Deepseek API: Nine Models and Two That See

The smallest number on the DeepSeek page is 2. The catalogue lists 9 models, and 2 of them can see — the rest are text-only. On OrcaRouter, the Deepseek API page prints that census in one FAQ line: across the catalogue, vision on 2, tool calling on 8, reasoning on 8 and JSON mode on 9. Read that distribution and the catalogue describes itself: structured text out and hard reasoning are the point, and image understanding is an edge shelf with two entries.

The other numbers on the page line up behind that reading. The header lists 9 models and a median p50 of 1816 ms [OURS], with a max context of 1M [VENDOR]. What the header does not list is a cheapest-input statistic, and what the page does not carry is a cost FAQ or a cost calculator — it is the only provider page in this set with none of the three [OURS]. That is a shape difference, not a defect, and it changes the reading order: capability first, then price by hand.

Why the census is the strongest column

A model count tells you a shelf exists; the census tells you what it is for. DeepSeek’s counts are unusually tight — JSON mode on all nine rows, tool calling and reasoning on eight, vision on two — and the tightness is the information.

JSON mode on all nine rows means structured output is the floor of the catalogue, not a feature of its upper tiers. Every model here can be asked for typed, schema-shaped responses, which makes the page the natural shelf for extraction, classification and structured generation work. Tool calling on eight is a hair narrower and removes one row from a tool-using agent’s shortlist. Reasoning on eight does the same. The counts compress the decision: a text-and-structure workload keeps essentially the whole shelf in play, and only vision cuts it.

Vision on two rows is the number that matters most, precisely because it is small. For any job that must take an image as an input, the nine-model catalogue narrows to two entries before a single price is read, and the other seven rows are out of the running no matter what they cost. That is not a criticism of the catalogue — a vendor is entitled to decide vision is a niche — it is the single most useful planning datum on the page. It tells you exactly where this shelf’s coverage ends.

The pricing side of the page cooperates with that reading, in its own way. The eight token-billed rows run from $0.150 to $0.660 per million input, a spread of only 4.4 times [VENDOR], which is gentle next to the three-order-of-magnitude spreads on some of the other provider pages. On a narrow spread, price stops being a filter and becomes a tiebreaker, and the version and capability columns carry the choice instead. The catalogue is telling you: within DeepSeek, the decision is which model does the job, not which one you can afford.

The nine rows, and the one that bills per request

The pricing table lists each model with an input and an output rate per million tokens, a context figure and a quality digit. The eight token-billed rows:

Rates are the provider’s published per-million-token figures as printed on the page [VENDOR]. The quality column is a 0-to-10 composite we build ourselves from the dimensions we grade each model on — it is our own measure, not a third-party benchmark score, and we label it that way wherever the number appears [OURS].

The ninth row bills per request rather than per token: DeepSeek V4 Flash (Free) prints $0.00 per request, which is a per-request billing mode, not a promise of free tokens — the request price is zero and the tokens inside it are not separately metered [VENDOR].

Notice the two vision rows. DeepSeek V4 Flash Vision (Exp) is one of them, sitting in the flash band at $0.220 / $0.660 with the vision capability that the plain V4 Flash and the 0731 snapshot lack. The second vision-capable row is the reason the census counts two rather than one, and it is the row a vision workload actually shortlists alongside it.

Reading the page in the order it asks for

Because there is no cost calculator, the page hands you a rate card and leaves the multiplication to you, and because the spread is narrow, the arithmetic stays cheap. Take the same modest shape the other provider pages use — 10 million tokens a month, 70% input and 30% output:

• DeepSeek V4.1 Flash at $0.150 / $0.600 — $1.05 plus $1.80 = $2.85 a month.
• DeepSeek V4 Flash Vision (Exp) at $0.220 / $0.660 — $1.54 plus $1.98 = $3.52 a month.
• DeepSeek V4 Pro at $0.660 / $1.98 — $4.62 plus $5.94 = $10.56 a month.

The whole shelf spans under $8 a month at that volume, which is the point of the narrow spread made concrete: for a text-and-structure workload, the model choice inside DeepSeek barely moves the bill, so the capability cut and the quality digit decide. For a vision workload, the arithmetic is even shorter, because the census already cut the shelf to two rows before the multiplication started.

The two prose registers on the same page

The page mixes two registers, and quoting them as one number is the standard error. The context column of the pricing table prints 1M for every row — the table’s short form [VENDOR]. The FAQ answers the largest-context question in spoken prose: DeepSeek V4.1 Flash, at 1048K tokens. Both are true; they are not the same figure. Quote the register you are reading — 1M from the table, 1048K from the FAQ — and never splice the two into one sentence.

The largest-context answer and the vision answer share the FAQ block, and together they round out the census. The model with the biggest window is DeepSeek V4.1 Flash, the same row that opens the price table at $0.150, which makes the cheapest text row also the longest-memory row — the output and context columns do not reorder the list the way the input column orders it.

The one row behind the vision shelf

The model page for DeepSeek V4 Pro is the text backbone of the shelf, and reading it next to the census completes the two-step method: the catalogue sorts, the model page prices and measures one row. The page states that DeepSeek V4 Pro is text-only, prints its own rate card at $0.66 per million input tokens and $1.98 per million output, and quotes a p50 time-to-first-token of 1.84 seconds on production requests — 1840 ms on the catalogue’s scale [OURS] [VENDOR]. One self-consistency note, in the spirit of reading the page closely: the model page’s rate card prints $0.66 / $1.98 while the same page’s body text states $0.44 / $0.87 in four places [VENDOR]. The two figures cannot both describe the same rate, and a careful reader has to pick one — which is exactly the kind of internal contradiction the page itself prints without comment.

The takeaway

The DeepSeek page’s smallest number is its most useful one. Two of nine models see, eight reason and call tools, all nine produce structured JSON — so the catalogue’s center of gravity is text and structure, and the two vision rows are the whole answer for anyone whose job needs an image. Read the census before the prices, keep one context register per sentence, and treat the missing calculator as an invitation to run the arithmetic yourself against your own volume. At a spread this narrow, the bill barely moves; the capability cut decides. For more information, visit our website.Edit