DeepSeek Review: V4.1-Flash and V4-Pro, From the V4 Preview to Today

If you run LLM workloads on a budget, this DeepSeek review sets out what the published evidence says about the two models DeepSeek’s API serves on 22 October 2026: deepseek-flash (DeepSeek-V4.1-Flash-1017) and deepseek-v4-pro (DeepSeek-V4-Pro-0813). We first published it on 24 April 2026, the day DeepSeek shipped the V4 Preview, whose API names at launch were deepseek-v4-pro and deepseek-v4-flash; V4-Pro has since reached general availability (13 August 2026) and V4-Flash has been retired for V4.1-Flash (10 September 2026). This version draws on DeepSeek’s official documentation and reports, dated below; we do not run our own evaluations.

Our verdict at a glance

Our assessment, based on DeepSeek’s pricing, rate-limit and Vision pages (22 October 2026) and its Change Log: deepseek-flash is the sensible default for most API workloads. It is cheaper on every token bucket (deepseek-v4-pro costs 3.3× as much per output token, 4.4× per cache-miss and about 7.33× per cache-hit input token, at either period), it alone accepts images, and its concurrency limit is five times higher. DeepSeek also claimed on 10 September 2026 that V4.1-Flash benchmarks ahead of V4-Pro, a vendor claim. Keep deepseek-v4-pro mainly for workflows already validated on it. Neither hosted model suits workloads whose compliance rules forbid processing in the People’s Republic of China, where DeepSeek’s Privacy Policy says personal data is collected, processed and stored; self-hosting the MIT-licensed weights is the alternative.

Who should use DeepSeek today, and who shouldn’t

This DeepSeek review is aimed at engineers and technical buyers who have real workloads to cost out. Our assessment splits as follows (prices as of 22 October 2026; volumes are illustrative arithmetic; ratios rounded):

• Use deepseek-flash if you run high-volume chat, classification, extraction, RAG or image-understanding pipelines where cost per million tokens drives your P&L. 10,000,000 output tokens cost 10,000,000 ÷ 1,000,000 × $1.20 = $12.00 at peak and 10,000,000 ÷ 1,000,000 × $0.60 = $6.00 off-peak. On Claude Opus 4.6, a legacy model at $25 per 1M output tokens on Anthropic’s pricing page, they cost $250.00: 20.8× the peak and 41.7× the off-peak figure.
• Consider deepseek-v4-pro if agentic coding or long-horizon workflows are already validated on it. At $3.96 peak / $1.98 off-peak per 1M output tokens, the same volume costs $39.60 / $19.80, against $250.00 on Opus 4.6, which costs 6.3× / 12.6× as much.
• Don’t use DeepSeek’s hosted API if data-residency obligations rule out processing in the People’s Republic of China (see the privacy section); self-hosting the open weights is the alternative.

How we assessed it

This review was rewritten on 22 October 2026. We do not run our own evaluations: benchmark scores are vendor-reported (rival scores as DeepSeek reported them), and costs are arithmetic on published prices, split into cache-hit input, cache-miss input and output at peak and off-peak rates. Sources, all read on 22 October 2026:

• DeepSeek Models & Pricing page (last modified 19 October 2026).
• The DeepSeek API Change Log (entries of 24 April, 13 August, 10 September and 17 October 2026) and release notes of 24 April, 13 August, 10 September and 17 October 2026.
• DeepSeek’s Thinking Mode, Vision, JSON Output and Responses API guides and Rate Limit page.
• The DeepSeek-V4 technical report (Preview), arXiv 2606.19348 (v1, 26 April 2026), and the DeepSeek-V4-Pro (Preview) model card: the source of every April 2026 figure here.
• Hugging Face model cards for DeepSeek-V4-Pro-0813 and DeepSeek-V4.1-Flash-1017.
• Hugging Face model cards for DeepSeek-V3 and DeepSeek-Coder-V2-Instruct (licence sections).
• DeepSeek’s Privacy Policy (last updated 10 February 2026).
• Apple’s US App Store listing for DeepSeek – AI Assistant (listed as free).
• The Anthropic, OpenAI and Google Gemini API pricing pages.

