benchrPricing breakdown
| Tier | Rate / 1M tokens |
|---|---|
| Standard input | $0.14 |
| Standard output | $0.28 |
| Cached input | $0.0028 |
| Context window | 1,000,000 tokens |
The cost-quality position
DeepSeek V4-Flash scores 79% on SWE-bench Verified — within 0.6 points of Claude Sonnet 4.6's 79.6%, and 4+ points above Claude Haiku 4.5's 73.3%. The pricing differential is staggering: Sonnet 4.6 costs $3/1M input (21× more), Haiku 4.5 costs $1/1M (7× more). On paper, V4-Flash delivers Sonnet-tier coding at Haiku-or-cheaper prices. The practical caveat is that SWE-bench is a benchmark, not a guarantee of performance on your specific task distribution — run your own eval before switching production traffic.
MIT license and self-hosting
DeepSeek V4-Flash weights are MIT licensed — fully open source and commercially usable without restriction. For teams with A100 or H100 GPU access, self-hosting eliminates API costs entirely. At $0.14/1M via the API, the breakeven point for self-hosting is at sustained high volume — roughly 500M+ tokens/month where infrastructure costs begin to rival API fees. Below that threshold, the API is usually cheaper than maintaining dedicated GPU capacity. Above that volume, self-hosting provides cost certainty and eliminates data residency concerns.
Cache pricing: $0.0028 per million
DeepSeek V4-Flash cached input costs $0.0028/1M — 98% off the $0.14 standard rate. For agentic pipelines with large, repeated system prompts, the effective input cost with high cache utilization is approximately $0.017 per million at a 90% hit rate. At this price, input cost becomes essentially negligible for most pipeline designs.
Cost scenarios
At 50M input + 15M output per month: $7 + $4.20 = $11.20/month. Claude Sonnet 4.6 at the same volume: $150 + $225 = $375/month — 33× more expensive. At 500M input + 150M output — large production scale: DeepSeek V4-Flash at $70 + $42 = $112/month. The cost savings compared to Sonnet at scale ($1,500 + $2,250 = $3,750) are not marginal — they're transformative for cost-constrained engineering teams.
Use-case fit
Best for: High-volume coding automation where 79% SWE-bench is sufficient; cost-sensitive startups where Sonnet pricing is prohibitive; self-hosted deployments on existing GPU infrastructure; batch processing pipelines where latency to China doesn't matter.
Skip if: Latency to DeepSeek's China-based API is a constraint for real-time user-facing applications. Also skip when frontier coding quality (88%+ SWE-bench) is required — DeepSeek V4-Pro at $0.43/1M or Claude Opus 4.8 at $5/1M are the appropriate upgrades.
Decision checklist
Before switching production coding traffic to V4-Flash: run a 100-task blind eval on your specific coding task distribution and compare pass rates against Claude Sonnet 4.6. SWE-bench measures overall coding ability; your specific task mix may show a larger or smaller quality gap than the benchmark suggests.
Consider your data residency requirements: DeepSeek's API routes through Chinese infrastructure. For EU GDPR or US government workloads with data locality requirements, use a US-hosted inference provider or self-host on your own infrastructure.