Two Flash Rows, One Price, 9,216 Tokens Apart: What the 393K Output Cap Actually Is

The OrcaRouter Deepseek API page lists eleven rows, and exactly two of them are named V4.1 Flash: `DeepSeek V4.1 Flash` and `DeepSeek V4.1 Flash 1017`. Both read $0.150 per million input tokens, $0.600 per million output, $0.0030 cached — identical to the cent [VENDOR: the provider’s published rates, passed through with no markup]. Both carry a 1M context window and both score 8 on the page’s own quality composite [OURS].

They differ in one cell. The first says 384K maximum output. The second says 393K.

Nine thousand two hundred and sixteen tokens is not a number anyone would put in a headline, and it is not one a product team would plan around. But it is a very specific number, and the specificity is the useful part: 9,216 is exactly 9 × 1024, which means the two caps are not a product decision at all. They are the same cap written in two different bases, and the table has rounded each one in the base it happened to arrive in.

The arithmetic, once

Underneath the table, the page carries the raw values: 384,000 and 393,216 [OURS].

• 384,000 = 375 × 1,024. As a binary figure this is 375 KiB.

• 393,216 = 384 × 1,024. As a binary figure this is 384 KiB.

So the newer row’s ceiling is exactly 384 binary kilobytes worth of tokens, and the older row’s ceiling is 375. The difference, 9,216, is precisely nine binary kilobytes. That is a rounding convention, not a capability.

The display follows the same convention downward: 384,000 ÷ 1,000 = 384.0, printed as 384K; 393,216 ÷ 1,000 = 393.2, printed as 393K. Both figures are decimal-rounded from their true value, so the column is internally consistent — it just isn’t consistent with the numbers that generated it.

That is the whole explanation. A reader comparing the two rows and concluding that the 1017 revision raised the output ceiling by two and a half percent is reading a unit artifact as a product change. The more useful conclusion is that someone upstream switched from expressing a cap in one base to expressing it in another, and the table faithfully propagated the switch.

Why this matters beyond two rows

A cap that moves by 9,216 tokens is trivial. A cap that a client reads and sizes against is not, and this page has eleven rows carrying caps in four different magnitudes: 384K (nine rows), 393K (one row), and nothing at all for the second free row [OURS]. If you are importing this table into a planner — a context-budget calculator, a chunker, a retry loop that splits a long generation — then a column that mixes 375 KiB and 384 KiB values without saying so will produce sizing decisions that differ by 2.4 percent for reasons nobody can see.

That is small enough to ignore once and large enough to matter at the boundary. A job that splits a document into 384,000-token chunks fits the older row exactly and overflows the newer one by zero; a job tuned to the newer row’s ceiling will send 393,216 and get rejected by the older row by 9,216 tokens. The failure only appears at the edge, which is the worst place for it to appear.

There is a second, larger reading in the same column. Every DeepSeek row on the page — nine of the eleven — reports 384K, in the 375 KiB sense. One reports 393K. The two free rows carry 384K. So the catalogue’s output ceiling has not changed across V4, V4 Flash, V4 Flash Vision, V4 Pro, V4 Pro 0813, V3, reasoner, V4.1 Flash and V4.1 Flash 1017, except in the one row where the base changed. A generation of models, one ceiling.

What did change, and where to look for it

If the output cap is a rounding artifact, what is the actual difference between the two V4.1 Flash rows? The page does not say, and it does not pretend to. Both rows carry the same rates, the same context window, the same cache price and the same quality score. What the row identifier adds is a date: `1017`, which the release notes attach to a revision dated 2026-10-17 [VENDOR].

That is the honest structure of a catalogue row: it tells you the price, the envelope and the score, and it leaves the version story to the release notes. When two rows differ only by a date suffix and a rounded number, the date is the signal and the rounded number is noise.

There is a practical consequence worth stating plainly. If you are choosing between the two rows, the price is not a tiebreaker and the output cap is not a tiebreaker; both are at the same level. What separates them is which version you want pinned, and that is a question about reproducibility rather than cost. A team that wants the October build pins `deepseek-v4.1-flash-1017`; a team that wants whatever the family currently resolves to uses the plain name.

The 2× window nobody puts next to the base rate

While the two Flash rows are identical on price, the price itself is not constant through the day, and this is the one place on the page where the table is systematically incomplete.

Nine of the eleven DeepSeek rows carry a time-based multiplier in the page’s underlying data, and the other eight provider catalogues carry none at all [OURS]. The windows are 01:00–04:00 UTC and 06:00–10:00 UTC, each with a 2× multiplier. Seven hours a day, and during those hours the printed $0.150 / $0.600 becomes $0.300 / $1.20.

Neither figure appears in the visible table. The table shows the off-peak base rate, which is the rate you pay for the other seventeen hours, and the page’s FAQ says nothing about a peak window at all. A reader who budgets from the table and runs their workload at 02:00 UTC will be off by a factor of two on every line, and the table will have been technically correct the whole time.

This is not a criticism of the page — a single rate column cannot hold two rates without doubling every row — but it is the thing to check before you size a budget off it. The peak window is 7 hours out of 24, so a workload spread evenly through the day pays the multiplier on 29 percent of its requests, which raises the blended per-token cost by roughly 29 percent above the printed figure. A workload that runs entirely inside those seven hours pays double.

What a 2.4 percent difference is worth in practice

Suppose you are generating the longest response the model allows — an unusual thing to do, but the ceiling exists for a reason, and long-form extraction, translation of a whole document and single-pass code generation all reach for it.

At 384,000 tokens of output, the older row bills 384,000 ÷ 1,000,000 × $0.600 = $0.2304. At 393,216 tokens, the newer row bills 393,216 ÷ 1,000,000 × $0.600 = $0.2359. The difference is $0.0055 per request at the absolute ceiling, and proportionally less at any realistic length.

Set that against the peak multiplier. The same request run inside a 2× window costs $0.4608 or $0.4719. The version difference is worth a fifth of a cent; the time-of-day difference is worth twenty-three cents. One of those is worth checking before you ship and the other is not.

That comparison is the practical summary of this whole page. The catalogue’s most visible columns — the rates — are the ones that do not move between these two rows. The columns that do move are either rounding artifacts (the output cap) or invisible in the table entirely (the peak window). Reading a catalogue well means knowing which column is carrying information and which one is carrying a formatting decision.

The lesson in the column, not the row

Most of what makes a price table useful is what it leaves out, and this page leaves out its peak window while including a rounded output cap. Both choices are defensible; together they set a trap for the reader who trusts the table’s precision.

The rule that follows is simple enough to carry to any catalogue: a number that differs between two otherwise identical rows is not automatically a feature. Before you treat a delta as a capability change, look at whether it survives a change of base. 9,216 does not; it is nine binary kilobytes wearing a decimal hat. The date suffix on the row does, and so does the time-of-day multiplier hidden behind the base rates — but only one of those three things is visible in the column.

Sourcing note: all rates, output caps and row counts in this article are read from the OrcaRouter DeepSeek provider page on 2026-10-21. The rates are the provider’s published prices passed through with $0 markup and the two output caps are the raw values carried by the same page [VENDOR and OURS respectively]; the peak-window multiplier of 2× across 01:00–04:00 and 06:00–10:00 UTC is carried in the page’s own model data and is DeepSeek-specific — the other eight provider catalogues carry no time-based pricing. The two cost figures at the ceiling are our own arithmetic on those rates. 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.

You May Also Like