Select a current DeepSeek identifier from the official catalog and verify the features your task requires. Compare pricing and context only after the request interface and acceptance criteria are clear.
Model families at DeepSeek
The current hosted catalog uses deepseek-flash and deepseek-v4-pro. The provider’s model reference distinguishes their backing versions and supported capabilities. Use the live table for numeric capacity and price details rather than transferring figures from an older family name. Official DeepSeek documentation.
Start selection from the input and artifact you need. For a code repair, preserve the relevant source and the failing check. For a document answer, include the material that supports the response and a case where the answer is absent. For a conversational feature, include follow-up references that depend on prior messages. These examples test whether a candidate fits the application rather than whether it can produce a fluent demonstration.
Treat mode as part of the selected configuration. DeepSeek’s thinking guide describes controls and conversation-state requirements that matter independently of the identifier. Record the mode and required optional fields in your evaluation. If changing a mode changes the output contract or the way tools are handled, compare those operational differences along with task quality. Official DeepSeek documentation.
The catalog’s feature matrix should guide targeted tests. When your application needs image input, structured output or tool interaction, read the exact feature reference for the selected model. A family-level description is not enough to justify sending unsupported input or assuming all related identifiers behave identically. Keep the feature source next to the fixture that verifies it.
All models
| Model | Official identifier | Modalities | Context tokens | Input USD / 1M | Output USD / 1M | Status |
|---|---|---|---|---|---|---|
| deepseek-flash | deepseek-flash | text, image | 1,000,000 | $0.3 | $1.2 | Active |
| deepseek-v4-pro | deepseek-v4-pro | text | 1,000,000 | $1.32 | $3.96 | Active |
Last verified · Source ↗
The table joins available official model metadata with verified pricing. Missing information is left unresolved rather than filled from a marketing label. Open the source link when a capability is essential to your application, and keep the selected official identifier in configuration rather than spreading it across request-building code.
For an existing integration, compare the current live record with the configuration already deployed. A request that still succeeds may be served through an alias whose underlying model has changed. Preserve a small regression set before migration so you can evaluate the new behavior against the same expected artifacts and edge cases.
Cheapest / strongest / longest-context
Only active candidates with documented values for this ranking are included. Prices retain the tier and deployment condition shown below.
| Model | Provider | Input USD / 1M | Output USD / 1M | Context tokens | Price condition | Cost |
|---|---|---|---|---|---|---|
| deepseek-flash | DeepSeek | $0.3 | $1.2 | 1,000,000 | peak | Estimate cost |
| deepseek-v4-pro | DeepSeek | $1.32 | $3.96 | 1,000,000 | peak | Estimate cost |
Last verified · Source ↗
A price ranking compares documented rates under stated conditions. It does not measure code quality, factual reliability or interactive speed. Use the AI API cost calculator to evaluate the input-output mix of your workload, then combine that estimate with acceptance results. A lower unit price is useful only when the produced artifact meets the task’s requirements.
For a longest-context comparison, examine how the application uses the supplied material. Test retrieval of a relevant fact, resolution of conflicting passages and preservation of the required answer structure. More capacity does not remove the need to select relevant input. Avoid treating a model’s advertised window as an instruction to send every available file for every question.
For a strongest candidate, define the criterion before running the comparison. A coding test can require passing the project’s checks; an extraction task can compare fields with a known answer set; a support answer can require support in supplied documentation. Preserve failures and incomplete responses in the result set. Removing them from the comparison produces an optimistic picture of actual application behavior.
If you also evaluate another provider, retain the same task while adapting each provider’s documented request contract. Keep source-provider prices and third-party hosting prices separate. A model with a similar name through a router can have different availability, limits and billing. Read DeepSeek direct pricing and direct account capacity for the endpoint used here.
Keep the evaluation dataset independent of the model’s generated explanation of its own performance. For a coding task, inspect the changed files and run the relevant check. For an extraction, compare the required fields with the source record. For an explanation, ask a reviewer to identify unsupported statements and missing qualifications. Record these outcomes consistently across candidates. If you change the prompt after seeing a failure, retain the earlier result and mark the new run as a different configuration rather than overwriting the evidence that motivated the change. This creates a selection decision that another developer can reproduce and challenge.
Deprecations and replacements
DeepSeek’s current quickstart explicitly distinguishes accepted legacy Flash names from their retired backing models. Requests through those aliases are served by the current Flash service according to the notice. The same reference also gives the current continuation notice for V4 Pro; follow later official updates instead of inferring an end date from an older announcement. Official DeepSeek documentation.
For migration, make the new identifier an explicit configuration change and run the saved regression set. Inspect message formatting, output validation, tool handling, latency and actual usage. If the application stores model names with results, retain the original name on historical records. Rewriting old records to the new alias would erase useful evidence about what the application requested at the time.
A replacement should have a rollback plan that reflects actual availability. If the old backing model is retired, reverting to an old alias may not restore the earlier behavior. Keep the previous passing fixture and a documented alternative, but do not promise a rollback route that the provider no longer serves. The DeepSeek error guide can help separate availability problems from malformed requests.
Naming and versions
Use official identifiers exactly, including punctuation. Keep an alias, a backing version and a provider marketing name as different concepts in your own records. The API request uses an accepted identifier; documentation may separately describe what is currently behind it. That separation is especially useful when a library example outlives the model generation it originally demonstrated.
When building a selector, show the human-readable name while storing the exact request identifier. Avoid deriving identifiers by lowercasing arbitrary page titles. Validate the selected value against your allowed configuration before dispatch and display a clear unavailable-model state if a previously saved choice can no longer be used.
After selection, follow DeepSeek key setup and the Python tutorial to establish a reproducible baseline. Return to the DeepSeek API hub for service status and recent changes. Keep the selection decision connected to the actual workload so a later model change can be assessed with the same criteria.
Frequently asked questions
Which identifier should new Flash integrations use?
Does an accepted old alias mean the old model is still served?
Is the longest-context model automatically best for documents?
How is strongest determined here?
Should I save model names with completed jobs?
Can I compare a routed model with direct DeepSeek pricing?
Sources
Last verified · Source ↗