DeepSeek V4.1 Flash: Pricing, Benchmarks and Pro Migration

DeepSeek V4.1 Flash is no longer a beta model. DeepSeek officially released it on September 10, 2026 and made it available through the API as deepseek-flash. This is more than a routine checkpoint: text, vision, thinking and agent capabilities now share one Flash route, while V4 Flash and Vision-Exp are retired and V4 Pro is scheduled to migrate.

This article first appeared on September 8, when only a temporary test route was public. We have now rechecked the DeepSeek launch announcement, changelog, official pricing and open model card. The beta verdict has been replaced with official-release facts. Mengbi did not run a paid API benchmark, so all performance numbers below remain attributed to DeepSeek.

Cover: original AI-generated Mengbi editorial artwork, not an official DeepSeek launch image.

DeepSeek V4.1 Flash at a glance

V4.1 Flash has moved from an interesting fast-model candidate to DeepSeek's default economical multimodal API. V4 Pro remains on the price page for the moment, but DeepSeek has published a firm date for rerouting that model name.

The legacy names deepseek-v4-flash and deepseek-v4-flash-vision-exp are temporarily accepted. Their original models have been retired, and requests now route to V4.1 Flash at Flash prices. Compatibility keeps older systems running, but it does not provide a durable version pin.

Why the new architecture changes input and output economics

DeepSeek describes V4.1 Flash as the smallest member of its new architecture family. The official model card and technical report specify 552B backbone parameters, while only about 8B parameters activate per token during prefill and 16B during decoding.

The model uses a Causal Encoder-Decoder design with a 20-layer causal encoder followed by a 20-layer decoder. The encoder processes the prompt first, then the decoder builds its global KV cache from the encoded state. Together with CSA2 sparse attention and FP4 KV caching, DeepSeek reports a global cache footprint of 890 bytes per token, roughly one quarter of V4 Flash. Persistent SSD storage for SWA KV state falls to about one eighth of the previous generation.

That split matters for agent workloads with long inputs and shorter outputs. Repositories, tool logs and conversation history repeatedly occupy the prompt. Lower prefill activation and cache costs can improve throughput without giving up the 1M-token context window. They do not guarantee that every request finishes four times faster. Cache hits, reasoning effort, output length, tool execution and service load still determine end-to-end time.

Vision is part of pre-training rather than a late adapter. The model card says DeepSeek-ViT image embeddings and text were trained together on a 45T-token multimodal corpus. The documented interface accepts text and images and produces text. There is no official basis for extending “native multimodal” to audio or video input.

Official benchmarks: stronger agents, but not a clean sweep over Pro

DeepSeek says V4.1 Flash has comprehensively surpassed V4 Pro in performance, cost, speed and total time. Its technical report supports a large agent and coding improvement, but it also records exceptions in knowledge-heavy, tool-free reasoning.

These are DeepSeek-run evaluations, not independent Mengbi results. The V4.1 Instruct scores use maximum reasoning effort, temperature=1.0 and top_p=0.95. Code-agent tests also use named DeepSeek Harness, mini-SWE or Claude Code scaffolds. Changing the agent, reasoning budget or stopping rule can change the outcome.

The practical verdict is narrower than the launch slogan. V4.1 Flash should be tested first for agent and coding work, but the published averages do not prove it wins every research, knowledge or long-horizon task. Model selection still depends on acceptance rate, total completion time and the cost of failures. Our AI coding agents comparison explains why a model score cannot substitute for testing the surrounding toolchain.

DeepSeek V4.1 Flash API pricing

The new prices took effect at 04:00 UTC on September 10, 2026. DeepSeek retains peak and off-peak billing: off-peak rates are half of peak rates.

Peak windows are 01:00–04:00 and 06:00–10:00 UTC, Monday through Friday. All other hours are off-peak. Prices can change, so production cost models should read the current official price page instead of treating this table as a permanent contract.

For an uncached request with 10,000 input tokens and 2,000 output tokens:

• Off-peak: 0.01 × $0.15 + 0.002 × $0.60 = $0.0027
• Peak: 0.01 × $0.30 + 0.002 × $1.20 = $0.0054

One thousand equivalent requests would cost about $2.70 or $5.40 before retries, tool-generated context or other services. Cost per accepted result remains the useful measure. A low token rate does not help if the model needs repeated corrections.

