DeepSeek Performance Review: V4.1-Flash and V4-Pro on the Published Evidence

Is DeepSeek’s current API lineup capable enough, and cheap enough, to build on? This DeepSeek performance review is the editors’ assessment of DeepSeek’s official documentation, its prices as of 22 October 2026 and the benchmark figures DeepSeek itself reports; we do not run our own evaluations. The current models are deepseek-flash (DeepSeek-V4.1-Flash-1017) and deepseek-v4-pro (DeepSeek-V4-Pro-0813). This page first appeared on 24 April 2026, the day DeepSeek released the V4 Preview as deepseek-v4-pro and deepseek-v4-flash. The V4-Flash line (re-post-trained as V4-Flash-0731 on 31 July 2026) was retired on 10 September 2026; the deepseek-v4-flash name is still accepted but served by DeepSeek-V4.1-Flash-1017 at the Flash price.

Our assessment at a glance (published evidence, 22 October 2026)

Facts only, no scores. Prices are USD per 1M tokens, peak / off-peak, as of 22 October 2026.

Who should use it, who shouldn’t

Use DeepSeek if you run high-volume chat, document processing, agentic coding or RAG where repeated (cache-hit) input dominates the bill, and your data rules allow a provider whose Privacy Policy says it stores personal data in the People’s Republic of China (or you can self-host open weights). Cached input is the standout: $0.006 per 1M tokens for deepseek-flash at peak, $0.003 off-peak. deepseek-v4-pro undercuts every model in our price table below charging $2.00 or more per 1M input tokens, while deepseek-flash costs more than GPT-4o mini on uncached input and output at peak.

Don’t use DeepSeek if your compliance rules exclude that data location or frontier-level factual recall is your main job (see the knowledge gap below). For latency-sensitive chat, remember that thinking is on by default; disable it with "thinking": {"type": "disabled"}.

If the privacy trade-off is a blocker, browse the open-source AI like DeepSeek list for models you can run on your own infrastructure.

How we assessed it

The site does not run its own evaluations, so no score, timing or success rate here is a measurement of ours. Benchmarks are DeepSeek’s own claims, and its benchmark versions and harnesses change between releases, so most cross-release figures are not like-for-like.

Sources and dates

• Models & Pricing (fetched 22 October 2026, last modified 19 October 2026) and the API docs home page (checked 22 October 2026).
• API Change Log entries of 1 December 2025 and 24 April, 31 July, 13 August, 10 September and 17 October 2026, the V4.1-Flash release note (10 September 2026) and the DeepSeek-V3.1 release note (21 August 2025).
• Thinking Mode, JSON Output, FIM Completion and Responses API guides and the Rate Limit & Isolation page: checked 22 October 2026.
• DeepSeek-V4 technical report, arXiv 2606.19348 (26 April 2026), and the V4-Pro model card: April 2026 Preview figures.
• Hugging Face model cards and file listings for DeepSeek-V4-Pro-0813 (13 August 2026), DeepSeek-V4.1-Flash-1017 (17 October 2026), the V4 Preview repositories and DeepSeek-V3.2: checked 22 October 2026.
• Privacy Policy (last updated 10 February 2026) and Open Platform Terms of Service (effective 29 April 2026).
• Anthropic, OpenAI and Google Gemini API pricing pages, OpenAI’s deprecations page and Anthropic’s models overview: as of 22 October 2026.

Minimal example

A minimal thinking-mode request with the OpenAI Python SDK, per DeepSeek’s Thinking Mode guide, shown for reference only (we did not run it). Both models share the endpoint (POST https://api.deepseek.com/chat/completions) and the thinking and effort fields, but image input works on deepseek-flash only, and top_p differs: Flash applies it only in thinking mode, Pro only in non-thinking mode.

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": "system", "content": "You are a senior staff engineer."},
        {"role": "user", "content": "Review this diff and flag regressions."},
    ],
    reasoning_effort="high",  # "low" | "high" | "max"
    extra_body={"thinking": {"type": "enabled"}},
    max_tokens=8000,
)
print(resp.choices[0].message.reasoning_content)
print(resp.choices[0].message.content)

Thinking mode returns reasoning_content alongside the final content; temperature and the penalty parameters have no effect in thinking mode, so the sample omits them. The API is stateless: the client resends the full messages array each turn. Coming from DeepSeek V3.2? Its API names, deepseek-chat and deepseek-reasoner, were retired in July 2026 (see the FAQ).

What DeepSeek’s published figures show

Everything here is vendor-reported and dated; April 2026 Preview figures describe models the API no longer serves.

