benchrHead-to-head: the rate card against the closed mid-tier is in V4-Pro vs Sonnet 4.6 and V4-Pro vs GPT-5.
Pricing breakdown
| Tier | Rate / 1M tokens |
|---|---|
| Standard input | $0.435 |
| Standard output | $0.870 |
| Cached input | $0.004 |
| Context window | 1,000,000 tokens |
| Max output | 384,000 tokens |
| License | MIT (self-hostable) |
The cost gap is not subtle
At medium production volume — 10M input tokens and 3M output tokens per month — the numbers are stark: DeepSeek V4-Pro costs $4.35 + $2.61 = $6.96/month. GPT-5 at the same volume: $12.50 + $30 = $42.50/month. Claude Opus 4.8: $50 + $75 = $125/month. DeepSeek saves 84% over GPT-5 and 94% over Opus 4.8 — at identical context window size and comparable coding benchmarks.
At 100M input + 30M output, the annual difference against Claude Opus 4.8 exceeds $60,000. For a funded team, that's an engineering hire.
Caching at near-zero cost
Cached input on DeepSeek V4-Pro costs $0.004/1M — 99% off the standard input rate. Compare: Anthropic charges $0.50/1M cached (90% off), Google $0.15/1M (90% off), OpenAI doesn't publish a cached rate for GPT-5. For agent workloads with large, repeated system prompts, DeepSeek's cache effectively reduces input cost to rounding error. A 100K-token system prompt cached 9 times out of 10 costs about $0.04 total, not $3.92.
Where DeepSeek V4-Pro excels
The model's benchmark profile is strongest on math and long-context reasoning. MATH benchmark: 96.1% — highest of any model tracked here. GPQA Diamond (graduate-level science questions): 90.1%, ahead of GPT-5.5. HumanEval: 93.5%. MMLU: 92.8%. These are not budget-model numbers.
The 384K max output token limit also stands out: nearly 3× the 128K cap on GPT-5, GPT-5.5, and Claude. For workloads that generate large artifacts — full codebases, long documents, synthetic datasets — this matters in practice.
What it can't do
DeepSeek V4-Pro has no vision or multimodal capability. Image inputs are not supported. If your pipeline touches screenshots, PDFs with figures, or any visual data, you need a different model for that step — Gemini 3.5 Flash (vision score 92) or Claude Opus 4.8 (86).
Enterprise SLAs, US-based data residency, and the ecosystem integrations of the OpenAI or Anthropic APIs are also absent. For regulated industries or infrastructure teams with strict data-handling requirements, these constraints matter regardless of the price advantage.
The MIT license and self-hosting path
DeepSeek V4-Pro weights are publicly available under the MIT license. Self-hosting on your own GPU infrastructure eliminates API fees entirely — you pay only for compute. For teams already running GPU clusters, the marginal cost of inference on a model you've already deployed approaches zero. The practical barrier is GPU memory: the full model requires significant VRAM, and inference optimization (quantization, batching) demands engineering time.
Most teams start with the hosted API and evaluate the self-hosting case only when monthly API costs exceed the annualized cost of the GPU capacity required.
Use-case fit
Best for: Frontier-grade coding and math at low cost; long-context document processing; agentic workloads where caching eliminates most input cost; self-hosted production at near-zero API spend.
Skip if: You need vision or image input; your infrastructure requires US-based data residency or enterprise SLAs; your team is already deeply integrated into the OpenAI or Anthropic API ecosystems and switching cost exceeds the savings.
The honest framing: DeepSeek V4-Pro benchmarks are real, not cherry-picked. The model genuinely competes with models costing 3–12× more. The cost case is so strong that the burden of proof falls on the reasons not to use it, not on the reasons to use it.
Decision checklist
Before routing production traffic to DeepSeek V4-Pro: confirm your workload is text and code only (no image inputs); verify latency from your infrastructure region against the DeepSeek API endpoint; run a small eval on your actual task type — math and coding benchmarks translate well, but instruction-following style differs from OpenAI/Anthropic models.
If switching from GPT-5 or Claude: test system prompt compatibility, function-calling format, and JSON mode behavior. API surface is similar but not identical. Prompt migration typically takes hours for simple pipelines, longer for complex tool-use agents.