DeepSeek Models

Q: Which DeepSeek model should I start with?

deepseek-v4.1-flash at $0.1408/M input — the cheapest entry here that declares tool calling, and it carries a 1M context. Move up to deepseek-ai/DeepSeek-R1-Zero when answer quality matters more than cost, or to deepseek-v4-flash for long-form reasoning.

Q: How is cached input billed?

The Cache read column is the rate for input tokens served from the prompt cache — for example deepseek-v4-flash-vision-exp bills cache hits at 2% of the input rate and deepseek-v4.1-flash bills cache hits at 2% of the input rate. Cache write is the surcharge for putting a prompt into the cache in the first place, and only a few upstreams bill it separately. A dash in either column means the catalog carries no cache rate for that model, so plan on paying the full input rate.

37 modelsGeneral models from $0.1408/M inputUp to 1M context

Usage

Last 30 days · 2026-10-02 to 2026-10-31

Tokens

980B

Requests

9.6M

Models in use

32 of 37

Tokens per day, stacked by model

deepseek-v4.1-flash
deepseek-v4.1-flash-1017
deepseek-v4-flash
deepseek-v4-flash-0731
deepseek-v4-pro-0813
deepseek-v4-pro
deepseek-v4-flash-vision-exp
deepseek-v3.2
7 more models

Which models that traffic went to

Share of 980B tokens. 17 models with traffic report no token counts and cannot be ranked here, including deepseek-r1-distill-llama-70b and DeepSeek-V3.1-Think — they are in the request view.

The two views disagree on purpose: a model can take a large share of the calls and a small share of the tokens — many short requests — or the reverse. Which one matters depends on whether your cost is driven by call volume or by prompt length. Measured on AIHubMix over the last 30 days, counting the 37 model IDs listed on this page; traffic routed through upstream-specific IDs that are not in the public catalog is not included.

All 37 DeepSeek Models

Open in model list

DeepSeek models on AIHubMix with input and output modalities, context length, maximum output, price per million tokens including cache read and cache write rates, and measured throughput and latency.
	Modalities
deepseek-v4.1-flash	Takes text, vision, returns text.	1M	384K	$0.1408$0.5632/M	$0.0028/M	—	82 tok/s	1.11 s
deepseek-v4-flash	Takes text, vision, returns text.	1M	393K	$0.142$0.284/M	$0.0284/M	—	133 tok/s	1.24 s
deepseek-v4-flash-0731	Takes text, returns text.	1M	384K	$0.142$0.284/M	$0.0284/M	—	76 tok/s	0.47 s
deepseek-v4-flash-vision-exp	Takes text, vision, returns text.	1M	393K	$0.155$0.62/M	$0.0031/M	—	38 tok/s	1.01 s
deepseek-v4.1-flash-1017	Takes text, vision, returns text.	1M	393K	$0.155$0.62/M	$0.0031/M	—	131 tok/s	0.85 s
deepseek-v4-flash-0731-fast	Takes text, returns text.	1M	384K	$0.28$1.4/M	$0.07/M	—	42 tok/s	0.47 s
deepseek-v4-pro-0813	Takes text, returns text.	1M	384K	$0.6918$2.0754/M	$0.0231/M	—	35 tok/s	0.62 s
deepseek-v4-pro	Takes text, returns text.	1M	384K	$1.69$3.38/M	$0.1403/M	—	97 tok/s	1.18 s
deepseek-v3.2-think	Takes text, returns text.	164K	64K	$0.302$0.453/M	$0.0302/M	—	23 tok/s	1.96 s
deepseek-v3.2	Takes text, returns text.	164K	64K	$0.58$1.68/M	—	—	106 tok/s	2.55 s
cc-deepseek-v3	Takes , returns text.	164K	—	$0.3$0.3/M	—	—	58 tok/s	1.61 s
cc-deepseek-v3.1	Takes text, returns text.	160K	—	$0.56$1.68/M	—	—	59 tok/s	0.69 s
deepseek-r1-distill-llama-70b	Takes text, returns text.	131K	—	$0.8$1.6/M	—	—	101 tok/s	1.21 s
DeepSeek-V3.1-Think	Takes text, returns text.	128K	32K	$0.56$1.68/M	—	—	32 tok/s	1.29 s
DeepSeek-V3-Fast	Takes text, returns text.	32K	—	$0.56$2.24/M	—	—	150 tok/s	1.46 s
DeepSeek-OCR	Takes text, vision, returns text.	8K	—	$0.02$0.02/M	—	—	—	—
deepseek-ocr	Takes text, vision, returns text.	8K	—	$0.02$0.02/M	—	—	—	—
deepseek-ai/DeepSeek-R1-Distill-Llama-8B	Takes , returns text.	—	—	$0.01$0.01/M	—	—	—	—
deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B	Takes , returns text.	—	—	$0.01$0.01/M	—	—	—	—
deepseek-ai/DeepSeek-R1-Distill-Qwen-7B	Takes , returns text.	—	—	$0.01$0.01/M	—	—	—	—
tngtech/DeepSeek-R1T-Chimera	Takes , returns text.	—	—	$0.02$0.02/M	—	—	—	—
deepseek-ai/DeepSeek-Prover-V2-671B	Takes , returns text.	—	—	$0.1$0.1/M	—	—	—	—
deepseek-ai/DeepSeek-R1-Distill-Qwen-14B	Takes , returns text.	—	—	$0.1$0.1/M	—	—	—	—
deepseek-ai/DeepSeek-Coder-V2-Instruct	Takes , returns text.	—	—	$0.16$0.32/M	—	—	—	—
deepseek-ai/deepseek-llm-67b-chat	Takes , returns text.	—	—	$0.16$0.16/M	—	—	—	—
deepseek-ai/DeepSeek-V2-Chat	Takes , returns text.	—	—	$0.16$0.32/M	—	—	—	—
deepseek-ai/DeepSeek-V2.5	Takes , returns text.	—	—	$0.16$0.32/M	—	—	—	—
deepseek-ai/deepseek-vl2	Takes , returns text.	—	—	$0.16$0.16/M	—	—	—	—
deepseek-ai/DeepSeek-R1-Distill-Qwen-32B	Takes , returns text.	—	—	$0.2$0.2/M	—	—	—	—
DeepSeek-v3	Takes , returns text.	—	—	$0.272$1.088/M	—	—	67 tok/s	1.17 s
deepseek-v3	Takes , returns text.	—	—	$0.272$1.088/M	—	—	67 tok/s	1.17 s
alicloud-deepseek-v3.2	Takes , returns text.	—	—	$0.274$0.411/M	$0.0548/M	$0.3425/M	35 tok/s	1.01 s
azure-deepseek-v3.2	Takes , returns text.	—	—	$0.58$1.68/M	—	—	106 tok/s	2.55 s
azure-deepseek-v3.2-speciale	Takes , returns text.	—	—	$0.58$1.68/M	—	—	176 tok/s	1.60 s
deepseek-ai/DeepSeek-R1-Distill-Llama-70B	Takes , returns text.	—	—	$0.6$0.6/M	—	—	—	—
deepseek-ai/Janus-Pro-7B	Takes , returns text, vision.	—	—	$2$2/M	—	—	—	—
deepseek-ai/DeepSeek-R1-Zero	Takes , returns text.	—	—	$2.2$2.2/M	—	—	—	—

