DeepSeek V4 Flash Ships at $0.14/M: 79% SWE-Bench Verified on 13B Active Parameters
DeepSeek shipped both V4 variants live today, April 24. The Flash is the story worth watching closely.
The Flash Numbers
DeepSeek-V4-Flash is a 284B parameter Mixture-of-Experts model that activates only 13B parameters per forward pass. That architecture choice is why the pricing works: $0.14/M input, $0.28/M output, with cache hits at $0.028/M — a 97.5% discount on repeated context.
Performance against V4-Pro shows a narrow gap on most benchmarks:
| Benchmark | V4-Flash | V4-Pro | Delta |
|---|---|---|---|
| SWE-Bench Verified | 79.0% | 80.6% | -1.6 pp |
| MMLU-Pro | 86.2% | 87.5% | -1.3 pp |
| LiveCodeBench | 91.6% | 93.5% | -1.9 pp |
| Codeforces | 3052 | 3206 | -4.8% |
| Terminal-Bench 2.0 | 56.9% | 67.9% | -11 pp |
| SimpleQA-Verified | 34.1% | 57.9% | -23.8 pp |
The 1-2 point gap on coding benchmarks is within run variance. The 11-point drop on Terminal-Bench 2.0 and the 24-point drop on SimpleQA-Verified are the real tells: Flash weakens on complex multi-step tool use and detailed factual recall. For most automated coding pipelines, that tradeoff is acceptable.
Against the Field
V4-Flash’s 79.0% SWE-Verified puts it ahead of Gemini 2.5 Pro (78%) and Claude Sonnet 4.6 (73%), level with GPT-5.4 within error bars, and 1.6 points behind V4-Pro and Claude Opus 4.6 (both at ~80.6-80.8%).
Pricing comparison against models at similar capability:
| Model | Input $/M | Output $/M | SWE-Verified |
|---|---|---|---|
| DeepSeek V4-Flash | $0.14 | $0.28 | 79.0% |
| DeepSeek V4-Pro | $1.74 | $3.48 | 80.6% |
| GPT-5.4 | $2.50 | $15.00 | ~80% |
| Claude Opus 4.6 | $5.00 | $25.00 | 80.8% |
V4-Flash is 18x cheaper on input and 54x cheaper on output than Claude Opus 4.6 while sitting 1.8 percentage points behind on SWE-Verified.
Architecture: DSA and Long Context
Both V4 models introduce DeepSeek Sparse Attention (DSA) with token-wise compression, which DeepSeek says achieves “world-leading long context with drastically reduced compute and memory costs.” The 1M token context is standard across both models with no surcharge — a direct contrast to GPT-5.4 and Claude Opus 4.6, which apply 2x pricing past 272K and 200K tokens respectively.
Maximum output is 384K tokens. Both models support thinking and non-thinking modes, JSON output, tool calls, and OpenAI-compatible as well as Anthropic-compatible APIs.
Retirement Timeline
DeepSeek is retiring the legacy deepseek-chat and deepseek-reasoner endpoints on July 24, 2026. Both are currently routing to V4-Flash (non-thinking) and V4-Flash (thinking) respectively. Developers on the old endpoints will be migrated automatically until that date; explicit model selection is available now.
Open Weights
Both models are MIT-licensed on Hugging Face. V4-Flash weights at 284B total are significantly more practical to self-host than V4-Pro at 1.6T. For organisations running private inference on current-generation hardware, Flash is the version that fits.