If you’re trying to work out whether DeepSeek can replace Claude or GPT-5 in your stack, the answer changed on April 24, 2026. That is when DeepSeek shipped V4 Preview as two open-weight models — `deepseek-v4-pro` and `deepseek-v4-flash` — and the DeepSeek capabilities reset that came with them is the biggest shift since R1. Million-token context became the default, thinking mode became a parameter rather than a separate model, and prices stayed at a fraction of frontier closed-source rates. This guide is what I wish I’d had on launch day: every capability that matters in production, with the benchmark numbers, pricing math, and limitations stated plainly so you can decide what to build on.

The short answer: what DeepSeek can do in April 2026

DeepSeek today is a family of two frontier-adjacent open-weight models with a chat app, a developer API, and a Hugging Face release. DeepSeek-V4-Pro has 1.6T parameters (49B activated) and DeepSeek-V4-Flash has 284B parameters (13B activated) — both supporting a context length of one million tokens. Both are released under the MIT license, both support a three-way thinking-mode switch, and both expose the same OpenAI- and Anthropic-compatible API surface.

In practical terms, here is what that buys you:

• Long-document reasoning. 1,000,000-token input window with up to 384,000 tokens of output.
• Frontier-tier coding. V4-Pro achieves 80.6% on SWE-bench Verified — within 0.2 points of Claude Opus 4.6.
• Agentic workflows. V4 has special optimizations for mainstream Agent frameworks such as Claude Code, OpenClaw, and CodeBuddy.
• Three reasoning effort levels. Non-thinking, thinking (high), and thinking-max — selected per request, not per model.
• Open weights. Both V4 tiers ship on Hugging Face under MIT, so you can self-host if compliance demands it.

The two-model family, in plain English

Treat V4 as a tier choice, not a model choice. Picking between Pro and Flash is the most consequential decision a new DeepSeek developer makes, because it dictates both cost and ceiling.

V4-Pro: the frontier-tier model

DeepSeek-V4-Pro is the flagship: 1.6 trillion total parameters, 49 billion active per token, pre-trained on 33 trillion tokens. This is the model DeepSeek calls ‘the best open-source model available today.’ Use Pro when you need agentic coding work, hard reasoning over long contexts, or competition-grade math.

V4-Flash: the cost-efficient sibling

DeepSeek-V4-Flash is the efficiency play: 284 billion total parameters, 13 billion active per token, trained on 32 trillion tokens. The 13B active parameter count is in the same ballpark as many mid-range models — but Flash has access to 284B worth of specialized expert knowledge. Flash is what I run for chat, customer-facing assistants, and high-volume retrieval pipelines.

For a deeper dive on each tier, see the DeepSeek V4-Pro and DeepSeek V4-Flash model pages, or the broader DeepSeek V4 overview.

Benchmark capabilities: where V4-Pro lands

I treat self-reported benchmarks as a starting point, not a verdict. Here are DeepSeek’s published numbers from the V4 technical report, cross-checked against third-party reporting from launch day. Frontier numbers shift monthly — verify before you procure.

The honest read: V4-Pro does not appear to dethrone GPT-5.5 or Claude Opus 4.7 on the benchmarks that can be directly compared, but it gets close enough on several of them — especially BrowseComp, Terminal-Bench 2.0 and MCP Atlas — that its much lower API pricing becomes the headline. Where V4-Pro clearly leads is in coding-arena results: Vals AI noted that DeepSeek V4 is “now the #1 open-weight model on our Vibe Code Benchmark, and it’s not close.”

The architecture that makes 1M context affordable

A million-token window is meaningless if every forward pass costs ten times what it did at 128K. The reason V4 ships 1M as the default is an attention rewrite. DeepSeek-V4 designs a hybrid attention mechanism combining Compressed Sparse Attention (CSA) and Heavily Compressed Attention (HCA) to dramatically improve long-context efficiency. In the 1M-token context setting, DeepSeek-V4-Pro requires only 27% of single-token inference FLOPs and 10% of KV cache compared with DeepSeek-V3.2.

V4-Flash drops further to 10% of FLOPs and 7% of KV cache. That is what makes long agent trajectories — SWE-bench tasks, multi-hour browse sessions, terminal sessions with hundreds of commands — economically viable on an open model. For a refresher on what changed between generations, the DeepSeek V3.2 page covers the previous-generation baseline.

Thinking modes: a parameter, not a model

In V4, you no longer pick between deepseek-chat and deepseek-reasoner. You pick a tier (deepseek-v4-pro or deepseek-v4-flash) and set the reasoning effort:

• Non-thinking (default). No chain-of-thought; fastest and cheapest.
• Thinking (high). Set reasoning_effort="high" and pass extra_body={"thinking": {"type": "enabled"}}. The response returns reasoning_content alongside the final content.
• Thinking-max. Set reasoning_effort="max" for the heaviest reasoning budget; for the Think Max reasoning mode, set the context window to at least 384K tokens.

