大模型网摘DeepSeek's Current Model Lineup: How deepseek-flash and deepseek-v4-pro Divide the Work
If you have not opened DeepSeek's API docs in a while, the model list may surprise you: it is down to two lines — deepseek-flash and deepseek-v4-pro. As of November 2026, that is the entire on-sale lineup: the former is served by DeepSeek-V4.1-Flash-1017, the latter by DeepSeek-V4-Pro-0813. The legacy names deepseek-v4-flash and deepseek-v4-flash-vision-exp are still accepted, but those models are retired; their requests are routed to V4.1-Flash-1017 and billed at flash prices. This article puts the two tiers side by side, per the official Models & Pricing page.
What they share: 1M context as standard
The spec sheet is nearly identical across both tiers: a 1M-token context window and up to 384K output tokens; JSON Output, Tool Calls, Responses API, Anthropic API and Chat Prefix Completion (Beta) are supported on both, while FIM completion (Beta) is non-thinking mode only. Both default to thinking mode and can be switched to non-thinking via a parameter — meaning "does it reason deeply?" is no longer a reason to pick one model over the other; it is a switch inside a single call.
Difference one: only flash sees images
The starkest line in the feature table is Vision: deepseek-flash accepts image input, deepseek-v4-pro explicitly does not. If your workload involves screenshot understanding, chart Q&A or document layout analysis, v4-pro is out — not a pricing trade-off, but a hard capability boundary. V4.1-Flash shipped with native vision back at its September 2026 launch (see our launch coverage), and the October 17 "1017" update iterated further from there.
Difference two: price and concurrency — pro costs 4x more
Per 1M tokens, off-peak: flash costs $0.15 input on a cache miss and $0.60 output, with cache-hit input as low as $0.003, at a concurrency limit of 2,500. v4-pro costs $0.66 in and $1.98 out, cache-hit $0.022, with a concurrency limit of 500. Roughly speaking, v4-pro's input runs about 4.4x flash's and its output about 3.3x, at one-fifth the concurrency allowance. Peak hours (01:00–04:00 and 06:00–10:00 UTC, Monday to Friday, excluding Chinese public holidays) double prices on both tiers.
How to choose: by workload, not by habit
Three rules of thumb. Vision, high concurrency, cost-sensitive — pick deepseek-flash; it is the default answer for most workloads. Text-only, stability-critical, budget-flexible — trial v4-pro against flash on your own data; the official docs publish no per-benchmark gap between the tiers, so a side-by-side on your workload beats any second-hand claim. Batch jobs that can run off-peak — both tiers support peak/off-peak pricing, and flash's $0.60 per 1M off-peak output is already low; stack it with context caching and the bill shrinks further (see our long-doc caching + off-peak tutorial). For the technical backstory of the flash tier, read our V4.1-Flash tech-report explainer.
Note that this is an unofficial news site; the parameters above are compiled from official public documentation, and real-world behavior should be validated against your own tests.