DeepSeek V4 Pro vs V4 Flash: Which Model Should You Use?
DeepSeek V4 Pro vs V4 Flash: 1.6T vs 284B params, $0.435 vs $0.14 per 1M input, 96.4% vs 88.8% SWE-bench. How to pick.
The Two Models at a Glance
V4 Pro is DeepSeek's 1.6T-parameter flagship with 49B active per token; V4 Flash is the 284B/13B workhorse that officially cleared Pro-Preview on nine agent benchmarks back in July[2]. Both share a 1M-token context window, 384K max output, MIT licensing, and the same low/high/max reasoning-effort controls introduced with the August 13 GA[1].
The August 13 release changed the comparison: V4 Pro went from preview to GA as the 0813 checkpoint, while Flash points to the 0731 build. Neither model name changes — you call deepseek-v4-pro and deepseek-v4-flash as always[1].
Concurrency is a real difference: Pro allows 500 concurrent requests, Flash 2,500 — Flash is built for high-throughput pipelines, Pro for depth-first reasoning[1].
| Spec | DeepSeek V4 Pro (0813) | DeepSeek V4 Flash (0731) |
|---|---|---|
| Total parameters | 1.6T | 284B |
| Active per token | 49B | 13B |
| Context window | 1M tokens | 1M tokens |
| Max output | 384K tokens | 384K tokens |
| Concurrency limit | 500 | 2,500 |
| Reasoning effort | low / high / max | low / high / max |
| GA status (Aug 16, 2026) | GA (0813) | Official build (0731) |
Benchmark Differences
On software engineering, third-party runs put the gap at about 8 points: V4 Pro 0813 scores 96.40% on SWE-bench Verified (Codersera/Vals AI, neutral harness) versus 88.80% for Flash-0731[7]. Both are strong — Flash alone beats GPT-5.6 Sol-adjacent models on patch-style coding at a fraction of the cost.
Official numbers come with a caveat: DeepSeek runs code-agent benchmarks through DeepSeek Harness minimal mode at max reasoning effort, and the harness was not published when the preview numbers appeared. Independent end-to-end agent runs land lower — Vals measured Pro at 54.68% on Terminal-Bench with a reference harness[4].
What matters for you: on patch-style SWE tasks the two models are close enough that cost decides; on long-horizon agentic work the flagship's deeper reasoning shows up more.
| Benchmark | V4 Pro 0813 | V4 Flash 0731 | Notes |
|---|---|---|---|
| SWE-bench Verified (3rd-party) | 96.40% | 88.80% | Codersera / Vals AI[7] |
| Terminal-Bench 2.1 (official) | 87.9 | 82.7 | DeepSeek Harness minimal mode[5] |
| DeepSWE (official) | 62.7 | 54.4 | Vendor-reported[5] |
| AutomationBench (official) | 31.8 | 25.1 | Vendor-reported[5] |
Price Gap & the 8/16 Peak Pricing
At the old flat rates, Flash ran about one-third of Pro's price on input and output: $0.14 vs $0.435 per 1M input (cache miss) and $0.28 vs $0.87 per 1M output[1]. On August 16, 16:00 UTC, DeepSeek switched both models to peak/off-peak pricing — and both got more expensive[3].
Peak hours are 01:00-04:00 and 06:00-10:00 UTC; everything else is off-peak at half the peak rate[3]. The ratio between the models survives the change — Flash stays roughly one-third of Pro — but the absolute floor moved up, so caching and off-peak scheduling matter more than before[3].
The cache-hit discount is the biggest lever for both: Pro input on a cache hit is $0.022 off-peak versus $0.66 on a miss — a 96.7% saving. Long, repetitive agent contexts that reuse a shared prefix benefit enormously from context caching.
| Per 1M tokens | Pro off-peak | Pro peak | Flash off-peak | Flash peak |
|---|---|---|---|---|
| Input (cache miss) | $0.66 | $1.32 | $0.22 | $0.44 |
| Input (cache hit) | $0.022 | $0.044 | $0.007 | $0.014 |
| Output | $1.98 | $3.96 | $0.66 | $1.32 |
Capability Fit: Agentic vs Knowledge Work
Flash is the default for high-volume pipelines: large-scale extraction, classification, chat, and code generation where a 3x price difference dominates. Pro earns its premium on hard reasoning, long-context synthesis, and agentic loops where a single wrong step wastes many tokens[9].
Mindstudio's independent 8-question eval found 0813 at 76.25%, with a real quirk: it over-thinks simple questions and over-rewrites code. Their practical take: 'sometimes little brother V4 Flash is better for daily use'[6]. Match the model to the task, not to the flagship label.
A useful mental model: Flash is the batch worker — huge concurrency, cheap cache hits, 88.8% SWE-bench accuracy for a tenth of a cent per test[7]. Pro is the senior engineer — 96.4% accuracy, deeper reasoning, but 500 concurrency and a higher floor price. If a task fails once in a blue moon and retries are cheap, Flash wins. If a single wrong answer is expensive (code review gates, medical-style QA, long agent runs), Pro's extra eight points justify the premium.
- Batch summarization over millions of tokens → Flash; 2,500 concurrency and cheap cache hits.
- Multi-step coding agent with tool calls → Pro; official TB2.1 87.9 vs 82.7 and better DeepSWE[5].
- Math / logic / world-knowledge Q&A → Pro; GPQA Knowledge 90.1% and higher HLE scores[8].
- Cost-sensitive routing with a quality floor → Flash for easy, Pro for hard; see Step 5.
Independent eval results from Mindstudio's August 13 review[6].
The Hybrid Routing Pattern
The most cost-effective setup is hybrid: default every request to Flash, escalate to Pro only for hard reasoning and agentic tasks. A Flash-first workload runs roughly 3x cheaper end-to-end, and DeepSeek's own 22-benchmark aggregation puts Flash at about 83% of Pro quality[9].
In practice this means one client, two model strings, and a simple classifier in front. Keep cache-friendly prompts identical across calls so the shared prefix hits, and schedule batch jobs into off-peak hours to cut the new peak rates[3].
A concrete routing rule that works well: route anything with tool calls or multi-turn agent state to Pro, route single-shot generation, extraction, and summarization to Flash. Add a fallback: when Pro times out or the request is below a difficulty threshold, downgrade to Flash automatically. Teams using this pattern report roughly 60-70% of tokens landing on Flash while keeping 95%+ of quality-critical tasks on Pro[9].
Quality-ratio claim is DeepSeek's own 22-benchmark aggregation[9]. Verify against your workload before committing.
# pseudo-code for hybrid routing
def route(task: str, agentic: bool) -> str:
if agentic or task.difficulty > THRESHOLD:
return "deepseek-v4-pro"
return "deepseek-v4-flash"
# both models share the same base_url and API key:
# https://api.deepseek.com- [1]DeepSeek Models & Pricing
- [2]DeepSeek V4 Pro GA Release Notes
- [3]DeepSeek V4 Pro 0813 Model Card (Hugging Face)
- [4]ExplainX: V4 Pro 0813 Terminal-Bench Analysis
- [5]Mindstudio: DeepSeek V4 Pro 0813 Review
- [6]Codersera: V4 Pro 0813 Guide & Benchmarks
- [7]BenchLM: V4 Pro vs GPT-5.5 Comparison
- [8]Rohit AI: V4 Pro 0813 GA Benchmarks & Pricing