A minimal Python call against V4-Pro in thinking mode, using the OpenAI SDK:

from openai import OpenAI

client = OpenAI(
    base_url="https://api.deepseek.com",
    api_key="YOUR_KEY",
)

resp = client.chat.completions.create(
    model="deepseek-v4-pro",
    messages=[{"role": "user", "content": "Plan the migration from V3.2 to V4."}],
    reasoning_effort="high",
    extra_body={"thinking": {"type": "enabled"}},
)
print(resp.choices[0].message.content)

Chat requests hit POST /chat/completions, the OpenAI-compatible endpoint at https://api.deepseek.com. Both models support the OpenAI ChatCompletions format and the Anthropic API format. Swap the SDK and you can reuse Anthropic-flavoured client code against the same base URL. For full request/response details, see the DeepSeek API documentation.

Important: the API is stateless

The web chat at chat.deepseek.com keeps your session history. The API does not. Every POST /chat/completions request must include the entire conversation in the messages array — DeepSeek’s servers do not remember prior turns on your behalf. That is a contrast with the app, and it is the single most common source of confusion for first-time integrators.

Developer features worth knowing

Temperature guidance from DeepSeek’s own docs: 0.0 for code and math, 1.0 for data analysis, 1.3 for general chat and translation, 1.5 for creative writing. Default top_p is 1.0.

Pricing and a worked cost example

As of April 2026, DeepSeek’s published rates per 1M tokens are:

A V4-Flash worked example — 1,000,000 calls per month, with a 2,000-token system prompt cached across calls, a 200-token user message (uncached), and a 300-token response:

• Cached input: 2,000 × 1,000,000 = 2,000,000,000 tokens × $0.028/M = $56.00
• Uncached input: 200 × 1,000,000 = 200,000,000 tokens × $0.14/M = $28.00
• Output: 300 × 1,000,000 = 300,000,000 tokens × $0.28/M = $84.00
• Total: $168.00

The same workload on V4-Pro costs ~$1,682.00 — roughly ten times more, because Pro’s output rate is $3.48/M versus Flash’s $0.28/M. Default to Flash unless your workload demonstrably needs the Pro lift on agentic coding or hard reasoning. Off-peak discounts are not active; DeepSeek discontinued them on 2025-09-05 and did not reintroduce them with V4. Build a more detailed model with the DeepSeek pricing calculator or read the dedicated DeepSeek API pricing breakdown.

Legacy IDs and the migration window

If you have older integrations using deepseek-chat or deepseek-reasoner, they still work — for now. deepseek-chat and deepseek-reasoner will be fully retired and inaccessible after Jul 24th, 2026, 15:59 (UTC Time) (currently routing to deepseek-v4-flash non-thinking/thinking). Migration is a one-line change to your model= parameter; the base_url stays the same. I’d schedule the swap in May rather than June — the Pro tier is worth testing before the deadline pressure.

What DeepSeek still cannot do well

Open-weight does not mean omnipotent. Honest limitations as of April 2026:

• Multimodal input. DeepSeek said the current version focuses on text, but it is preparing a multimodal AI roadmap. If you need vision, look at the older DeepSeek VL2 family or a different provider.
• Closed-model parity on the hardest reasoning. On Humanity’s Last Exam without tools, DeepSeek-V4-Pro-Max scores 37.7% versus 46.9% for Claude Opus 4.7 (VentureBeat, April 2026). It is competitive, not dominant.
• Self-hosting cost. Pro is 865GB on Hugging Face, Flash is 160GB. Pro needs tens of GPUs for any reasonable throughput; Flash is more realistic on a well-equipped single server.
• Privacy posture. Hosted requests are processed on servers subject to Chinese law. If that matters for your data, self-host the weights or pick another provider. Read the DeepSeek privacy guide before committing.
• Regional availability. Italy’s Garante restricted the consumer app in January 2025; some US states limit DeepSeek on government devices. The availability-by-country guide tracks the current status.

For a fuller list, the DeepSeek limitations page is the deeper companion to this section.

Where DeepSeek capabilities fit best

After three weeks of running V4 in production alongside Claude and GPT-5, my heuristics:

1. High-volume chat or RAG? V4-Flash. The output rate is the cheapest of any frontier-adjacent API I’ve tested in April 2026 — pair it with the DeepSeek context caching guide to push effective costs lower.
2. Agentic coding loops? V4-Pro with thinking mode enabled. The Codeforces and LiveCodeBench numbers translate well to real PR-fixing work.
3. Long-document analysis? Either tier with the 1M context. Flash is fine for summarisation; Pro earns its premium when you need cross-document reasoning.
4. Anything requiring vision or audio? Use a different model.

For role-specific recipes, see DeepSeek for coding and DeepSeek for developers, or browse the wider DeepSeek beginner guides hub.

Last verified: 2026-04-25. 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.