Prices are USD per million tokens; cache read and cache write are the rates for prompt-cache hits and for writing a prompt into the cache. Throughput and latency are measured on AIHubMix — the same figures the model detail page shows — not vendor claims. A dash means the catalog does not publish that field for that model, which is not the same as the model not supporting it.

Common Comparisons

DeepSeek-V3.1-Think vs alicloud-deepseek-v3.2DeepSeek-V3.1-Think vs azure-deepseek-v3.2DeepSeek-V3.1-Think vs azure-deepseek-v3.2-specialeDeepSeek-V3.1-Think vs qwen3.8-max-previewDeepSeek-V3.1-Think vs kimi-k3

DeepSeek on AIHubMix

Which DeepSeek model should I start with?

deepseek-v4.1-flash at $0.1408/M input — the cheapest entry here that declares tool calling, and it carries a 1M context. Move up to deepseek-ai/DeepSeek-R1-Zero when answer quality matters more than cost, or to deepseek-v4-flash for long-form reasoning.

Which of these models reason before answering?

11 of the 37 models here declare a reasoning phase — they work through the problem before producing an answer, which helps on multi-step problems at the cost of extra output tokens. Use the Reasoning filter above the table to see them. The catalog does not record anything further about how they differ, so this page does not sort them into families.

Why are there several entries for the same model?

Because each row is a route you can call, not a model release. Some IDs name an upstream (azure-, alicloud-, cc-), some are the open-weight repository form (deepseek-ai/…), and some differ only in capitalisation, kept so older integrations keep working.

The catalog does not carry a field saying which of those a given row is, so this page does not sort them into buckets it would have to invent. Every row shows that route’s own price, context and speed — compare those directly, and open a model to see the upstreams that serve it.

How is cached input billed?

The Cache read column is the rate for input tokens served from the prompt cache — for example deepseek-v4-flash-vision-exp bills cache hits at 2% of the input rate and deepseek-v4.1-flash bills cache hits at 2% of the input rate. Cache write is the surcharge for putting a prompt into the cache in the first place, and only a few upstreams bill it separately. A dash in either column means the catalog carries no cache rate for that model, so plan on paying the full input rate.

Do I need a separate DeepSeek account?

No. One AIHubMix key covers every model on this page, and switching between them is a change to the model string — billing, rate limits, and logs stay in one place.