Coding and agentic work

DeepSeek’s 10 September 2026 note says V4.1-Flash benchmarks ahead of flagship models including V4-Pro, a vendor claim. Its Change Log figures: 77.0 on DeepSWE v1.1, 68.2 on NL2Repo-Bench and 42.3 on Terminal-Bench 4.0 for V4.1-Flash-1017 (17 October 2026), against NL2Repo 61.5 and DeepSWE 62.7 for V4-Pro-0813 (13 August 2026). The two DeepSWE scores used different harnesses (Mini-SWE-V2 for V4.1-Flash-1017, DeepSeek Harness minimal mode for V4-Pro-0813, per DeepSeek’s model cards); the table below shows where versions differ.

Launch-day figures belong to the April 2026 Preview: V4-Pro-Max scored 80.6 on SWE-bench Verified (80.8 for Claude Opus 4.6 Max in DeepSeek’s table) and V4-Flash-Max 79.0 (DeepSeek-reported, Max effort).

Our assessment, based on these figures and the price list: for day-to-day review and refactoring, deepseek-flash is the sensible default. For IDE-style completion, FIM is a beta feature that works in non-thinking mode only. See the DeepSeek for coding walkthrough.

Long-context retrieval

Both models take a 1M-token context, but DeepSeek’s Change Log gives no long-context benchmark for them; in the April 2026 Preview report, V4-Pro-Max scored 83.5 (MMR) on MRCR 1M, against 76.3 for Gemini 3.1 Pro High and 92.9 for Claude Opus 4.6 Max. The report credits a hybrid of Compressed Sparse Attention and Heavily Compressed Attention for its efficiency figures; see the FAQ.

JSON extraction and structured output

DeepSeek’s JSON Output guide says the feature is meant to “ensure the model outputs valid JSON strings”, but warns that the API “may occasionally return empty content”. Its advice: set response_format={"type": "json_object"}, include the word “json” and an example in the prompt, and set max_tokens high enough to avoid truncation. See DeepSeek API JSON mode and DeepSeek API best practices for a retry-and-validate pattern.

Reasoning and math

DeepSeek reported GPQA Diamond 90.9 for the September V4.1-Flash build and HLE without tools 42.6 for V4.1-Flash-1017, against 42.7 for V4-Pro-0813. On HMMT 2026 Feb, the April 2026 Preview table put V4-Pro-Max at 95.2, against 96.2 for Claude Opus 4.6 Max and 97.7 for GPT-5.4 xHigh: gaps of 1.0 and 2.5 points.

Speed and latency: what is documented

DeepSeek’s documentation gives no latency or throughput figures for the current models, so neither do we. Documented behaviour: the chain of thought (reasoning_content) comes before the answer, so for latency-sensitive chat switch thinking off or use reasoning_effort="low". DeepSeek’s 17 October 2026 note claims V4.1-Flash-1017 cut task runtime and output length versus the previous V4.1 build.

Value for money: an illustrative cost calculation

An illustrative, hypothetical workload: 1,000,000 calls, each with a 2,000-token system prompt (assumed to hit the cache every time; a lower hit rate raises the totals), a 200-token user message (cache miss) and a 300-token response (thinking disabled; the default thinking mode adds chain-of-thought tokens to output). Each line is tokens ÷ 1,000,000 × rate, at prices as of 22 October 2026; every line comes out to whole cents, so nothing is rounded.

deepseek-flash at official rates (peak and off-peak)

• Cached 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
• Uncached 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

deepseek-v4-pro at official rates (peak and off-peak)

• Cached input: 2,000,000,000 tokens ÷ 1,000,000 × $0.044 = $88.00 peak; × $0.022 = $44.00 off-peak
• Uncached input: 200,000,000 tokens ÷ 1,000,000 × $1.32 = $264.00 peak; × $0.66 = $132.00 off-peak
• Output: 300,000,000 tokens ÷ 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 about 3.6× deepseek-flash at either period ($1,540.00 ÷ $432.00 = $770.00 ÷ $216.00 ≈ 3.56, rounded at the end), and running entirely off-peak halves either bill.

Prices checked against DeepSeek’s official Models & Pricing page on 22 October 2026 (page last modified 19 October 2026); DeepSeek may adjust them. The DeepSeek pricing calculator shows peak and off-peak costs side by side for your own token volumes and cache-hit ratio.

How DeepSeek compares on price

Standard (non-batch) list prices, USD per 1M tokens, at each vendor’s default (shortest) context tier, from the official pricing pages as of 22 October 2026. DeepSeek input is the cache-miss rate.

