DeepSeek V4 in 2026: Benchmarks, Pricing & API Guide

DeepSeek skipped the rumored R2 entirely and shipped what actually matters: DeepSeek V4, a two-tier open-weight family with a 1M-token context window. I tested the docs, the pricing page and the API surface so you don’t have to guess.

In this guide you’ll get the spec sheet, what V4 improves over V3.2, DeepSeek-reported benchmarks with honest caveats, official pricing, and a verified quickstart you can run in five minutes.

1. Spec Sheet: V4 Pro vs V4 Flash

DeepSeek V4 Preview launched on April 24, 2026 with two Mixture-of-Experts models. Both are open-weight under the MIT license, both support a 1M-token context with up to 384K output tokens, and both expose Thinking and Non-Thinking modes through the same API.

🚀

V4 Pro — the flagship

1.6T total · 49B active · GA build 0813

MoE flagship for hard reasoning and agentic coding. Served as DeepSeek-V4-Pro-0813 (GA since August 13, 2026). Weights on HuggingFace under MIT.

⚡

V4 Flash — the efficient one

284B total · 13B active · build 0731

Fast, economical MoE for high-volume work. Served as DeepSeek-V4-Flash-0731 (public beta since July 31, 2026). Same 1M context and API surface as Pro.

👁️

Flash Vision Exp — experimental

Image input · experimental

Experimental multimodal variant (deepseek-v4-flash-vision-exp, August 21, 2026) that additionally accepts image input. Expect rough edges.

📜

License & weights

MIT · HuggingFace

Both tiers ship open weights under the MIT license (collection deepseek-ai/deepseek-v4). You can self-host, fine-tune and distill without per-token fees.

Official API pricing (per 1M tokens)

Peak cache-miss input costs $0.44 (Flash) and $1.32 (Pro); output costs $1.32 (Flash) and $3.96 (Pro). Off-peak rates are exactly half. Cache-hit input drops to $0.014 (Flash) and $0.044 (Pro) at peak. Peak hours are 01:00–04:00 and 06:00–10:00 UTC, Monday to Friday. Always re-check the official pricing page before budgeting — earlier promo-era figures ($0.14/$0.28 Flash, $0.435/$0.87 Pro) no longer reflect the current table.

Context & API surface

Base URL https://api.deepseek.com (OpenAI format) plus an Anthropic-compatible endpoint. Both models support JSON output, tool calls, the Responses API, chat prefix completion and FIM (non-thinking mode only). Legacy names deepseek-chat and deepseek-reasoner were retired on July 24, 2026.

2. What V4 Improves Over V3.2

V4 is an efficiency release as much as a capability release. The headline architectural change is a hybrid attention mechanism combining Compressed Sparse Attention (CSA) and Heavily Compressed Attention (HCA), plus Manifold-Constrained Hyper-Connections and the Muon optimizer from the April 26 technical report.

🧠 Million-token context that’s actually affordable

At a 1M-token context, V4 Pro needs only about 27% of the single-token FLOPs and 10% of the KV cache of V3.2. Flash pushes further: roughly 10% of the FLOPs and 7% of the KV cache. Long-horizon agents and whole-repo analysis become economically feasible.

💻 Frontier-adjacent coding, open-weight

DeepSeek-reported V4-Pro-Max hits 80.6% on SWE-bench Verified, 93.5 on LiveCodeBench, 87.5 on MMLU-Pro and 90.1 on GPQA Diamond. Flash-Max trails closely at 79.0% on SWE-bench Verified — a 1.6-point gap that rarely shows up on routine tasks.

🔌 Modern API: Responses + Anthropic compatible

V4 natively supports the Responses API and an Anthropic-compatible endpoint, with dual Thinking/Non-Thinking modes on one model ID. No more juggling separate chat vs reasoner endpoints.

A note on R2

You may have seen “DeepSeek R2” rumors. As verified in August 2026, there is no official R2 release, model card, API ID, pricing entry or HuggingFace page — the official line is V4 Pro and V4 Flash. I cover V4 here because it’s what you can actually call today.

3. Quickstart: Call V4 in Five Minutes

The API is OpenAI-compatible, so you keep your tooling and only change the base URL and model name. Grab a key from the DeepSeek platform, then run one of these verified snippets.

cURL — first chat completion

curl https://api.deepseek.com/chat/completions \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer $DEEPSEEK_API_KEY" \
  -d '{
    "model": "deepseek-v4-flash",
    "messages": [
      {"role": "system", "content": "You are a helpful assistant."},
      {"role": "user", "content": "Explain MoE routing in two sentences."}
    ]
  }'

Python — OpenAI SDK with thinking mode

from openai import OpenAI

client = OpenAI(
    api_key="your-key-here",
    base_url="https://api.deepseek.com",
)

response = client.chat.completions.create(
    model="deepseek-v4-pro",
    messages=[
        {"role": "system", "content": "You are a senior coding assistant."},
        {"role": "user", "content": "Write a retry decorator with exponential backoff."},
    ],
    extra_body={"reasoning_effort": "high"},
)

print(response.choices[0].message.content)

My routing rule

Default to Flash for chat, classification, extraction and standard coding help — it keeps ~80–85% of Pro’s quality at a fraction of the price. Switch to Pro for agentic coding loops, hard reasoning and 20+-step workflows where reliability pays back the premium. Schedule batch jobs off-peak to halve the bill.

4. Honest Limits

V4 is the best value in open weights right now, but it’s not magic. Here are the limits I’d plan around before putting it in production.

• Benchmarks above are DeepSeek-reported on their own harness — independent reproduction of the 0813/0731 builds is still pending, so verify on your own tasks.
• Preview-era caveat still applies: capabilities may expand and behavior can shift between builds (Flash-0731 vs Pro-0813 are different post-trains).
• Peak pricing is 2x off-peak and Pro costs ~3x Flash at identical usage — a careless always-Pro default burns budget fast.
• DeepSeek’s hosted API terms allow training on user inputs — route proprietary data through a US inference provider or self-host under MIT.
• Known sensitivity around Chinese political topics; no published system card or formal red-teaming disclosure as of mid-2026.
• Text and tool-call output only — no audio or video generation, and vision is still an experimental Flash-only path.

Golden rule

Benchmark the 1.6-point SWE-bench gap on your own repo before deciding. If Flash passes your CI at the same rate as Pro, you just cut your bill to a third — pocket the difference and only escalate the hard cases.

5. Verdict

For most engineering teams in 2026, DeepSeek V4 Flash is the default and V4 Pro is the escalation path. Flash covers the high-volume majority after the 0731 post-train, while Pro earns its premium on agentic coding, competitive programming and long-horizon reasoning.

That combination — 1M context, MIT weights, 80.6% SWE-bench-class coding at open-model prices — puts V4 on every shortlist I build. Just respect the preview labels, verify on your workload, and watch the peak/off-peak clock.

Who should pick what

Pick Flash if: you run high-volume chat, review, refactor or extraction workloads where latency and cost dominate.

Pick Pro if: agentic coding is the bottleneck, you need max reasoning effort, or long-context analysis quality gates revenue.

Sources

Every version name, spec and price above comes from DeepSeek’s own docs — never from rumor threads. Benchmark figures are DeepSeek-reported and labeled as such until independent reproductions land.

If you found this useful, check the two related open-model guides below — they use the same quickstart pattern so you can compare Qwen, GPT-OSS and DeepSeek on equal footing.