- DeepSeek-V4.1-Flash is available under the deepseek-flash API name.
- Current documentation retains V4 Pro as a separate model, superseding the originally announced migration plan.
- Flash pricing varies by cache use and weekday UTC peak windows; Pro remains a separate service.
The release remains, the migration plan changed
DeepSeek launched V4.1 Flash on September 10, with an initial plan to redirect V4 Pro API requests to the new model. That migration should no longer be presented as the current policy. DeepSeek's documentation now lists V4 Pro separately alongside Flash, with its own model version, pricing and concurrency limit.
The supported Flash identifier is deepseek-flash. The older V4 Flash and experimental Flash vision identifiers are still accepted, but their traffic is served by V4.1 Flash. These legacy Flash aliases are a different case from V4 Pro, which remains listed as DeepSeek-V4-Pro-0813.
Why the architecture emphasizes input cost
The September 17 technical report frames long-running agents as increasingly input-heavy workloads. An agent repeatedly consumes conversation history, files and tool results before producing its next action. Reading and retaining that growing history can become a substantial part of serving cost, even when an individual answer is short.
DeepSeek describes a multimodal mixture-of-experts backbone with 552 billion parameters and a one-million-token context. Its causal encoder-decoder design activates 8 billion parameters per token during input processing and 16 billion during output generation. The paper reports training on a 45-trillion-token multimodal corpus.
A smaller cache in two kinds of memory
The paper combines compressed sparse attention, reuse of cached state across layers and FP4 cache storage. It reports a global cache footprint of 890 bytes per token in high-bandwidth memory, about one quarter of V4 Flash's corresponding footprint. A separate replay optimization reduces persistent cache storage to approximately one eighth of the earlier model's level.
Those ratios refer to particular cache components. They are not a statement that the entire model fits into one quarter of the hardware, or that every request costs one eighth as much. Model weights, temporary computation, batching and output length remain separate contributors. The result is specifically about reducing the state carried through long contexts.
What an API migration changes in practice
DeepSeek lists both OpenAI-format and Anthropic-format base URLs, along with Responses API support. Both current models allow up to 384,000 output tokens within the documented one-million-token context. The developer has to select the supported model identifier as well as the compatible request format; an older alias is not a way to pin the retired Flash checkpoint.
Peak billing applies on weekdays from 01:00 to 04:00 and 06:00 to 10:00 UTC, excluding Chinese public holidays. Other periods, including whole weekends, use the lower rates. For a workload containing one million uncached input tokens and 100,000 output tokens, the listed Flash rates imply $0.21 off-peak or $0.42 at peak, before any cache savings. This is an arithmetic example of the price schedule, not a measurement of how many tokens a particular task needs.
Current API terms
As checked on September 24, Flash's off-peak prices are $0.15 per million uncached input tokens, $0.003 for cached input and $0.60 for output. Peak rates are twice those amounts. The documentation defines weekday peak windows in UTC and excludes Chinese public holidays.
Both listed models support tools, structured JSON and thinking modes. Vision support is listed for Flash and not for V4 Pro. This makes the retained Pro service more than a naming footnote: callers choosing between the models face different capabilities as well as different prices. The original launch and its subsequent API policy therefore need to be read together.
Sources & context
Go to the original material. Company claims remain attributed to their sources.
01Updates & corrections
— Updated the original migration announcement: current DeepSeek documentation retains V4 Pro as a separate API model. Added the subsequently published V4.1 Flash technical report.
