DeepSeek V4 Flash 0731 vs V4 Pro: Cheaper Coding Model (August 2026)
The Short Answer
For cheap coding agents in August 2026, DeepSeek V4 Flash 0731 is the better pick than V4 Pro. Released July 31, 2026, Flash 0731 scored 82.7 on Terminal-Bench 2.1 against V4 Pro Preview’s 72.1, and costs about 6× less ($0.14/$0.28 vs $0.435/$0.87 off-peak). Reach for V4 Pro only on the hardest deep-reasoning tasks.
The Comparison
| DeepSeek V4 Flash 0731 | DeepSeek V4 Pro | |
|---|---|---|
| Released | Jul 31, 2026 (GA) | Jul 2026 |
| Price (off-peak) | $0.14 / $0.28 per MTok | $0.435 / $0.87 per MTok |
| Cache-hit input | $0.0028 / MTok | ~$0.043 / MTok |
| Context | 1M tokens | 1M tokens |
| Terminal-Bench 2.1 | 82.7 | 72.1 (Preview) |
| AA Coding Index | 69.1 | Slightly lower on agents |
| Best at | Cheap fast coding agents | Hardest deep reasoning |
What Changed in 0731
V4 Flash 0731 is a re-post-training of the same 284B-parameter MoE (13B active), not a new architecture. DeepSeek reports better multi-file code generation, stronger repository understanding, more reliable bug fixing, and cleaner tool calling. It scored an Elo of 1559 on GDPval-AA v2 (real-world agentic tasks) and 50 on the Artificial Analysis Intelligence Index v4.1 (#2 of 162 models). Native Codex + Responses API support means it slots straight into modern agent tools.
Cost Per Task (30K in / 5K out)
- V4 Flash 0731: ~$0.0056 off-peak
- V4 Pro: ~$0.017 off-peak
Both double during peak hours (roughly 1–4 & 6–10 UTC). Flash 0731 is close to a 3× cost saving while beating Pro on agent benchmarks — a rare case where cheaper is also better for the target job.
Which Should You Pick?
- Cheap, high-volume coding agents → V4 Flash 0731
- The hardest one-shot reasoning problems → V4 Pro
- Best value overall (Aug 2026) → V4 Flash 0731
Sources
- DeepSeek API pricing: api-docs.deepseek.com/quick_start/pricing
- OpenRouter — DeepSeek V4 Flash 0731: openrouter.ai/deepseek/deepseek-v4-flash-0731