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 | Peak / 1M tokens | Off-peak / 1M tokens |
|---|---|---|
| Input (cache miss) | $1.32 | $0.66 |
| Output | $3.96 | $1.98 |
| Input (cache hit) | $0.044 | $0.022 |
| Context window | 1,000,000 tokens | 1,000,000 tokens |
| Max output | 384,000 tokens | 384,000 tokens |
| License | MIT (self-hostable) | MIT (self-hostable) |
When the request runs is now a pricing decision
DeepSeek moved its V4 models onto peak and off-peak rates at 16:00 UTC on August 16, 2026. Peak hours are 01:00–04:00 and 06:00–10:00 UTC, Monday to Friday; everything else bills at half. That is 35 peak hours in a 168-hour week, so about 79% of the week sits at the lower rate — our arithmetic from the published hours, not a DeepSeek figure. Work you can schedule pays the off-peak rate. Work tied to European or Asian business hours mostly does not.
Listed price compared with commercial alternatives
At medium production volume — 10M input tokens and 3M output tokens per month — V4-Pro costs $13.20 + $11.88 = $25.08/month at peak, or $12.54/month entirely off-peak. Before August 16 the same volume was $6.96. GPT-5 at that volume is $12.50 + $30 = $42.50/month and Claude Opus 4.8 is $50 + $75 = $125/month, both unchanged. So V4-Pro still undercuts GPT-5 by about 41% at peak and 70% off-peak, and Opus 4.8 by 80% and 90% — but the gap against GPT-5 used to be 84%.
At 100M input + 30M output tokens per month, the listed rates model to $250.80/month at peak and $125.40/month off-peak for V4-Pro, against $1,250/month for Claude Opus 4.8 — an annual difference of about $11,990 at peak and $13,495 off-peak, before caching, batch discounts, infrastructure, support, or retries. Treat this as a rate-card scenario, not a forecast of total cost.
Caching still pays, but it is no longer a rounding error
Cached input on V4-Pro is listed at $0.044/1M at peak and $0.022 off-peak, up from a flat $0.003625 — the steepest proportional rise on the sheet, about twelve times the old rate. As a simple model, a 100K-token system prompt sent ten times, with one cache miss and nine hits, costs about $0.17 of input at peak and $0.086 off-peak, against $1.32 and $0.66 if all ten requests miss the cache. The ratio is as good as it ever was; the absolute number is the part that changed. Eligibility, writes, expiry, and observed hit rate can change the result.
Published benchmark profile
The cited scorecard reports 96.1% on MATH, 90.1% on GPQA Diamond, 93.5% on HumanEval, and 92.8% on MMLU. These are benchmark-specific, dated figures that benchr did not independently measure, and they do not guarantee quality on your workload. Check the evaluation setup and run representative prompts before relying on the comparison.
The 384K max output token limit is another practical difference: nearly 3× the 128K cap on GPT-5, GPT-5.5, and Claude. That extra room matters for workloads that generate large artifacts, including full codebases, long documents, and synthetic datasets.
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 other visual data, evaluate a current vision-capable model on those exact inputs. Benchr's editorial capability profiles are not provider benchmarks or measured guarantees.
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 listed as publicly available under the MIT license. Self-hosting replaces hosted API charges with GPU, storage, networking, monitoring, security, and engineering costs. The practical barrier is GPU memory: the full model requires significant VRAM, and inference optimization such as quantization and 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
Candidate for: cost-sensitive coding and math, long-context text processing, workloads that can benefit from verified cache hits, and teams prepared to evaluate self-hosting operations.
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 cited benchmarks and rate card justify a workload-specific trial, not an automatic production choice. Compare exact model checkpoints under the same prompts and acceptance criteria, then include retry, support, security, and infrastructure costs in the decision.
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.