DeepSeek V4.1 Flash vs V4 Pro: review and pricing

DeepSeek V4.1 Flash, often searched as DeepSeek 4.1 Flash, is the small model of DeepSeek's new generation: a 552B-parameter MoE with a 1M-token context that understands images. The current build is V4.1-Flash-1017 from 17 October 2026. Flash covers most agentic and text work. Move to V4 Pro when Flash fails your own tests and a mistake costs more than the gap in token prices.

What DeepSeek V4.1 Flash is, and what 1017 means

In its announcement of 10 September 2026 DeepSeek called V4.1-Flash the smallest model of the new line. It reads images natively, with no separate adapter. The model has 552 billion parameters in total: 8 billion are active while it reads the input and 16 billion while it generates the answer. The key-value cache takes four times less video memory and eight times less SSD space than in the previous generation. That matters for agents. The vendor says cache hits make up a large share of their costs, so the smaller the cache, the lower the bill.

The 1017 in the name is the update date, 17 October 2026. The changelog describes it as a "Pareto" improvement: benchmarks went up, tasks finish faster and answers got shorter. The way you call the model and its price stayed the same, and the API name is still deepseek-flash. DeepSeek published the weights on Hugging Face under the MIT license.

The timeline, according to DeepSeek:

The old names deepseek-v4-flash and deepseek-v4-flash-vision-exp still work: the vendor answers them with the new model and bills at the Flash rate. DeepSeek planned to retire V4 Pro on 14 September but changed its mind. According to the changelog, the V4 Pro API keeps running at its old pricing.

V4, V4 Pro and V4.1 Flash: how DeepSeek's models differ

The data comes from the Hugging Face model cards (V4-Pro, V4.1-Flash-1017), the announcement and the pricing table.

Thinking is switched on and off with a request parameter, and the depth is chosen from the levels low, high and max. DeepSeek's documentation has a ready example call with "thinking": {"type": "enabled"} and reasoning_effort.

DeepSeek R1 and deepseek-reasoner: what happened to them

R1 came out on 20 January 2025, and its update R1-0528 followed on 28 May 2025. Hugging Face lists R1-0528 at 685 billion parameters under the MIT license. In the API its place was taken by the name deepseek-reasoner, but with every generation that name was redirected to a newer model. On 24 April 2026 DeepSeek announced that it would shut down deepseek-chat and deepseek-reasoner on 24 July 2026. The pricing table of 29 October no longer lists them: only deepseek-flash and deepseek-v4-pro remain.

The standalone reasoning model became a mode inside V4 and V4.1. If you specifically want R1, run its weights yourself. The Ollama library has the deepseek-r1 family in sizes 1.5b, 8b, 32b, 70b, 671b and others, but that is the previous generation, not V4.

Prices: DeepSeek and codexkey

The vendor's official prices per 1M tokens as of 29 October 2026 (pricing page). Off-peak, everything costs half as much:

Peak hours run from 01:00 to 04:00 and from 06:00 to 10:00 UTC, Monday to Friday, except Chinese public holidays. In Moscow time that is 04:00-07:00 and 09:00-13:00. All other times, including weekends and Chinese holidays, are billed at the reduced rate.

The codexkey price list as of 29 October 2026, per 1M tokens. For current figures see the models page:

From the names, deepseek-v4.1-flash matches the September release and deepseek-v4.1-flash-1017 the build of 17 October. The price is the same, so pick -1017 for new projects. Check the model page to see exactly what answers under the ID deepseek-v4-flash on the gateway: at the vendor that name now serves V4.1-Flash. We cover the free deepseek-v4.1-flash-1017-free in DeepSeek for free: limits and status.

Let us run the numbers with explicit assumptions: a month uses 30M input tokens with no cache hits and 5M output tokens. On codexkey deepseek-v4.1-flash comes to 30 × $0.04 + 5 × $0.16 = $2.00, and deepseek-v4-pro to 30 × $0.12 + 5 × $0.34 = $5.30. The same volumes at DeepSeek cost $7.50 off-peak and $15.00 at peak on Flash; on Pro they come to $29.70 and $59.40. A real bill depends on the share of cache hits, which the vendor prices tens of times below misses.

When Flash is enough and when you need Pro

DeepSeek publishes its own measurements. They are the vendor's self-report with its own settings, and we found no independent measurements in official sources.

