DeepSeek V4 Pro: a 1M-token reasoning model
DeepSeek's flagship V4 reasoning model. Within 0.2 points of Claude Opus 4.6 on SWE-bench Verified at one-seventh the vendor cost. Claude-compatible API.
1M tokens · Text / Code · Prompt cache
VM0 no longer runs DeepSeek V4 Pro. This page is kept as a reference for its specs, pricing and benchmarks. For the same kind of work, use GPT 5.6 Luna.
See GPT 5.6 LunaDeepSeek V4 Pro is the flagship of DeepSeek's V4 generation — an open-weight 1.6T-parameter MoE under the MIT license. The headline is the price-to-quality ratio: vendor-reported SWE-bench Verified is 80.6%, within a fraction of a point of Claude Opus 4.6, at roughly one-seventh of Anthropic's vendor cost. That makes reasoning-heavy agents — bulk PR review, batch document analysis, scheduled summarisation — affordable at high volume.
Vendor list price is $1.74 / $3.48 per 1M tokens with cache reads at $0.028 / 1M and free cache writes (unique in the lineup). 1M-token context, Anthropic-compatible API. Reach for Sonnet 4.6 when production tool-routing reliability is the deciding factor, and for GPT-5.4 Mini or Kimi K2.7 Code when single-shot bulk work does not need V4 Pro's reasoning depth.
What is DeepSeek V4 Pro?
April 24, 2026 · Reasoning variant of the DeepSeek V4 family, focused on maximum reasoning quality.
DeepSeek V4 Pro is the flagship of DeepSeek's V4 generation, released April 24, 2026 under the MIT License. It's an open-weight Mixture-of-Experts model with 1.6T total parameters and 49B active per token.
V4 Pro supports a 1M-token context window, 384K maximum output, three reasoning effort modes (standard, think, think-max), JSON output, tool calls, and FIM completion in non-think mode. The Pro model adds a hybrid attention architecture (Compressed Sparse Attention + Heavily Compressed Attention) for dramatically improved long-context efficiency. 27% of single-token inference FLOPs and 10% of KV cache vs DeepSeek V3.2 at 1M context.
What's notable about DeepSeek V4 Pro
Headline architecture and capability features.
V4 Pro is a Mixture-of-Experts model with 1.6T total parameters and 49B active per token, fronted by a hybrid attention stack (Compressed Sparse Attention plus Heavily Compressed Attention) that keeps long-context inference cheap. It supports a 1M-token context window with 384K of maximum output, three reasoning effort modes (standard, think, and think-max), and uses Manifold-Constrained Hyper-Connections for stable signal propagation. The model was trained on 32T+ tokens with the Muon optimizer and is released under the MIT License with open weights.
Specs at a glance
DeepSeek V4 Pro benchmarks
Vendor-reported scores from DeepSeek's V4 Pro release. Independent reviews (Geeky Gadgets, Code Arena) place V4 Pro third on Code Arena behind GLM-5.1 and Kimi K2.7 Code. The strongest benchmark claims come from DeepSeek's own materials. Treat directionally rather than as absolute truth.
DeepSeek V4 Pro pricing
Provider list price, per 1M tokens.
How DeepSeek V4 Pro behaves in practice
Observed behaviour from production agent runs.
Reasoning
Strongest sub-Sonnet reasoning in our lineup. Holds up on multi-step work where cheaper models start to drift. Vendor-reported MMLU-Pro matches GPT-5.4.
Coding benchmarks
Vendor-reported SWE-bench Verified 80.6% (within 0.2 of Opus 4.6), Terminal-Bench 2.0 67.9% (leads Opus 4.6), LiveCodeBench 93.5%.
Cost efficiency
The standout property. ×0.1 credit cost with reasoning that competes well with Sonnet 4.6 makes V4 Pro the cost-optimisation default, far cheaper than Claude Opus 4.7.
Speed
Around 36 tokens/sec at max effort per Artificial Analysis. Slower than Kimi K2.7 Code, slightly slower than Opus 4.6.
Best agent tasks for DeepSeek V4 Pro
The PR-review agent that runs on every commit
Sonnet-tier accuracy at a fraction of Sonnet's vendor cost is what makes "review every commit, not just the big PRs" actually viable. V4 Pro reads the diff, the related files, and the linked issue, then writes a structured comment — and the per-call price is low enough that running it as a CI step on every push doesn't show up as a noticeable line item.
The scheduled summariser that runs every night
Pulls yesterday's customer conversations, support tickets, or sales calls and writes a digest. The system prompt and tool schema don't change between runs, and DeepSeek doesn't bill cache writes — so the long fixed prefix is paid for once and cached reads cost a fraction of normal input. This is where V4 Pro's pricing model genuinely changes what's affordable.
The whole-repo code agent that costs less than Opus
1M-token context with hybrid attention (Compressed Sparse Attention plus Heavily Compressed Attention) means a mid-sized codebase fits in one prompt and inference cost stays manageable as the window fills up. For cross-file refactors and architecture-level reviews, this is where you get the Opus-style "see everything at once" workflow without the Opus-style invoice.
When to skip DeepSeek V4 Pro
Skip V4 Pro on the hardest tool-routing edge cases where Sonnet 4.6 still leads, and on bulk single-shot work where GPT-5.4 Mini or Kimi K2.7 Code is sufficient at lower cost.
DeepSeek V4 Pro vs other models
DeepSeek V4 Pro vs Claude Sonnet 4.6
Sonnet 4.6 (×1) wins on tool-routing edge cases and English-language reasoning. V4 Pro (×0.1) wins on cost and is competitive on coding benchmarks (vendor-reported). Worth A/B-testing on a real agent before committing.
Bottom line: should you use DeepSeek V4 Pro?
Pre-filter with GPT-5.4 Mini or Kimi K2.7 Code, escalate to V4 Pro for reasoning, escalate to Sonnet 4.6 only when V4 Pro stalls on tool-routing edge cases.
Frequently asked questions
When was DeepSeek V4 Pro released?
DeepSeek released V4 Pro on April 24, 2026 under the MIT License with open weights.
Why are cache writes free?
DeepSeek doesn't bill the cache-write portion. Only cache reads bill, at $0.145 per 1M tokens. Stable system prompts and large reference contexts cost nothing extra to cache.
What's V4 Pro's context window?
1 million tokens with up to 384K tokens of output. The hybrid attention architecture makes the full window usable at much lower inference cost than V3.2.
How does V4 Pro compare to Claude Opus 4.6?
Vendor-reported SWE-bench Verified is within 0.2 points (80.6 vs 80.8). Terminal-Bench 2.0 favours V4 Pro (67.9 vs 65.4). Opus 4.6 leads on HLE (40.0 vs 37.7) and HMMT 2026 math (96.2 vs 95.2). At ~7× lower vendor cost, V4 Pro is the right call when reasoning quality is the bar but cost matters.
Availability of DeepSeek V4 Pro on VM0
DeepSeek V4 Pro has been removed from the VM0 lineup, so it can no longer be selected in chat or in a workflow, and it is not available with your own API key. GPT 5.6 Luna covers the same cost-saving tier.