DeepSeek V4 Flash API Pricing
The cheapest frontier-class agent model available. Here is exactly what it costs and how it compares.
DeepSeek V4 Flash vs Competitors — Price Per Million Tokens
| Model | Input/M | Output/M | Context | Active Params | License |
|---|---|---|---|---|---|
| DeepSeek V4 Flash | $0.14 | $0.28 | 1M | 13B | MIT |
| DeepSeek V4 Pro | $0.435 | $0.87 | 1M | 49B | MIT |
| Qwen3.7 Flash | $0.10 | $0.50 | 256K | ~10B | Apache 2.0 |
| GPT-5.5 Mini | $0.60 | $2.40 | 128K | N/A | Closed |
| Claude Sonnet 4.6 | $3.00 | $15.00 | 200K | N/A | Closed |
Self-Hosting Cost Math
Community benchmarks report 400 tokens per second at approximately $10 per hour on standard GPU rental hardware. That translates to roughly 1.44M tokens per hour, or about $6.94 per million tokens — more expensive than the API for light use, but dramatically cheaper at sustained throughput since you pay per hour, not per token.
The crossover: if you process more than roughly 70M tokens per hour consistently, self-hosting beats API pricing. Below that, the API is simpler and cheaper. The 13B active parameter count means V4 Flash runs on hardware that would choke on V4 Pro's 49B active.
# Quick self-hosting estimate
400 tps × 3600s = 1,440,000 tokens/hour
$10/hour ÷ 1.44M tokens = ~$6.94/M tokens (all-in)
API: $0.14/M input + $0.28/M output ≈ $0.21/M blended
# API wins until ~70M tokens/hour sustained
Pricing FAQ
Is DeepSeek V4 Flash the cheapest frontier model?
Among models with verified agent-tier benchmark scores, yes. At $0.14/M input it undercuts V4 Pro by 3x, GPT-5.5 Mini by 4x, and Claude Sonnet by over 20x. Only Qwen3.7 Flash ($0.10/M) is cheaper but lacks comparable agent benchmarks.
Is V4 Flash cheaper to self-host or use via API?
For light usage, the API wins — no infrastructure to manage. For sustained heavy workloads (millions of tokens per hour), self-hosting at ~$10/hour becomes cheaper than API pricing. The crossover is roughly 70M tokens per hour.
Does V4 Flash have rate limits?
The API is in public beta. DeepSeek has not published specific rate limits for V4 Flash yet. Third-party providers like ZenMux and Infermatic also offer V4 Flash endpoints with their own rate policies.
Will pricing change after beta?
Possibly. Beta pricing is often promotional. However, DeepSeek has historically maintained aggressive pricing — V3's pricing barely moved from beta to GA. The MIT license means self-hosting is always an escape valve.