Current status: October 1, 2026
Which DeepSeek API models are available?
Use deepseek-flash for V4.1 Flash and deepseek-v4-pro for V4-Pro-0813. V4.1 Flash was released September 10, 2026; DeepSeek continues Pro service. The original V4 Flash identifiers temporarily route to V4.1 Flash.
What are the current DeepSeek API limits?
Both current API models list 1M context and a maximum output of 384K tokens. They support thinking and non-thinking modes, tool calls, and Chat Completions, Responses and Anthropic API formats. Flash accepts images; Pro is text-only.
Current DeepSeek rate card
| API model | Cached input | Uncached input | Output |
|---|---|---|---|
| deepseek-flash | $0.006 / $0.003 | $0.30 / $0.15 | $1.20 / $0.60 |
| deepseek-v4-pro | $0.044 / $0.022 | $1.32 / $0.66 | $3.96 / $1.98 |
Source: https://api-docs.deepseek.com/quick_start/pricing/
USD per million tokens, shown as peak / off-peak. Peak hours are Monday–Friday, 01:00–04:00 and 06:00–10:00 UTC, excluding Chinese public holidays. Other hours use the off-peak rate. These are the checked October 1 rates; recheck the provider page before budgeting.
Can I self-host DeepSeek V4 weights?
Yes. DeepSeek publishes V4 Pro and V4.1 Flash weights under MIT. Use the deployment instructions for the exact checkpoint; a public API version and a downloaded checkpoint need separate verification.
Check the DeepSeek changelog, feature and pricing table, Flash model card and Pro license. The guide below covers release history and migration steps.
Release history and API migration details, reviewed October 1, 2026
Set new Flash calls to deepseek-flash. Test the current API’s tool calls and output handling before moving traffic. The original V4 Preview specifications below describe that release, and the current rate card above covers today’s hosted models.
- V4 Preview was released April 24, 2026 with Pro and Flash variants.
- Pro’s GA API update was released August 13, 2026.
- V4.1 Flash was released September 10, replacing the original Flash API model.
- Pro API service continues after the previously planned September shutdown.
- Downloaded checkpoints have their own deployment instructions.
Choose an API Model
Try Flash for an initial cost baseline. Keep Pro as a comparison candidate on tasks where your own tests show an improvement. Evaluate output accuracy, tool execution and total token spend separately; the model name alone does not establish which will finish your task better.
DeepSeek deployment choices
| Need | Route | Check |
|---|---|---|
| Lower listed API rates | deepseek-flash | Completed-task cost including retries |
| A Pro comparison | deepseek-v4-pro | Task quality and higher token cost |
| Image input | deepseek-flash | Image handling in your client |
| Your own inference stack | Downloaded checkpoint | Model version, memory and prompt format |
Version check
Pin or record the weight revision for local evaluations. API aliases can change their served version without changing the identifier in your application.
Official Release History
DeepSeek lifecycle
| Date | Published change |
|---|---|
| April 24, 2026 | V4 Preview API and open weights announced |
| August 13, 2026 | V4 Pro GA API update |
| September 10, 2026 | V4.1 Flash release; original Flash API retired; Pro service continues |
Source: https://api-docs.deepseek.com/updates/
The V4 Preview release links the weights and technical report. The changelog documents later API changes. Use both when comparing an original checkpoint with today’s hosted service.
Original V4 Pro and Flash Specifications
April 24 V4 Preview architecture sizes
| Original model | Total parameters | Active parameters |
|---|---|---|
| V4 Pro | 1.6T | 49B |
| V4 Flash | 284B | 13B |
Source: https://api-docs.deepseek.com/news/news260424/
These are the original V4 sizes. V4.1 Flash has a different architecture; the old Flash row does not describe it. Read the current model card for its architecture and inference implementation.
API Access and Migration
Current endpoints and features
| Item | Published value |
|---|---|
| OpenAI-format base URL | https://api.deepseek.com |
| Anthropic-format base URL | https://api.deepseek.com/anthropic |
| Model names | deepseek-flash; deepseek-v4-pro |
| JSON and tool calls | Supported by both current API models |
| Responses API | Supported |
| FIM completion | Beta; non-thinking mode only |
Source: https://api-docs.deepseek.com/quick_start/pricing/
Compatibility names
The old deepseek-v4-flash and deepseek-v4-flash-vision-exp identifiers currently route to V4.1 Flash. Set deepseek-flash explicitly in new integrations. DeepSeek’s April announcement gave a July retirement deadline for deepseek-chat and deepseek-reasoner; use the current documented identifiers.
Estimate API Spend
Use the current rate card above. Separate cached from uncached input, and account for whether a request falls in peak hours. Reasoning and failed attempts can increase the tokens needed for a successful result.
Inputs to a cost estimate
| Billing item | What to record |
|---|---|
| Uncached input | New prompt and conversation tokens |
| Cached input | Provider-reported cache hits |
| Output | Generated tokens, including reasoning where billed |
| Schedule | The peak or off-peak rate that applies |
Check before purchase
The original promotional Pro prices have been replaced by the current rate card. Use live provider rates for a purchasing decision.
Provider-Published Coding Results
DeepSeek reports 90.6 on Terminal-Bench 2.1 for V4.1 Flash in the September 10 changelog. The model card identifies the DeepSeek Harness Minimal setup. This is a provider evaluation; it does not establish an independent result for your agent or codebase.
How to use a benchmark result
| Check | Reason |
|---|---|
| Benchmark version | Different versions contain different tasks |
| Agent setup | Tool environment and execution logic affect completion |
| Own repository tasks | Measure accepted patches and total cost |
Architecture and Weight Deployment
Use the technical report and inference instructions for the exact checkpoint. The V4.1 Flash model card publishes a prompt encoder and minimal inference instructions. An API-compatible client still needs the correct prompt format when serving weights itself.
Weight deployment checks
| Item | Source to inspect |
|---|---|
| Architecture | Checkpoint model card and technical report |
| Prompt encoding | Provider’s encoding implementation |
| Inference | Provider’s deployment instructions |
| Permissions | Checkpoint license file |
Migration Checklist
Before moving API traffic
- 1Replace old Flash compatibility names with deepseek-flash.
- 2Record which served model your application is evaluating.
- 3Set the required thinking mode and validate client support.
- 4Test JSON output, tool execution and FIM if you depend on them.
- 5Check output handling against the documented limits.
- 6Measure tokens and retries using the current peak and cache rates.
Deployment Checks
Run a representative workload through your existing client and verify response parsing before rollout. Recheck the current pricing page when estimating monthly spend.
For a cross-provider choice, see the DeepSeek and Qwen comparison. For broader task selection, use the canonical model guide.
Founder of Spectrum AI Labs — testing AI tools and models, and writing up what actually ships.
More about Paras →Need help choosing an AI model?
Use our free AI Model Picker to compare models by use case, budget, and workflow.
Try the AI Model Picker