How to call the API and what happens to old model names

New integrations should use deepseek-flash; the base URL remains api.deepseek.com. This is a minimal OpenAI SDK example:

import osfrom openai import OpenAI
client = OpenAI(    api_key=os.environ["DEEPSEEK_API_KEY"],    base_url="https://api.deepseek.com",)
response = client.chat.completions.create(    model="deepseek-flash",    messages=[{"role": "user", "content": "List three migration risks."}],    reasoning_effort="high",)
print(response.choices[0].message.content)

The official thinking-mode guide says Chat Completions enables thinking by default at high. OpenAI-format requests can set reasoning_effort to low, high or max; disabling thinking uses thinking.type. Test streaming, reasoning_content, tool calls and usage parsing rather than stopping at an HTTP 200 response.

“Temporarily” matters. DeepSeek does not give an end date for the compatibility aliases in its current documentation. Production systems should migrate explicitly to deepseek-flash and monitor the served model, billing and response shape.

What V4 Pro users should do before September 14

Under the current official plan, every deepseek-v4-pro request will route to V4.1 Flash from September 14, 2026 at 12:00 Beijing time (04:00 UTC). Those calls will be billed at Flash prices until V4.1 Pro launches. DeepSeek has not announced a V4.1 Pro release date.

If production depends on Pro, do not wait for the silent route change to reveal behavioral differences. Complete four checks first:

1. Replay an accepted Pro evaluation set through deepseek-flash, including hard questions, long context, tool failures and structured output.
2. Record accuracy, correction time, time to first token, complete job time and total tokens instead of comparing price alone.
3. Validate image input, thinking effort, streaming parsing and tool-call fields so the adapter does not silently drop data.
4. Preserve request IDs, model names and scores before and after the change. An alias cannot prove which model served a request after routing changes.

The official account-level concurrency limit is 2,500 for V4.1 Flash and currently 500 for V4 Pro. Which limit a Pro alias inherits after September 14 should be verified against the live rate-limit documentation and API responses. A pricing change does not by itself define scheduling behavior.

Verdict: prioritize the release, but validate the migration

The official release fills the gaps that made the beta difficult to judge: a stable API name, peak and off-peak prices, a technical report, MIT-licensed weights and a dated legacy-routing plan. V4.1 Flash makes large gains over the old Flash on DeepSeek's agent and coding evaluations and beats V4 Pro on most of the reported tasks. Native vision and a smaller KV cache also fit long-context, tool-heavy workflows.

The launch announcement is still not your production acceptance test. GPQA and the text-only HLE subset show that Pro does not lose every dimension. The sensible move is to run deepseek-flash against your regression set now, preserve a Pro comparison before September 14 and place price, accuracy and end-to-end time in the same report.

For another low-cost, long-context model, add our GLM-5.3-Flash API analysis to the same evaluation. Compare completed, accepted work rather than a single launch score.

FAQ

Is DeepSeek V4.1 Flash officially released?

Yes. DeepSeek officially released V4.1 Flash on September 10, 2026 and exposed it through the API as deepseek-flash. The temporary beta route reported on September 8 is no longer the current integration path.

How much does the DeepSeek V4.1 Flash API cost?

Off-peak prices per million tokens are $0.003 for cached input, $0.15 for uncached input and $0.60 for output. Two weekday peak windows double those rates. Recheck the official pricing page before funding or deploying an account.

Does V4.1 Flash support images?

Yes. The official API table marks Vision as supported, and the model card says it natively processes text and images to produce text. The release does not document audio or video input.

Can I still use deepseek-v4-flash?

Temporarily, yes, but the original model is retired. The name now forwards to V4.1 Flash. New applications should use deepseek-flash; existing applications should migrate and verify response parsing and billing.

Does V4.1 Flash beat V4 Pro everywhere?

DeepSeek's overall claim says it does, and most official agent and coding results point that way. V4 Pro still scores higher on GPQA Diamond and the text-only HLE subset. These are also vendor-run evaluations, so your workload remains the deciding evidence.

Can I keep calling V4 Pro after September 14?

The deepseek-v4-pro name will still accept requests under the current plan, but the backend will route them to V4.1 Flash and charge Flash prices. This continues until V4.1 Pro launches; no date has been announced.

Last updated: September 11, 2026. Core facts were checked against DeepSeek's official launch announcement, changelog, pricing page, thinking-mode documentation and official Hugging Face model card. Mengbi did not run a paid API call or independent benchmark for this article.