00. $4.00 and $20.00. $3.00 and $15.00. $1.00 and $5.00. Every pair divides to 5.00 [VENDOR: the provider’s published rates, passed through with $0 markup], and there is no exception in the catalogue — not on the two older 200K models, not on the flagship, not on the cheapest row.

No other catalogue on the site behaves this way. The OpenAI page spans 1.00 to 20.00 across its rows; the Qwen page 3.00 to 11.67; Z.ai 3.14 to 5.50; DeepSeek 3.00 to 4.00 [OURS]. A constant ratio across a whole provider is a design decision, and unpacking it is more useful than the number itself.

What a constant output ratio does to a comparison

The practical effect is that one number tells you the whole rate card. Name the input price and the output price follows: $6.00 in means $30.00 out, $7.50 means $37.50. Two of the five rate dimensions collapse into one, which makes this the easiest catalogue in the set to hold in your head and the hardest to accidentally misread.

It also means input price and output price can never disagree about which model is more expensive. On a catalogue with variable ratios, a row can be cheaper on input and dearer on output, and the “cheaper” label depends on your workload’s shape. Here it never does: Opus 5.5 at $4.00/$20.00 is cheaper than Fable 5.1 at $10.00/$50.00 on both columns, and cheaper than Opus 5 at $5.00/$25.00 on both, and there is no workload mix that flips the order.

The corollary is that a five-to-one ratio makes the output line dominant in almost any realistic workload. A million tokens of traffic at a typical 70/30 input/output split spends 30 percent of its volume on a line that is five times as expensive. Do the arithmetic on the Opus 5.5 row: 700,000 input tokens at $4.00 per million is $2.80; 300,000 output tokens at $20.00 is $6.00. Output is 68 percent of the bill on 30 percent of the tokens.

At a one-to-one ratio those two terms would be equal for the 70/30 split at equal rates; at five-to-one the output line is more than twice the input line, and every decision that shortens the answer saves roughly twice as much per token as the equivalent decision that shortens the prompt.

The catalogue’s three blocks

Read the thirteen rows as a price ladder and they sort into three blocks, all of them maintaining the ratio.

The $10 / $50 block. Claude Fable 5 and Fable 5.1, both quality 10 on the page’s own composite [OURS]. Two rows at the top of the ladder, identical rates, identical context and output limits, differing only in cached-input price — $1.00 against $0.250.

The $5 / $25 block. Opus 4.7, 4.8, 5 and 4.6, four rows at one rate. All four carry 1M context and 128K max output. Three score 10 and one scores 9.

The $2–$4 block. Sonnet 5 and 5.5 at $2.00/$10.00, Opus 5.5 at $4.00/$20.00, Sonnet 4.6 and Opus 4.5 at $3.00/$15.00, Sonnet 4.5 at $3.00/$15.00, Haiku 4.5 at $1.00/$5.00.

Within each block the prices are identical and the rows are separated by something else: cached rate, context window, output ceiling, quality score. That is the shape of a mature catalogue — the price is a tier label and the differentiation lives in the other columns. A reader who ranks these rows by price has ranked them into three groups and learnt nothing about which one to use.

The two columns that do separate the $5 block

Four Opus rows at $5.00/$25.00 is the clearest test of that idea.

Four rows, one rate card, and the envelope widens from 200K/64K to 1M/128K between the oldest and the rest. The quality score climbs 8 → 9 → 10 → 10 across the same span.

Everything about the price is identical; everything about the product is not. That is what a rate card looks like when the vendor has decided that model generation is not a pricing dimension — the same dollars buy a bigger window and a higher score, and the customer’s decision is entirely about capability rather than cost.

It also means the $5 block contains a latent upgrade: a workload sized against Opus 4.5’s 200K window has five times the room on 4.6 and later at the identical rate. That is the sort of migration a price comparison will never surface, because the prices do not differ.

What the ratio does not tell you

A constant output-to-input ratio says nothing about absolute cost, and the temptation to read it as a discount is worth heading off.

Five-to-one is not generous or stingy on its own. It is a ratio between two rates that both move with the model. Fable 5.1’s output at $50.00 per million is ten times Opus 5.5’s input at $4.00, and both figures are true at once because they describe different rows. The comparison that matters is between rows, not between columns.

The ratio is also not a statement about how many tokens a model emits. Two models at five-to-one can produce wildly different answer lengths for the same prompt, and on a fixed budget the chattier model costs more per task even at identical rates. The page cannot tell you that; only your own traffic can.

And the ratio says nothing about the cached rate, which is the one place this catalogue does diverge — eleven rows at ten percent of input, one at five, one at two and a half [OURS]. A model that is uniform on the output ratio is not uniform on the cache ratio, and the cache ratio is what determines the cost of a repeated prefix.

How to use a uniform rate card

Four habits, all of which fall out of the constant ratio.

Learn one number per row. Input price is the key and output is five times it. That halves the arithmetic in any comparison and removes the class of error where a cheap-input row turns out to be an expensive-output one.

Rank by envelope, not by price. With the ratio constant, price only tells you the tier. The columns that decide fitness are context, max output and the quality composite — and on this catalogue they vary within a tier by as much as five times on context.

Model the output share first. A five-to-one ratio makes the answer the dominant cost term. Before tuning anything else, estimate how many output tokens a task actually needs; a 20 percent cut in answer length is worth more than most input-side optimisations.

Read the cache column separately. It is the one column not governed by the rule, and the two rows that break the ten-percent pattern are the two newest additions — which is where a stale assumption will bite first.

A rate card that repeats one ratio thirteen times is easier to read than it looks, provided you stop reading it as a list of prices and start reading it as three tiers with four differentiating columns. The price tells you which tier; the rest of the table tells you whether the tier is the right one.

Sourcing note: every rate, context window, max-output value and cached-input price quoted here is read from the OrcaRouter Anthropic provider page on 2026-10-21, where the rates are the provider’s published prices passed through with $0 markup [VENDOR]. The observation that all thirteen output-to-input ratios are exactly 5.00, the three-block reading, the cross-provider ratio ranges and the monthly arithmetic are ours [OURS]. The 0-to-10 quality scores are our own composite, not a third-party benchmark. No third-party evaluation figures are used. Rates change; all values checked 2026-10-21.