benchrPricing breakdown
| Tier | Peak / 1M tokens | Off-peak / 1M tokens |
|---|---|---|
| Input (cache miss) | $0.44 | $0.22 |
| Output | $1.32 | $0.66 |
| Input (cache hit) | $0.014 | $0.007 |
| Context window | 1,000,000 tokens | 1,000,000 tokens |
The clock is now part of the bill
DeepSeek moved every V4 model onto peak and off-peak rates at 16:00 UTC on August 16, 2026, alongside the V4 lineup release. Peak hours are 01:00–04:00 and 06:00–10:00 UTC, Monday to Friday. Everything outside that window — evenings, nights, and all weekend — bills at half the peak rate.
That leaves 35 peak hours in a 168-hour week, so about 79% of the week sits at the off-peak rate. That share is our arithmetic from the published hours, not a DeepSeek figure. The catch is where the peak window falls: the European morning and the Asian afternoon, which is exactly when an interactive product serves its users. Scheduled work — nightly indexing, evaluation runs, bulk summarisation — can be moved. A chat feature cannot.
What the published figures show
DeepSeek's V4-Flash model card reports 79% on SWE-bench Verified. This is a provider-published figure, not an independently reproduced result. Scores published by different providers may use different model settings, scaffolds, or evaluation dates, so a cross-provider score comparison does not establish equivalent production quality. Run a controlled task evaluation before switching traffic.
MIT license and self-hosting
DeepSeek's official weights repository labels V4-Flash under the MIT License. The license permits use, copying, modification, distribution, sublicensing, and sale, subject to retaining the copyright and permission notice; it also includes an “as is” warranty disclaimer. Self-hosting replaces API charges with hardware, serving, staffing, and energy costs, and its economics depend on the deployment. Security, privacy, and data-residency obligations still require a deployment-specific review.
Cache pricing: $0.014 peak, $0.007 off-peak
Cached input costs $0.014/1M at peak and $0.007/1M off-peak — about 97% below the matching cache-miss rate, roughly the same proportion as before the change. At a 90% cache-hit rate the blended input rate works out to about $0.057/1M at peak and $0.028/1M off-peak. Both are still cheap, but the cache-hit line rose five-fold in absolute terms, from $0.0028, so a prompt-caching design that was effectively free is now a line item worth measuring. Actual savings depend on which tokens qualify for cache hits and the hit rate your workload achieves.
Cost scenarios
At 50M uncached input + 15M output tokens per month, the listed rates produce $22 + $19.80 = $41.80/month if every request lands in the peak window, or $11 + $9.90 = $20.90/month if none do. The same volume cost $11.20 before August 16. At 500M input + 150M output, the range is $209 to $418/month, against $112 before. These are token-only illustrations calculated from the listed rates; taxes, platform fees, retries, tools, and hosting overhead may change the production total.
Use-case fit
Consider for: High-volume coding automation with a validated task pass rate, cost-sensitive batch work, or self-hosted deployments where the team can operate the weights securely.
Do not adopt on price alone: Measure output quality, latency, throughput, error behavior, support, and governance requirements. A provider-reported benchmark score is not a production acceptance test.
Decision checklist
Before switching production coding traffic to V4-Flash, run a blind evaluation on a representative task set and compare accepted outputs, tool-call success, latency, and failure recovery against your incumbent model.
Confirm data processing, retention, residency, security, and support terms with the provider used for deployment. Self-hosting changes the control boundary but does not remove those obligations.