Between the September Flash and the 1017 build the vendor reports gains on Terminal-Bench 4.0, from 31.2 to 42.3 (measured with Claude Code), and on DeepSWE v1.1, from 74.2 to 77.0. In the 10 September announcement DeepSeek goes further: it says tests by several parties rank V4.1-Flash above V4-Pro on quality, price, speed and total completion time. That is the vendor's claim, and the table neither confirms nor refutes it. On HLE the 1017 build (42.6) nearly caught up with Pro (42.7), while on Terminal Bench 2.1 the September Flash (90.6) came out ahead of Pro (87.9).

How to choose in practice:

• Give bulk classification, data extraction and summarization to Flash: on codexkey its input is three times cheaper than Pro's.
• Screenshots, diagrams and documents with images also go to Flash, because Pro does not accept image input.
• With an agent in your editor, start on Flash and switch to Pro when tests stop passing.
• A costly mistake justifies a second pass: run Flash's result through Pro and compare the answers.
• Need high concurrency? Take Flash: the vendor allows it 2500 simultaneous requests against 500.

You can check these conclusions in one evening. Take ten real tasks from your repository that have tests, run both models and count how many tests each passes. Then divide the token spend by the number of tasks solved. That gives the cost per solved task, a fairer number than the price per million tokens. We rank models by scenario in Best AI for programming.

How to call DeepSeek through codexkey

codexkey gives you one sk_cdx_... key for every model in the catalog. You pay with a Russian bank card, SBP (the Russian Faster Payments System) or crypto, and you need neither a DeepSeek account nor a VPN. The address for OpenAI-compatible clients is https://api.codexkey.online/v1. Model names on the gateway differ from the vendor's (there is no deepseek-flash there), so take IDs from GET /v1/models.

from openai import OpenAI

client = OpenAI(base_url="https://api.codexkey.online/v1", api_key="sk_cdx_YOUR_KEY")

for model in ("deepseek-v4.1-flash-1017", "deepseek-v4-pro"):
    reply = client.chat.completions.create(
        model=model,
        messages=[{"role": "user", "content": "Find the bug: def avg(xs): return sum(xs) / len(xs) - 1"}],
    )
    print(model, "->", reply.choices[0].message.content)

An agent needs a config, not a script. A context-limit example that already lists deepseek-v4.1-flash is in the OpenCode guide, and Claude Code connects through a separate guide. For DeepSeek's own open agent, read what DeepSeek Harness is. If DeepSeek is not responding, the causes are covered in the piece on V4 outages, and access to the platform from Russia is covered in DeepSeek API from Russia.

FAQ

What is DeepSeek V4.1 Flash?

The small model of DeepSeek's new generation: a 552B-parameter MoE with a 1M-token context, answers up to 384K and image understanding. It came out on 10 September 2026, and the current build is V4.1-Flash-1017 from 17 October. In DeepSeek's API it is called deepseek-flash.

How does V4.1 Flash differ from V4 Pro?

Pro is larger, with 1.6T parameters and 49B active, but it does not accept image input and allows fewer parallel requests: 500 against 2500. Flash is cheaper and understands pictures. DeepSeek claims V4.1-Flash beats V4-Pro on its own tests, but it is better to decide by results on your own tasks.

How much does DeepSeek 4.1 Flash cost?

At DeepSeek, deepseek-flash costs $0.15 per 1M input tokens without a cache hit and $0.6 for output off-peak, twice that at peak. On codexkey, as of 29 October 2026, deepseek-v4.1-flash and deepseek-v4.1-flash-1017 cost $0.04 and $0.16.

What does 1017 in the DeepSeek name mean?

The update date, 17 October 2026. It is a small update to V4.1-Flash: DeepSeek reports higher benchmark scores, shorter answers and less time to finish tasks. The way you call it and the pricing stayed the same.

Is there DeepSeek R1 in 2026?

Not as a separate model in the API: the name deepseek-reasoner was redirected to newer generations and announced for shutdown on 24 July 2026. The weights of R1 and R1-0528 stay open under MIT, and reasoning is now a mode inside V4 and V4.1.

Does DeepSeek V4.1 Flash understand images?

Yes. In DeepSeek's model table, deepseek-flash has a mark for support in the Vision row, while deepseek-v4-pro says Not supported. If the task involves screenshots or diagrams, take Flash.

How do I use DeepSeek V4.1 Flash from Russia?

Through a gateway that takes payment in rubles: get a codexkey key and call https://api.codexkey.online/v1 with the ID deepseek-v4.1-flash-1017. Topping up the DeepSeek platform with a Russian card runs into the payment step, which is covered in DeepSeek API from Russia.