Architecture: from the V4 Preview to V4.1-Flash

At the 24 April 2026 Preview, DeepSeek listed V4-Pro at 1.6T total / 49B active parameters and V4-Flash at 284B total / 13B active, both Mixture-of-Experts models with a 1M-token context, and its model card put the repository and weights under the MIT License. That is broader than the original DeepSeek-V3 (December 2024) and DeepSeek-Coder-V2, whose model cards (V3, Coder-V2) put the code under MIT but the weights under the DeepSeek Model License. DeepSeek’s technical report estimates that at 1M tokens V4-Pro (Preview) needs 27% of the single-token inference FLOPs and 10% of the KV cache of DeepSeek-V3.2, and V4-Flash (Preview) 10% of the FLOPs and 7% of the KV cache size.

Today’s builds are newer. The DeepSeek-V4-Pro-0813 card calls it the official V4-Pro release, superseding the Preview, with a DSpark speculative-decoding module; it gives no parameter counts. V4-Flash was retired on 10 September 2026 for V4.1-Flash, which DeepSeek’s note that day described as a 552B-parameter MoE with 8B active parameters for input and 16B for output, native vision, and a KV cache needing 1/4 of the HBM and 1/8 of the SSD storage of the previous generation. DeepSeek calls the current DeepSeek-V4.1-Flash-1017 a minor update and has not restated those figures for it.

Both API models support JSON output, tool calls, the Responses and Anthropic APIs and Chat Prefix Completion (beta); FIM Completion (beta) is non-thinking only. Only deepseek-flash accepts images (JPEG, PNG, GIF or WebP, via base64, a URL or the Files API). Thinking is the default, at effort high: send "thinking": {"type": "disabled"} (in the OpenAI Python SDK, via extra_body) to turn it off, and set reasoning_effort to low, high or max, the three-level scale DeepSeek introduced on 13 August 2026.

What DeepSeek’s published benchmarks show

All figures here are vendor-reported. April 2026 benchmark-table numbers are for the V4 Preview in maximum-effort “Max” mode; most newer numbers use other benchmarks or versions and are not like-for-like, except HLE: DeepSeek’s V4-Pro-0813 card lists the Preview’s 37.7 beside the 0813 build’s 42.7 (both without tools).

Coding and agents

In DeepSeek’s April 2026 Preview report, V4-Pro-Max scored 67.9 on Terminal-Bench 2.0 (Claude Opus 4.6 Max: 65.4; GPT-5.4 xHigh: 75.1), 93.5 on LiveCodeBench (Opus 4.6: 88.8), a Codeforces rating of 3206 (no Opus score reported) and 80.6% on SWE-bench Verified (Opus 4.6: 80.8%). For the current builds, the Change Log reports Terminal Bench 2.1 87.9, DeepSWE 62.7 and NL2Repo 61.5 for V4-Pro-0813 (13 August 2026), and DeepSWE v1.1 77.0, Terminal-Bench 4.0 42.3 and NL2Repo-Bench 68.2 for V4.1-Flash-1017 (17 October 2026). For the earlier DeepSeek Coder models, see our DeepSeek Coder review and the head-to-head on DeepSeek Coder vs Copilot.

Reasoning and math

In the same April table, V4-Pro-Max scored 37.7 on HLE (Humanity’s Last Exam), behind GPT-5.4 (39.8), Claude Opus 4.6 (40.0) and Gemini-3.1-Pro (44.4; Google lists it as Gemini 3.1 Pro Preview), and 95.2 on HMMT 2026 (Opus 4.6: 96.2). On SimpleQA-Verified it scored 57.9 to Gemini-3.1-Pro’s 75.6, and DeepSeek’s 24 April 2026 note said V4-Pro trailed only Gemini-3.1-Pro on world knowledge. DeepSeek’s newer HLE figures without tools are 42.7 for V4-Pro-0813 (60.0 with tools; 13 August 2026) and 42.6 for V4.1-Flash-1017 (17 October 2026); the latter’s model card lists 56.3 for Opus 5 and 44.5 for GPT-5.6 Sol on the same test.

