Flash or Pro? A rule that saves time and money.
Two models, one easy heuristic. Pick wrong and you either overpay or under-think the problem. Here is how to choose in five seconds.
Default to Flash. Switch to Pro only when the task has real multi-step reasoning and the cost of a wrong answer is high. Most work is a Flash job; Pro is for the hard 10%.
01V4 Flash — the workhorse
Fast, extremely cheap, and strong enough for the large majority of tasks: drafting, summarizing, everyday coding, classification, and high-volume jobs where price per call matters. When in doubt, start here.
02V4 Pro — for the hard 10%
Slower and pricier, but it reasons more deeply: tricky math and logic, gnarly debugging, multi-constraint planning, and analysis where being right matters more than being fast. If Flash gives a shaky answer on a hard problem, escalate to Pro.
Flash just got a serious upgrade. DeepSeek released V4-Flash-0731, the official version of V4 Flash that supersedes the April preview. Same architecture, re-trained: on DeepSeek’s own published agent and coding benchmarks the new Flash now scores above the V4 Pro preview — Terminal Bench 2.1 jumps to 82.7 (the Pro preview scores 72.1), and it adds a third reasoning-effort level (low / high / max). The open weights are on Hugging Face under an MIT license. The 5-second rule still holds, but the “hard 10%” just got smaller: for agentic and coding work especially, give Flash the first shot before paying up for Pro.
Flash is now V4.1 — and it sees. DeepSeek officially released DeepSeek-V4.1-Flash, the smallest model in a new architecture family built for a higher capability ceiling and faster inference. Two things change the rule above. First, Flash now has native vision: it reads images and charts directly, which V4 Pro still cannot do. Second, on DeepSeek’s own published numbers V4.1 Flash clears the V4 Pro GA build on the agent and coding benchmarks that used to justify escalating (Terminal-Bench 2.1 at 90.6 vs 87.9, GPQA Diamond 90.9) — so the “hard 10%” is now mostly long, multi-constraint text reasoning where you want the bigger model’s depth. On the API the model name is simply deepseek-flash; the old deepseek-v4-flash and deepseek-v4-flash-vision-exp names still work but are routed to V4.1 Flash, and both previous-generation Flash models are retired. V4 Pro is not going away: DeepSeek reversed its September 14 cutoff and will keep serving deepseek-v4-pro at unchanged billing. Flash also got cheaper at launch — see what DeepSeek costs for the new rate card.
03The shared superpower: ~1M tokens
Both models share a very large context window — roughly a million tokens — so the choice is about depth of reasoning, not how much they can read. Either one can ingest an entire report or codebase; Pro just thinks harder about it.