DeepSeek V4 review, pricing and limits

For engineering teams chasing frontier coding quality without frontier pricing, DeepSeek V4-Pro is the open-weight pick: it trails Claude Opus 4.8 and GPT-5.5 by about 8 SWE-bench points while running open-source under MIT, though its August GA launch replaced flat pricing with a peak and off-peak schedule.

DeepSeek V4-Pro reached general availability on August 13, 2026, after an April preview. Built by DeepSeek under the MIT license, the open-weight Mixture-of-Experts model scores 80.6% on SWE-bench Verified and 90.1% on GPQA Diamond, with a 1-million-token context window and 384K max output.

Provider: DeepSeek · Family: DeepSeek V4

More about DeepSeek on HokAI

Context window: 1,000,000 tokens · Max output: 384,000

Input modalities: text, image, audio, video, tool-calls · Output: text, tool-calls

About DeepSeek V4

DeepSeek V4 launched as a public preview on April 24, 2026, from DeepSeek (a Chinese AI research lab), as a family of open-source Mixture-of-Experts models available in two sizes: V4-Pro (1.6 trillion total parameters, 49B active per token) and V4-Flash (284 billion total parameters, 13B active). Both ship under the MIT license with weights publicly available on HuggingFace, making them among the largest open-weight models to reach production-quality benchmark scores. V4-Pro left preview for general availability on August 13, 2026, and V4-Flash moved to public beta on July 31, 2026. V4 remains the first DeepSeek release with native multimodal architecture, processing text, images, audio, and video in a unified pipeline. The general-availability release strengthened agentic capabilities, added native support for the Responses API format with Codex-style tool integration, and introduced three selectable thinking-effort levels (low, high, max) in place of the preview's single standard/max-effort toggle. DeepSeek also shipped a new experimental model, V4-Flash-Vision-Exp, pairing V4-Flash's text performance with stronger visual understanding aimed at agent tasks. The GA milestone also brought a new peak/off-peak pricing schedule, replacing the flat preview-era rate (full breakdown in the pricing FAQ below). DeepSeek V4-Pro scores 80.6% on SWE-bench Verified (agentic coding), 90.1% on GPQA Diamond (graduate-level reasoning), 90.1% on MMLU (multitask language understanding), 73.5% on MMLU-Pro, approximately 96.4% on HumanEval, 79.5% on ARC-AGI-2, and 85.0% on AIME 2025. Some of these figures carry the disclaimer 'internal claim only' as of June 2026; third-party verification is still ongoing. At 80.6% SWE-bench Verified, V4-Pro is within 0.2 points of Claude Opus 4.6 and rivals GPT-5.5 on agentic coding at roughly one-thirtieth of GPT-5.5's per-token cost. One source also cites 91.2% SWE-bench on a harder evaluation variant; treat both with caution until full independent verification is published. Both V4-Pro and V4-Flash support a 1M-token (1,000,000) context window with up to 384,000 tokens of output per request. The V4 architecture introduces a hybrid attention mechanism combining Compressed Sparse Attention (CSA) and Heavily Compressed Attention (HCA), requiring only 27% of single-token inference FLOPs and about a tenth of the KV cache compared to DeepSeek-V3.2 at the 1M-token context setting. This efficiency gain makes very long-context inference feasible even on constrained hardware. DeepSeek V4 natively accepts text, images, audio, and video. Image understanding covers OCR and technical schematic interpretation, with performance comparable to GPT-5 on standard vision benchmarks. Video processing handles short clips under ten minutes for temporal question answering. Audio supports transcription plus speaker identification. V4 is not a bolt-on vision module: multimodality is part of the base architecture. Function calling, structured output, and tool use are confirmed available. Audio and video output are not supported; the model produces text and tool-calls only. V4-Pro API pricing moved to a peak/off-peak schedule as part of the August 2026 GA release, roughly doubling during weekday UTC business hours versus overnight and weekend rates (exact figures in the pricing