Writing

DeepSeek’s published benchmark lists for V4-Pro-0813 and V4.1-Flash-1017 include no writing-quality test, and we run no evaluations of our own, so we make no claim either way about their prose against Claude’s. DeepSeek’s April 2026 Preview report did include internal Chinese-writing evaluations (effort level not stated): V4-Pro led Gemini-3.1-Pro on functional writing (62.7% vs 34.1% win rate) and Claude Opus 4.5 led V4-Pro on the hardest, complex-constraint or multi-turn prompts (52.0% vs 45.9%). If writing is your primary need, weigh the alternatives in the DeepSeek vs Claude comparison.

Long context

The current builds’ Change Log entries and model cards list no long-context benchmark; in the April 2026 Preview, V4-Pro-Max scored 83.5 on MRCR 1M and 62.0 on CorpusQA 1M (Claude Opus 4.6 Max: 92.9 and 71.7). Illustrative arithmetic (22 October 2026 prices): one uncached 480,000-token prompt costs 480,000 ÷ 1,000,000 × $0.30 = $0.144 on deepseek-flash at peak ($0.072 off-peak at $0.15), and 480,000 ÷ 1,000,000 × $1.32 = $0.6336 on deepseek-v4-pro ($0.3168 off-peak at $0.66). Billed as cache hits, the same prefix costs $0.00288 peak / $0.00144 off-peak on deepseek-flash.

How to actually call the API

Chat requests hit POST https://api.deepseek.com/chat/completions, the OpenAI-compatible endpoint; both models also accept the Responses API format and the Anthropic format (the latter at base URL https://api.deepseek.com/anthropic). The API is stateless: your client resends the full conversation history on every request.

Minimal Python (OpenAI SDK):

import os
from openai import OpenAI

client = OpenAI(
    base_url="https://api.deepseek.com",
    api_key=os.environ.get("DEEPSEEK_API_KEY"),
)

resp = client.chat.completions.create(
    model="deepseek-flash",  # or "deepseek-v4-pro"
    messages=[{"role": "user", "content": "Refactor this service for async I/O."}],
    reasoning_effort="high",
    extra_body={"thinking": {"type": "enabled"}},
    max_tokens=8000,
)

print(resp.choices[0].message.reasoning_content)
print(resp.choices[0].message.content)

With thinking enabled, the response returns reasoning_content alongside the final content. In thinking mode temperature, presence_penalty and frequency_penalty have no effect (and raise no error), so the sample omits them; for local deployment, the V4-Pro-0813 and V4.1-Flash-1017 model cards recommend temperature = 1.0. For JSON mode, set response_format={"type": "json_object"}, include the word “json” and a small example in your prompt, and set max_tokens high enough to avoid truncation; DeepSeek’s JSON Output guide warns that the API may occasionally return empty content. Full walkthroughs live in the DeepSeek API documentation and the DeepSeek API getting started guide.

Older integrations: deepseek-chat and deepseek-reasoner were retired after 15:59 UTC on 24 July 2026; from 24 April until then they routed to deepseek-v4-flash (non-thinking and thinking respectively). deepseek-v4-flash is still accepted but retired, and its requests are served by DeepSeek-V4.1-Flash-1017 at the Flash price. To migrate, set model to deepseek-flash (or deepseek-v4-pro); base_url stays https://api.deepseek.com. Thinking is on by default, so pass {"thinking": {"type": "disabled"}} wherever you relied on the old non-thinking deepseek-chat behaviour.

Value for money: peak and off-peak cost arithmetic

These are the rates on DeepSeek’s official pricing page as of 22 October 2026, in USD per 1M tokens, peak / off-peak:

Off-peak rates are half the peak rates. Peak hours are 01:00–04:00 and 06:00–10:00 UTC, Monday to Friday, excluding Chinese public holidays; all other hours, including weekends and those holidays, are off-peak, so 35 of the 168 hours in a normal week are peak. Peak/off-peak pricing took effect at 16:00 UTC on 16 August 2026, and DeepSeek reduced API prices with the V4.1-Flash release on 10 September 2026.

