V4-Flash vs. V4-Pro: DeepSeek promised better and cheaper. It’s true, but not how I expected.

Sicherheitsmeldung mit Schweregrad noch nicht bewertet. Technische Details im Tab „Originaltext“; empfohlene Schritte in der Checkliste.

Originaltext der Quelle (nicht übersetzt) – die deutsche Zusammenfassung findest du im Tab „Kurzfassung".

DeepSeek, the Chinese AI lab that’s been undercutting OpenAI and Anthropic on price since it arrived, recently refreshed V4-Flash, moving it out of preview. DeepSeek documentation promised that this refreshed Flash model topped the V4-Pro preview on coding and agentic benchmarks.

The prices make that claim strange, though. V4-Flash runs $0.14 per million input tokens and $0.28 per million output. V4-Pro runs $0.435 and $0.87. That is almost exactly one-third the price on both input and output.

If the budget model beats the flagship at a third of the price, why does the flagship exist? I had to get to the bottom of this, so I gave both models the same three professional coding jobs and tracked every token.

The tests

I used rich, the Python terminal-formatting library from Textualize. I chose it because it’s real production Python with a massive install base and enough internal complexity (style spans, render pipelines, cell-width math) to make a model work for it.

Both models ran inside OpenCode, the open source terminal coding agent — like Claude Code except you can plug in any model. One setup note: OpenCode’s built-in DeepSeek list only ships with Flash, and there’s a closed-as-not-planned GitHub issue about it. To get Pro into the model picker, I had to add a small config file to the project folder (this guide has the JSON). It only took about two minutes, mentioning just in case you try to test pro on OpenCode and can’t find the model.

I built a separate clone of the repo for each model, started a fresh session for every test, and used identical prompts, so the …

Vollständiger Originalartikel ↗

• Betroffenheit im eigenen Stack prüfen: Versionen/Komponenten abgleichen.
• Originalquelle / Hersteller-Advisory lesen: https://thenewstack.io/deepseek-flash-pro-benchmark/
• Verfügbaren Patch oder Workaround einspielen und dokumentieren.