Which Dify plan should you buy? Count collaborators first, then calculate model costs

When a team first buys Dify, it is easy to start by comparing the number of models. What often determines the right plan, however, is who needs workspace access to maintain applications, and how many model calls a single business request involves. If one person is simply testing an internal question-answering workflow, start with Sandbox. If four colleagues all need to help maintain it, the three-member Professional plan should not be treated as a long-term solution.

The assessments below are based on the official pricing page and documentation consulted on November 3, 2026. They are not load-test results and do not guarantee that any particular plan can support your actual concurrent workload.

The official pricing page lists workspace member limits of one for Sandbox, three for Professional and fifty for Team. These should not be interpreted directly as the number of visitors an application can serve per day. When purchasing, first list the people who need to log in to the workspace, modify workflows or maintain knowledge materials. Then separately verify the access requirements of your application’s delivery method.

For example, if the business owner, workflow maintainer and knowledge-base maintainer are exactly three people, Professional can be a candidate. If a fourth person also needs an ongoing role in workspace maintenance, assess Team rather than having several people share one administrator account. If a manager only occasionally reads the output, first check how it will be delivered. Do not upgrade solely because there is one more reader.

Current monthly list prices are USD 59 per workspace for Professional and USD 159 per workspace for Team, both excluding applicable taxes. The monthly difference is USD 100. Whether that difference is worthwhile depends first on whether the additional collaborators really need workspace seats, rather than treating member counts as a performance metric.

Dify deducts credits per model response. One response means one model call, not one complete request submitted by a business user. A more complex workflow might classify first, retrieve information and generate an answer next, and check formatting last. If each of these three steps calls a model once, one request contains three calls. This is an example for cost calculation, not a claim that every Dify workflow always makes three calls.

Charges for your own model API key are billed to the corresponding provider account. Your budget therefore needs at least two separate lines: the Dify workspace subscription and model expenses incurred through your own key. Check the provider’s rate limits separately as well. Upgrading a Dify plan does not establish that all model calls are unlimited.

Choose a low-risk workflow, such as drafting customer-service responses from product documentation you are already authorized to use. Prepare twenty questions with reference answers, including questions involving missing information, ambiguous wording and matters outside the scope. This number is a pilot recommendation, not an official testing standard.

Run the same workflow on each question, recording whether the answer can be used, how many model calls each question triggers, whether human revision is needed and which step fails. The goal of the pilot is not to produce an attractive average, but to distinguish two types of problems. If the workflow frequently cites the wrong materials, fix the knowledge base and retrieval logic first. If the answers are already usable but collaboration or resource limits get in the way, then assess a paid upgrade. Upgrading the plan cannot replace improvements to content quality.

The official Docker Compose documentation specifies minimum hardware of a two-core CPU and 4 GiB of memory. These are only minimum deployment requirements, not a promise of production capacity. The documentation also requires checking new or changed environment variables during upgrades, showing that self-hosting does not mean maintenance ends once installation is complete.

If a team cannot identify who will handle upgrades, backup verification and incident resolution, starting with a cloud service usually makes responsibilities and required investment easier to estimate. If you have operational capabilities, assess self-hosting against actual data flows and control requirements. You must also check where data goes when using the selected models and external services. Hosting the application on your own server is not, by itself, grounds to conclude that all data stays in-house.

For teams just starting out, the safest purchasing sequence is to demonstrate that the workflow is useful with a small sample, confirm the actual number of collaborators, and then budget separately for the platform subscription and model bills. An upgrade buys an identifiable benefit only when a specific limit is blocking an already usable workflow.