Short answer: “DeepSeek 4” refers to a family of models and releases, not one API model name. DeepSeek’s current API catalog lists deepseek-flash, served by DeepSeek-V4.1-Flash-1017, and deepseek-v4-pro, served by DeepSeek-V4-Pro-0813. Flash is the option to evaluate if you need image input or lower listed token prices. Consider Pro for text-only work if your own evaluation justifies its higher price. The catalog lists a 1M-token context length and a maximum output of 384K for both; it does not explain how those limits interact for a particular request. These details do not establish which model will perform best on your workload. (Models & Pricing)
API catalog and pricing details reflect DeepSeek documentation reviewed November 12, 2026. Model availability and prices can change.
DeepSeek V4 model and version map
| Name | What it refers to | What matters for selection |
|---|---|---|
| V4-Pro and V4-Flash | The V4 research preview described in the April 2026 paper. It lists Pro at 1.6T total parameters (49B activated) and Flash at 284B (13B activated), with a 1M-token context for both. | Research-era specifications; do not assume the preview-era Flash figures describe the later V4.1-Flash service. (arXiv paper) |
| V4.1-Flash-1017 | The Flash version identified in DeepSeek’s October 17, 2026 API changelog. | The API catalog maps deepseek-flash to this version. (API changelog; Models & Pricing) |
| V4-Pro-0813 | The Pro version identified in the current API pricing table. | Its API identifier is deepseek-v4-pro. The catalog marks Pro as not supporting vision. (Models & Pricing) |
| V4-Pro-Max | A maximum-reasoning-effort mode described in the V4 paper, not a separate API identifier listed in the current pricing table. | Do not use “Pro-Max” as an API model name unless current documentation specifies it. (arXiv paper) |
The V4 paper reports that, at a one-million-token context, V4-Pro used 27% of V3.2’s single-token inference FLOPs and 10% of its KV cache. That is a paper comparison against V3.2—not an API price, latency measurement, or Pro-versus-Flash result. (arXiv paper)
Which DeepSeek V4 model should you evaluate?
- For image-based tasks, start by evaluating Flash. DeepSeek’s vision guide says
deepseek-flashaccepts images alongside text and lists JPEG, PNG, GIF, and WebP. The current catalog says Pro does not support vision. Check the guide for image-size and request requirements before building around image input. (Vision guide; Models & Pricing) - If token price is a priority, compare Flash’s and Pro’s listed rates against your expected token usage. Flash’s listed rates are lower, making it a cost-first candidate—not a guarantee that it will need fewer retries or less human review. (Models & Pricing)
- For text-only work, evaluate Pro if its results justify the higher cost. The available paper and vendor benchmark posts do not establish that the current Pro version will outperform current Flash on your prompts, tools, and success criteria. Compare both on representative tasks before committing. (Models & Pricing; API changelog)
- For long-context tasks, check both documented limits. The current catalog lists a 1M-token context length and a maximum output of 384K for both models. It does not clarify how input and output budgets interact for a particular request, so confirm applicable request limits in current API documentation rather than assuming the full context is available for input. (Models & Pricing)
DeepSeek’s October changelog reports benchmarks for V4.1-Flash-1017, including Terminal-Bench 4.0 and DeepSWE v1.1, and names different test frameworks for those evaluations. Treat these as vendor-reported, release-specific results—not a standardized ranking against Pro or a prediction of your production results. (API changelog)
Current API identifiers and listed prices
DeepSeek lists prices in USD per 1 million tokens, separating cached input, uncached input, and output. Its current catalog lists these model identifiers and served versions: (Models & Pricing)
| API model name | Served model version | Input, cache hit | Input, cache miss | Output |
|---|---|---|---|---|
deepseek-flash |
DeepSeek-V4.1-Flash-1017 | $0.006 peak / $0.003 off-peak | $0.30 peak / $0.15 off-peak | $1.20 peak / $0.60 off-peak |
deepseek-v4-pro |
DeepSeek-V4-Pro-0813 | $0.044 peak / $0.022 off-peak | $1.32 peak / $0.66 off-peak | $3.96 peak / $1.98 off-peak |
Peak hours: 01:00–04:00 and 06:00–10:00 UTC, Monday through Friday, excluding Chinese public holidays. All other hours—including weekends and Chinese public holidays—are off-peak. DeepSeek lists off-peak rates at half of peak rates. Recheck the official table before budgeting or deployment. (Models & Pricing)
Worked estimate: Suppose one request uses 800,000 uncached input tokens and generates 100,000 output tokens—900,000 tokens combined. At peak rates, the listed model-token charge is about $0.36 on Flash ($0.24 input + $0.12 output) or $1.452 on Pro ($1.056 input + $0.396 output). At off-peak rates, those estimates are $0.18 and $0.726. This assumes no cache-hit input and is an estimate based on listed rates; actual token counts depend on the request and response. (Models & Pricing)
Production API: names and basic access details
DeepSeek’s quick-start documentation lists https://api.deepseek.com as the OpenAI-format base URL, requires an API key, and shows deepseek-flash in its example. It also lists https://api.deepseek.com/anthropic for the Anthropic-compatible format. (Your First API Call)
For example, with DEEPSEEK_API_KEY set in your environment, a minimal text request is:
curl https://api.deepseek.com/chat/completions \
-H "Content-Type: application/json" \
-H "Authorization: Bearer ${DEEPSEEK_API_KEY}" \
-d '{
"model": "deepseek-flash",
"messages": [{"role": "user", "content": "Summarize this paragraph in one sentence."}]
}'
To try Pro, change the model value to deepseek-v4-pro. This is a request-format example, not a test of either model’s performance. (Your First API Call)
Older Flash names and Pro routing
DeepSeek’s current pricing catalog says the older identifiers deepseek-v4-flash and deepseek-v4-flash-vision-exp are still accepted, although the corresponding models are retired. The catalog says requests using those identifiers are served by V4.1-Flash-1017 and billed at the Flash price. For new configuration, use deepseek-flash, the catalog-listed identifier. (Models & Pricing)
A dated Pro notice is less straightforward: DeepSeek’s September 2026 announcement said deepseek-v4-pro requests would route to Flash until V4.1-Pro launched, while its September API changelog said Pro API service would continue after September 14. The current API catalog lists deepseek-v4-pro as V4-Pro-0813 with its own prices. That catalog entry does not independently establish live request routing; confirm routing and billing in your account before relying on Pro-specific behavior. (September announcement; API changelog; Models & Pricing)
Bottom line
Start with deepseek-flash if you need image input or want the lower listed token price. Test deepseek-v4-pro if your text-only workload benefits enough to justify its higher rates. For long-context work, both are listed with a 1M-token context and a maximum output of 384K; verify how the applicable limits work for your request. Before launch, run a task-specific evaluation and check the official model and pricing pages—especially if your application depends on a particular version, alias, routing behavior, or billing window.
Sources
Models & Pricing | DeepSeek API Docs
Your First API Call | DeepSeek API Docs
Change Log | DeepSeek API Docs
DeepSeek-V4: Towards Highly Efficient Million-Token Context Intelligence
Introducing DeepSeek-V4.1-Flash: smarter, faster, more efficient