User-selectable compute tiers with different token pricing for DeepSeek V4.1 Flash

[XuZhangxing]

Hi DeepSeek team,

I would like to suggest an API-level feature for DeepSeek V4.1 Flash: user-selectable inference compute tiers, such as economy, standard, and full, with different token pricing for each tier.

This is related to #5254, which proposes auto-adjusting reasoning effort based on task difficulty. I think this proposal is slightly different: instead of only controlling reasoning depth or token usage, it is about exposing a clearer compute / latency / reliability / cost trade-off at the model API level.

My personal use case is ordinary business coding. Most of my daily tasks are not extremely difficult: editing application logic, writing CRUD code, fixing small bugs, reviewing common code paths, or asking an agent to make routine changes. I would like to use DeepSeek models more often, but as a cost-sensitive user, using the full capability and full pricing all the time can feel too expensive for these everyday tasks.

Today, reducing reasoning effort or limiting output tokens can help reduce cost, but from a user perspective this is not as intuitive as choosing a compute tier with a different token price. Ideally, each tier would have its own token pricing, reflecting different inference compute allocation, latency, and reliability trade-offs.

For example, V4.1 Flash could expose something like:

{
  "model": "deepseek-v4.1-flash",
  "compute_tier": "economy"
}

or:

{
  "model": "deepseek-v4.1-flash",
  "compute_tier": "full"
}

Possible tiers could be:

Economy  - lower inference compute, lower token price, lower latency, suitable for simple coding and routine agent tasks
Standard - balanced default for normal development work
Full     - maximum capability, normal/full token price, suitable for complex reasoning, large refactors, difficult debugging, or high-stakes tasks

Ideally, this could support both modes of control:

1. The user can explicitly set a tier when they know the task is simple or difficult.
2. The agent can automatically choose a tier based on task difficulty, similar in spirit to Auto-adjust reasoning effort based on task difficulty #5254, but at the compute / pricing tier level.

I want to emphasize that I am not asking for a vague "80% performance" mode. Model capability is not a linear slider. A better framing may be reduced inference compute allocation, or an economy compute tier, where the trade-off is documented as lower cost and latency in exchange for a lower ceiling on hard tasks.

This would also be different from only reducing reasoning tokens. Reasoning effort controls how much the model thinks, and output limits control how much it says, but a compute tier could represent a broader product and API contract: capability, latency, reliability, and token pricing. For many developers, especially those doing routine business coding, that may be easier to understand and more useful in daily usage.

I think this could make each model family more flexible without necessarily replacing separate model lines such as Pro and Flash. For example, Flash could still focus on speed and cost, while Pro could focus on stronger capability. Within each family, compute tiers would give users a more practical way to choose the right cost / latency / capability trade-off for each task.

Thanks for considering this.