Worked example: 1M requests on deepseek-flash

Illustrative arithmetic, not a measured workload: 1,000,000 API calls, each with a 2,000-token cached system prompt, a 200-token uncached user message and a 300-token response.

• Cache-hit input: 2,000 × 1,000,000 = 2,000,000,000 tokens; ÷ 1,000,000 × $0.006 = $12.00 peak, × $0.003 = $6.00 off-peak
• Cache-miss input: 200 × 1,000,000 = 200,000,000 tokens; ÷ 1,000,000 × $0.30 = $60.00 peak, × $0.15 = $30.00 off-peak
• Output: 300 × 1,000,000 = 300,000,000 tokens; ÷ 1,000,000 × $1.20 = $360.00 peak, × $0.60 = $180.00 off-peak
• Total: $432.00 peak / $216.00 off-peak

Same workload on deepseek-v4-pro

• Cache-hit input: 2,000,000,000 ÷ 1,000,000 × $0.044 = $88.00 peak, × $0.022 = $44.00 off-peak
• Cache-miss input: 200,000,000 ÷ 1,000,000 × $1.32 = $264.00 peak, × $0.66 = $132.00 off-peak
• Output: 300,000,000 ÷ 1,000,000 × $3.96 = $1,188.00 peak, × $1.98 = $594.00 off-peak
• Total: $1,540.00 peak / $770.00 off-peak

Here deepseek-v4-pro costs $1,540.00 ÷ $432.00 = 3.56× as much as deepseek-flash, and the same off-peak ($770.00 ÷ $216.00), rounded to two decimals; all other figures are exact. Our assessment: start on deepseek-flash and move call sites to deepseek-v4-pro only where your own evaluation shows a gain worth the price. For a reusable calculator, use the DeepSeek pricing calculator; for a quick heuristic token estimate (the API’s usage block is authoritative), the DeepSeek token counter.

Privacy and regulatory considerations

DeepSeek’s Privacy Policy (last updated 10 February 2026, read 22 October 2026) states: “To provide you with our services, we directly collect, process and store your Personal Data in People’s Republic of China.” It also says it does not cover data that developers’ own apps collect from their end users; the developer, as controller, should disclose its own rules. For regulated workloads, processing in China is a real constraint, so check the rules in your jurisdiction before sending data to the hosted API. To keep data in-house, self-host the MIT-licensed weights covered in the FAQ below. See the DeepSeek privacy guide for the full picture.

Competitor context

For price context, as of 22 October 2026 the vendors’ pages list per 1M tokens: Claude Opus 4.6 at $5 input / $25 output (a legacy model at Anthropic); GPT-5.4 at $2.50 / $15 and GPT-5.5 at $5 / $30 (OpenAI’s short-context rates, up to 272K input tokens); and Gemini 3.1 Pro Preview at $2 / $12 (prompts up to 200K tokens). None is its vendor’s newest model. Price per token is not cost per task, because output lengths differ between models.

For deeper head-to-head comparisons, see DeepSeek vs ChatGPT and the full DeepSeek reviews hub. If you are evaluating alternatives on coding specifically, the DeepSeek alternatives for coding list is the right next read.

What we would like DeepSeek to change

Three limitations stand out in the published evidence. First, DeepSeek’s benchmarks are vendor-run and their versions shift between releases (Terminal-Bench 2.0 in April, 2.1 in August, 4.0 in October), which hampers like-for-like tracking. Second, deepseek-v4-pro has narrower limits (no image input; 500 concurrent requests per account against 2,500) and an uncertain future: DeepSeek’s 10 September 2026 note said its requests would be routed to V4.1-Flash, before the Change Log reversed that “in response to user demand”. Third, the V4-Pro (Preview), V4-Pro-0813 and V4.1-Flash-1017 model cards say there is no Jinja-format chat template; DeepSeek ships Python encoding scripts instead (the V4.1-Flash-1017 card also points to deepseek-recipe, Rust libraries with Python bindings), which suits Python stacks best.

DeepSeek AI Guide is an independent resource and is not affiliated with DeepSeek or its parent company. Model IDs, pricing and API behaviour change; check the official DeepSeek documentation and pricing page before committing to a production decision.