FAQ below). V4-Flash remains the cheaper sibling, at a fraction of V4-Pro's per-token rate on both input and output. For self-hosted deployments under the MIT license, there are no per-token API costs at all. A daily coding-agent workload now costs meaningfully more than under April 2026 preview pricing, though still well under what the same workload costs on GPT-5.5. V4-Pro is available via the DeepSeek API, Together AI, Fireworks AI (among the fastest peak throughput of any third-party provider on Artificial Analysis), Azure, and Vercel AI Gateway. Self-hosting is possible under the MIT license. V4-Pro requires roughly 865GB of storage for full-precision weights, necessitating multi-node GPU infrastructure across several datacenter-class accelerators. V4-Flash is more accessible: roughly 33GB VRAM heavily quantized (1x RTX 6000 Ada or 2x RTX 4090), 80GB on a single datacenter GPU at FP8, or 170GB for full weights plus KV cache. Community GGUF quantizations (antirez/deepseek-v4-gguf on HuggingFace) are available for V4-Flash. Safety considerations for DeepSeek V4 are material for enterprise users. DeepSeek's own API Terms of Service allow the company to train on user inputs, which is a significant data privacy risk for proprietary data. US-based inference providers (Together AI, Fireworks AI) avoid this by running DeepSeek model weights on US servers under their own data terms. As an open-source Chinese model, V4 carries known sensitivity around Chinese political topics while being relatively permissive on other content. For air-gapped or fully controlled deployments, self-hosting under the MIT license is the recommended path. The model does not have a published system card or formal red-teaming disclosure as of June 2026. DeepSeek V4-Pro is best suited for teams doing agentic coding, research, and long-context analysis who need strong benchmark results at a fraction of closed-model pricing. It trails Claude Opus 4.8 (88.6%) and GPT-5.5 (88.7%) on SWE-bench Verified, but as the only open-weight model in that tier, it is the only one of the three a team can self-host or fine-tune. For teams with data privacy requirements, routing through Together AI or Fireworks AI rather than DeepSeek's own API resolves the ToS training-data concern. Organizations with compliance needs (HIPAA, enterprise data governance) should self-host under the MIT license. Teams needing audio or video output, or requiring sub-second TTFT for real-time voice applications, should look elsewhere, as V4 produces text output only. DeepSeek V4 was pre-trained on more than 32 trillion diverse and high-quality tokens. The training pipeline uses the Muon optimizer for faster convergence and Manifold-Constrained Hyper-Connections (mHC) to strengthen residual connections. Training data cutoff has not been explicitly published; the April 2026 release suggests data through at least late 2025. As an MIT-licensed open-source model, V4 weights can be fine-tuned, distilled, or modified without restriction, making it a base for derivative models. DeepSeek V4 was previewed to select users before April 24, 2026, with the public preview shipping that day. V4-Pro-Max is a maximum reasoning effort mode with higher latency but stronger benchmark performance. The predecessor DeepSeek V3.2 introduced a hybrid MoE architecture in August 2025. DeepSeek V3-0324 (March 2025) improved AIME scores by 19.8 points over V3. Both V4-Flash and V4-Pro have been available on Fireworks AI since shortly after launch. DeepSeek's prior API endpoints, deepseek-chat and deepseek-reasoner, were retired on July 24, 2026, as previously announced.

Pricing

V4-Pro (DeepSeek API, effective August 16, 2026): $0.66/M cache-miss input off-peak ($1.32/M peak), $1.98/M output off-peak ($3.96/M peak). Cache-hit input is $0.022/M off-peak ($0.044/M peak). V4-Flash: $0.22/M cache-miss input off-peak ($0.44/M peak), $0.007/M cache-hit, $0.66/M output off-peak ($1.32/M peak). Peak hours run 01:00-04:00 and 06:00-10:00 UTC, Monday to Friday; every other hour is off-peak, at exactly half the peak rate. Third-party providers (Fireworks, Together) may vary. Self-hosting is free under the MIT license; VRAM costs apply.