On output, GPT-5.5 costs about 7.6× deepseek-v4-pro’s peak rate and 15.2× its off-peak rate; Claude Opus 4.6 about 6.3× and 12.6×. Their $5.00 input rate is about 3.8× Pro’s peak and 7.6× its off-peak rate. Among the low-cost rows, deepseek-flash at peak charges more for cache-miss input than GPT-4o mini, GPT-5.4 nano and Gemini 3.1 Flash-Lite; off-peak it undercuts the latter two and matches GPT-4o mini. Its clear lead is cached input ($0.006 peak, $0.003 off-peak, against $0.02–$0.075), and DeepSeek’s 10 September 2026 note says “Cache-hit charges often account for a large share of agent costs.” Each vendor also lists newer models; OpenAI has announced shutdowns for GPT-4o mini (14 January 2027) and GPT-5.4 nano (1 April 2027), and Anthropic lists Claude Opus 4.6 and Sonnet 4.6 as legacy models, still available (as of 22 October 2026). Treat “cheaper than” as a pricing comparison, not a quality verdict.

Where DeepSeek falls short

• Expert-level knowledge. DeepSeek’s own model cards still show a gap on HLE without tools: 42.6 for V4.1-Flash-1017 against 56.3 for Claude Opus 5 and 44.5 for GPT-5.6 Sol (17 October 2026 card), and 42.7 for V4-Pro-0813 against 49.8 for Claude Opus 4.8 (13 August 2026 card). In the April 2026 Preview table, V4-Pro-Max also trailed Gemini 3.1 Pro High on SimpleQA-Verified (57.9 vs 75.6). For factual Q&A at the edge of human knowledge, compare candidates on your own questions.
• Privacy. DeepSeek’s Privacy Policy (last updated 10 February 2026) says “we directly collect, process and store your Personal Data in People’s Republic of China”. It covers DeepSeek’s apps, websites and related services; for end users of apps built on the API, the developer is the controller. The Open Platform Terms of Service are governed by the laws of the People’s Republic of China (mainland). See DeepSeek privacy for the full trade-off.
• JSON mode is not a guarantee. It may occasionally return empty content. Build in a retry and a validator.
• Peak hours cost twice as much. Since 16 August 2026, off-peak rates have been half the peak rates. Peak is 01:00–04:00 and 06:00–10:00 UTC, Monday to Friday, excluding Chinese public holidays: 35 of the 168 hours in a normal week, about 20.8%. All other hours, including weekends and Chinese public holidays, are off-peak. (DeepSeek’s earlier off-peak discount stopped on 5 September 2025, as its V3.1 release note announced.)
• Model churn. V4-Flash-0731 arrived on 31 July 2026, V4-Pro went GA as V4-Pro-0813 on 13 August, V4-Flash was retired for V4.1-Flash on 10 September and V4.1-Flash-1017 followed on 17 October. deepseek-v4-flash and deepseek-v4-flash-vision-exp are still accepted but retired and served by DeepSeek-V4.1-Flash-1017 at the Flash price, and DeepSeek announced, then reversed, a plan to route deepseek-v4-pro to V4.1-Flash from 14 September 2026. Re-check the Change Log before relying on any model.

Competitor context

If you are cross-shopping, the two comparisons most relevant to this review are DeepSeek vs Claude and DeepSeek vs ChatGPT. For a broader sweep, start at the DeepSeek alternatives hub; the DeepSeek reviews hub collects every DeepSeek review.

Verdict

Our assessment, based on DeepSeek’s published figures and official prices as of 22 October 2026: deepseek-flash (DeepSeek-V4.1-Flash-1017) is the DeepSeek model to start with for cost-sensitive chat, classification, extraction and standard coding. It costs $1.20 per 1M output tokens at peak and $0.60 off-peak, it is the only one of the two that accepts images, and DeepSeek’s 10 September 2026 note claims V4.1-Flash benchmarks ahead of V4-Pro (a vendor claim; harnesses differ between releases). deepseek-v4-pro costs 3.3× as much on output and 4.4× on cache-miss input, though its rates sit well below the table’s competitor rows at $2.00 or more per 1M input; keep it for comparison or for workflows already validated on it. At the April 2026 launch, DeepSeek’s own table showed V4-Pro-Max behind Claude Opus 4.6 Max, GPT-5.4 xHigh and Gemini 3.1 Pro High on HLE, and behind the first two on HMMT 2026 Feb; its newer model cards still put today’s models behind Claude Opus 4.8 and Opus 5 on HLE without tools.

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.