01
Overview
DeepSeek V4.1 Flash is available through DeepSeek's production API under the model name deepseek-flash. The September 2026 release is also a migration event: older V4 Flash aliases now resolve to V4.1 Flash, and DeepSeek says V4 Pro requests will begin routing to V4.1 Flash on September 14 while the company works toward a future V4.1 Pro.
This guide focuses on the hosted API rather than local weights.
02
The model ID to use: deepseek-flash
DeepSeek's official Models & Pricing page lists deepseek-flash as the API model whose current version is DeepSeek-V4.1-Flash.
The OpenAI-format base URL remains https://api.deepseek.com. DeepSeek also documents an Anthropic-format base URL at https://api.deepseek.com/anthropic.
The model supports both thinking and non-thinking modes, a one-million-token context window, and a maximum output size currently listed as 384K tokens.
DeepSeek's table also lists JSON output, tool calls, the Responses API, Anthropic API compatibility, chat-prefix completion, FIM completion in non-thinking mode, and Vision support for deepseek-flash.
03
What happens to the old V4 Flash model names?
DeepSeek says deepseek-v4-flash and deepseek-v4-flash-vision-exp remain accepted as legacy names, but the models behind those identifiers have been retired.
Requests sent to those names are now served by DeepSeek V4.1 Flash and billed at the current Flash price. That means a request can keep succeeding even though the underlying model version has changed.
For production systems, use the current deepseek-flash identifier when practical and regression-test behavior if you previously depended on one of the retired aliases.
04
DeepSeek V4.1 Flash API pricing
DeepSeek's official pricing page currently uses peak and off-peak rates. Prices are per one million tokens.
Off-peak
- Cache-hit input: $0.003 per 1M tokens.
- Cache-miss input: $0.15 per 1M tokens.
- Output: $0.60 per 1M tokens.
Peak
- Cache-hit input: $0.006 per 1M tokens.
- Cache-miss input: $0.30 per 1M tokens.
- Output: $1.20 per 1M tokens.
DeepSeek defines peak periods as 01:00–04:00 UTC and 06:00–10:00 UTC, Monday through Friday. It states that all other hours use off-peak pricing.
These are launch-period prices retrieved from DeepSeek's official documentation on September 10, 2026. DeepSeek explicitly says prices may change, so applications and articles should link to the live pricing page rather than treating these values as permanent.
05
Cache-hit vs cache-miss input
DeepSeek prices cached input far below uncached input. The exact savings a workload achieves therefore depend not only on total prompt tokens but also on how effectively repeated context can use DeepSeek's context-caching system.
For agentic applications with large, repeated system instructions or stable document context, cache behavior can materially affect cost. Do not estimate a production bill using only the cache-hit rate unless your own logs show that the relevant tokens are actually hitting cache.
06
Concurrency limits
DeepSeek's current pricing/model table lists a concurrency limit of 2,500 for deepseek-flash, compared with 500 for deepseek-v4-pro. Those are current service limits, not promises of achieved requests-per-second or latency under every workload.
Throughput depends on request size, context length, output length and service conditions. Treat the listed concurrency as an account/service constraint, not a benchmark.
07
V4 Pro migration on September 14
DeepSeek says that from 12:00 Beijing Time on September 14, 2026, requests to deepseek-v4-pro will be routed to V4.1 Flash and billed at V4.1 Flash prices until V4.1 Pro is released in the future.
DeepSeek's explanation is that V4.1 Flash has surpassed V4 Pro in the company's testing on performance, cost, speed and total time. That comparative claim is vendor-reported; developers should still test their own tasks before assuming behavioral equivalence.
The migration has an operational consequence: an integration that keeps requesting deepseek-v4-pro can receive a different underlying model after the cutover without changing its model string.
08
How to prepare for the V4 Pro cutover
First, inventory every application, agent, workflow and environment variable that still requests deepseek-v4-pro. Include staging systems and scheduled jobs, not only the main production app.
Second, run a representative regression set against deepseek-flash now. Compare structured-output validity, tool-call behavior, reasoning mode, long-context performance, latency and any prompts that were tuned specifically for V4 Pro.
Third, check whether your product assumes text-only behavior. V4.1 Flash supports Vision, but gaining an extra capability does not guarantee identical text behavior to the older model.
Fourth, update cost models. The Flash prices are substantially lower than the currently listed V4 Pro prices, but actual spend still depends on cache hit rate and output length.
Finally, monitor DeepSeek's changelog and pricing documentation around the September 14 cutover. A migration date announced at launch is time-sensitive operational information.
09
V4.1 Flash vs V4 Pro pricing
DeepSeek's current table lists V4 Pro at higher rates. For comparison, the V4 Pro peak rates are $0.044 per 1M cache-hit input tokens, $1.32 per 1M cache-miss input tokens and $3.96 per 1M output tokens. Off-peak V4 Pro rates are half those amounts.
This does not by itself prove V4.1 Flash is cheaper for every completed business task. A cheaper token price can be offset by different output lengths, retries or task-success rates. Cost comparisons should be made on successful end-to-end workloads.
10
OpenAI and Anthropic API compatibility
DeepSeek's documentation lists both OpenAI-format and Anthropic-format access. Compatibility makes migration easier for software built around those request patterns, but 'compatible' should not be interpreted as identical to every vendor-specific feature.
Test system-message behavior, tools, reasoning controls, vision payloads and streaming in the exact SDK or gateway used by your application.
11
Vision through the API
V4.1 Flash supports DeepSeek's Vision API feature. This is a meaningful difference from the current V4 Pro row, which lists Vision as unsupported.
Possible workloads include screenshot analysis, chart/document understanding and image-assisted agent tasks. For applications moving from V4 Pro, image support is an added option rather than a requirement to change existing text-only requests.
12
When should you use the hosted API instead of self-hosting?
The hosted API is the simpler path when you want immediate access, elastic serving and predictable token billing without operating the 552B-backbone model yourself.
The open weights are attractive when infrastructure control, data boundaries or custom deployment are more important. The correct comparison should include accelerator cost, utilization, engineering time and operations — not only token price versus electricity.
13
FAQ
What is the API model name for DeepSeek V4.1 Flash?
deepseek-flash.
Do old V4 Flash model IDs still work?
DeepSeek says the legacy deepseek-v4-flash and deepseek-v4-flash-vision-exp names are still accepted, but requests now run on V4.1 Flash.
When does V4 Pro start routing to V4.1 Flash?
DeepSeek says 12:00 Beijing Time on September 14, 2026.
Is V4.1 Flash cheaper than V4 Pro per token?
Under the pricing table retrieved on September 10, V4.1 Flash has lower listed input and output token rates. Real task cost still depends on usage and success rate.
Does the API support a 1M context window?
Yes. DeepSeek currently lists a one-million-token context length for deepseek-flash.
Does deepseek-flash support images?
Yes. DeepSeek's current model table lists Vision support.
Are the prices permanent?
No. DeepSeek explicitly states that product prices can change. Recheck the official pricing page before making a budget or publishing fixed figures.
Sources
Primary and supporting sources
Facts were rechecked against the linked sources immediately before publication. Pricing, product availability and rollout status can change.