What a real job costs

Budgets: 20-page PDF = 30k in / 1k out · Support reply = 2k in / 300 out · Coding agent run = 200k in / 20k out. Computed from the vendor's per-token prices at render time; cached-input discounts are not applied.

Key Features

• MIT License and Open Weights: Fully open weights available on HuggingFace under MIT license. Self-host, fine-tune, distill, or modify without restriction. No per-token fees when self-hosted.
• 1.6T MoE with 49B Active Parameters: Sparse activation means only 49B of 1.6T parameters activate per token, enabling frontier-quality inference at dramatically lower compute cost than dense models of equivalent benchmark score.
• 384K Max Output Tokens: The highest max output of any frontier model: handles full codebase rewrites, long-form reports, or entire document translations in one API call.
• CSA/HCA Hybrid Attention: Novel attention mechanism using Compressed Sparse Attention and Heavily Compressed Attention reduces 1M-context KV cache to 10% of DeepSeek-V3.2, enabling long-context inference on constrained hardware.
• V4-Flash Variant for Cost Optimization: 284B MoE sibling priced well below V4-Pro on every tier even after the August 2026 GA increase, and runs on a single 80GB-class datacenter GPU at FP8. Trades a small benchmark delta for a large reduction in API cost versus V4-Pro.
• Three-Level Thinking Effort Control: The August 2026 GA release replaced the preview's single standard/max toggle with low, high, and max thinking-effort levels, plus native Responses API support for Codex-style tool integration.

Pros

• Matches frontier closed models on agentic coding benchmarks at a fraction of GPT-5.5's per-token cost, even after the August 2026 GA price increase.
• MIT license and open weights on HuggingFace enable self-hosting, fine-tuning, and full data control.
• The highest max output token limit of any frontier model, eliminating multi-turn workarounds for large generation tasks.

Cons

• DeepSeek's API ToS allows training on user inputs by default, creating enterprise data privacy risk when using api.deepseek.com directly.
• V4-Pro self-hosting requires a multi-node GPU cluster, a significant infrastructure commitment beyond a single server.
• The August 2026 GA pricing introduced a peak/off-peak split that doubles cost during certain weekday UTC hours (see the pricing FAQ); several headline benchmark figures also remain vendor-reported and have not been independently reproduced.

Benchmarks

• MMLU: 90.1% vendor-reported · 03 Sep 2026 — General-knowledge exam across 57 subjects, % correct.
• MMLU-Pro: 87.5% vendor-reported · 03 Sep 2026 — A harder version of the 57-subject knowledge exam, % correct.
• AIME 2025: 85% vendor-reported · 03 Sep 2026 — Competition-level maths problems from the 2025 exam, % solved.
• ARC-AGI 2: 79.5% vendor-reported · 03 Sep 2026 — Abstract visual puzzles built to resist memorisation, % solved.
• HumanEval: 76.8% vendor-reported · 03 Sep 2026 — Small programs that must pass hidden tests, % passing.
• GPQA Diamond: 90.1% vendor-reported · 03 Sep 2026 — PhD-level science questions that are hard to search for, % correct.
• SWE-bench Verified: 80.6% vendor-reported · 03 Sep 2026 — Real GitHub issues fixed end to end, % solved.
• AA Intelligence Index: 53 cited: Artificial Analysis · 03 Sep 2026 — Composite of 10 evaluations run by Artificial Analysis, 0 to 100.
• Output speed: 58 tok/s cited: Artificial Analysis · 03 Sep 2026 — Median tokens written per second as measured by Artificial Analysis.

A benchmark is an exam, not the job. Scores transfer unevenly between tasks, so weigh the one closest to your workload and read every figure with its source.

Visit DeepSeek V4 Official Page