DeepSeek V4.1 Flash Arrives: API Pricing, Multimodal Features, and Migration Risks

*This research-based analysis covers DeepSeek's September 10 release and API documentation checked September 16, 2026. DiscoverAI has not independently benchmarked the model. Performance and efficiency results attributed to DeepSeek are provider-reported.*

DeepSeek V4.1 Flash is a 552-billion-parameter mixture-of-experts model with native image input, a one-million-token context window, and a new asymmetric Causal-Encoder-Decoder architecture. It is available through the deepseek-flash API name and as open weights.

Peak API rates are $0.006 per million cached input tokens, $0.30 per million uncached input tokens, and $1.20 per million output tokens. Off-peak rates are half those amounts. The low prices are striking, but the immediate operational story is migration: retired V4 Flash aliases now route to V4.1 Flash, and DeepSeek says V4 Pro requests are being redirected during the transition to V4.1 Pro.

DeepSeek describes an asymmetric model that activates eight billion parameters while processing input and 16 billion for output. The company says its redesigned cache requires one-quarter of the HBM and one-eighth of the SSD storage of the previous generation. Those claims target the expensive part of long-running agents: repeatedly retaining and reading large contexts.

The model accepts text and images and targets coding, research, tool use, and other agent workflows. DeepSeek reports gains across coding and agent benchmarks, including results above its earlier V4 Pro in its internal comparison. These tables are useful release evidence, not independent validation. Tool scaffolds, inference settings, retry budgets, graders, and contamination can materially change agent scores.

For deepseek-flash, the published peak price is $0.006 per million cache-hit input tokens, $0.30 per million cache-miss input tokens, and $1.20 per million output tokens. Off-peak prices are $0.003, $0.15, and $0.60 respectively.

Cache-hit pricing is exceptionally low, but it should not anchor a forecast by itself. Real cost depends on hit rate, output length, retries, tool calls, multimodal inputs, concurrency, and whether the result passes review. Benchmark cost per accepted task using production-shaped prompts and record both latency and reviewer time.

DeepSeek retired deepseek-v4-flash and deepseek-v4-flash-vision-exp; those names temporarily resolve to V4.1 Flash. The company also announced that deepseek-v4-pro traffic would route to V4.1 Flash after September 14 until V4.1 Pro arrives, although the current pricing page continues to list V4 Pro separately. That documentation tension is a reason to inspect live response metadata and billing rather than assume a model string identifies fixed behavior.

Pin the supported deepseek-flash name, capture model/version metadata where available, and rerun golden tests for JSON schemas, tool choice, refusal behavior, vision extraction, long context, latency, and token use. Consequential automations need staged rollout and a rollback path. Compatibility routing prevents an HTTP error; it does not guarantee semantic compatibility.

The weights provide more inspection and deployment control than an API-only release, but 552B total parameters remain an infrastructure-scale system. DeepSeek explicitly asks organizations considering large deployments to contact it if they have resources on the order of 2,000 GPUs and storage clusters.

Self-hosting also transfers patching, inference optimization, access control, abuse prevention, observability, and incident response to the operator. Teams without that capacity should treat the weights as auditable research material and compare managed inference providers rather than equate “open” with inexpensive operations.

DeepSeek's public privacy terms and hosting jurisdiction require review before sending personal, regulated, or proprietary data. API affordability does not answer retention, training use, subprocessors, deletion, incident notification, or cross-border transfer requirements.

Keep secrets out of prompts, minimize and redact inputs, use scoped tool credentials, isolate generated code, and request contractual answers for sensitive deployments. For image input, remember that screenshots can expose names, messages, account identifiers, and confidential interface data even when the text prompt looks harmless.

DeepSeek V4.1 Flash is noteworthy for combining native vision, a very long context, unusually low public API prices, and architecture work aimed at lowering cache costs. It deserves evaluation for high-volume coding, document, and agent workloads where price per accepted task matters.

The release also demonstrates why model aliases are a governance issue. Existing applications may receive a different model without changing code. Test the current endpoint, verify invoices and response metadata, and preserve a rollback path before trusting provider benchmarks or compatibility routing in production.