DeepSeek pricing explained: cheap per token, not always per task

DeepSeek's list prices are low, but a low price per token does not guarantee a low price per task. DeepSeek's models generate thousands of reasoning tokens per answer, and those are billed as output. On my measured chat answer task, DeepSeek V4.1 Flash cost 0.76¢ per task, more than GPT-6.1 Sol at 0.67¢, even though Sol's list price is much higher.

Per-token prices are only half the bill

As of 2026-10-11, DeepSeek V4.1 Flash lists at $0.30 per 1M input tokens, $1.20 per 1M output tokens and $0.006 per 1M cached input tokens. Its context window is 1,048,576 tokens and reasoning is on by default. DeepSeek V4 Pro 0813 lists at $0.66 input and $1.98 output.

Compare that with GPT-6 Luna at $0.10 input and $0.50 output, or GPT-6.1 Sol at $2.00 input and $10.00 output. On paper, Luna is cheapest of these and Sol the most expensive. Per task, the ranking changes, because reasoning tokens are billed as output tokens. A model that thinks for a long time costs more than its list price suggests.

What a chat answer costs

In my measured chat run, DeepSeek V4.1 Flash used 65 input tokens, 859 output tokens and 5,461 reasoning tokens. That came to 0.76¢ per task, or $7.60 per 1,000 tasks. The median across all 326 models is 0.24¢.

GPT-6.1 Sol answered for 0.67¢ ($6.72 per 1,000), using 45 input tokens, 560 output tokens and only 103 reasoning tokens. Despite a $2.00 input price, it came out slightly cheaper per task than DeepSeek's lower-priced model.

#	Model	Cost per task		Text score	Tokens in / out	Basis
1	GLM 5.3 FlashZ.ai	0.04¢$0.36 / 1,000		1474	48 / 1,782	measured best value
2	GPT-6 LunaOpenAI	0.06¢$0.59 / 1,000		1443	45 / 1,179	measured
3	Claude Haiku 5.5Anthropic	0.07¢$0.66 / 1,000		-	61 / 1,299	measured
4	Qwen3.8 FlashAlibaba Qwen	0.08¢$0.76 / 1,000		-	88 / 1,598	measured
5	Mistral Small 4Mistral	0.08¢$0.83 / 1,000		-	40 / 1,377	estimated
6	MiMo-V2.6-ProXiaomi	0.09¢$0.90 / 1,000		1479	206 / 933	estimated best value
7	DeepSeek V4.1 FlashDeepSeek	0.09¢$0.91 / 1,000		1475	65 / 6,320	measured
8	Llama 4 MaverickMeta	0.11¢$1.06 / 1,000		1327	87 / 1,592	estimated

The V4 Pro 0813 figure is an estimate, not a measurement. See the chatbot answer task page for the full list.

Code review and RAG

Code review shows the reasoning load most clearly. I gave each model CPython's textwrap.py (497 lines). DeepSeek V4.1 Flash hit my 16,000-token output cap: 4,801 input tokens, 0 visible output tokens and 16,000 reasoning tokens. It cost 2.1¢ ($20.64 per 1,000) and returned no visible review. The median for this task is 1.4¢.

DeepSeek V4 Pro 0813 also hit the cap, at an estimated 3.5¢. Several other models hit the cap too, including Qwen3.8 Flash and GLM 5.3 Flash, so this is not unique to DeepSeek.

The lesson: output caps are part of the bill, and a model can spend its whole budget on reasoning and return nothing usable. My runs used default settings, so a different configuration could change the result. See the code review task page.

For RAG over five retrieved documents, V4.1 Flash did better: 3,696 input, 648 output and 3,102 reasoning tokens, for 0.56¢ ($5.61 per 1,000). The median is 0.50¢. GPT-6 Luna's measured run cost 0.05¢. V4 Pro 0813 was estimated at 0.99¢. See the RAG task page.

List price changes

DeepSeek prices on OpenRouter have moved a lot. This table compares first-seen prices with the 2026-10-11 snapshot, from weekly Internet Archive data.

Output prices rose for most of these models. DeepSeek V4 Flash output went from $0.28 to $1.28 per 1M, while V4 Pro fell.

Week-to-week moves can be large. On 2026-10-05, the V4.1 Flash listing showed input at $0.003 and output at $2.40 per 1M. On 2026-10-11 it showed $0.30 and $1.20, a 50% output cut. Some of these values look like feed artifacts, so treat any single snapshot with caution and check the provider's pricing page before you budget.

Output price cuts also matter less than reasoning volume. Halving an output price does little if the model generates several times more reasoning tokens than a rival.

Off-peak pricing

DeepSeek's official pricing page lists an off-peak price at half the peak price. Peak hours are 01:00-04:00 and 06:00-10:00 UTC on weekdays. Batch work such as overnight code review or document extraction can be scheduled outside those windows. I have not measured per-task cost at off-peak times, but you can model it in the calculator.

What to do with this

• Track reasoning tokens, not just list prices. A per-token comparison misleads when one model thinks far longer.
• Set output caps deliberately. A 16,000-token cap can become a 16,000-token bill with no answer.
• Measure your own task. My numbers come from one run per task with default settings. See the methodology.
• Schedule flexible work off-peak to use the discount.

FAQ

Is DeepSeek V4.1 Flash cheaper than GPT-6 Luna?

Per token, no: $0.30 input and $1.20 output versus $0.10 and $0.50 per 1M. Per task, Luna was also far cheaper in my chat and RAG runs, partly because DeepSeek generated thousands of reasoning tokens.

Why did DeepSeek V4.1 Flash cost 2.1¢ on code review with no output?

It spent all 16,000 output tokens on reasoning. Reasoning is billed as output, so the run was charged like a full answer while returning nothing visible.

Does the off-peak discount change the per-task cost?

DeepSeek lists off-peak rates at half of peak, outside 01:00-04:00 and 06:00-10:00 UTC on weekdays. I have not measured off-peak per-task costs, so I cannot give a figure from my own data.