All posts

DeepSeek V4-Flash: the cheapest capable model in 2026 (and when to use it)

· DeepSeek· Pricing· Guide

Not every prompt needs a flagship. A large share of real production traffic (classification, extraction, routing, short drafts) is simple enough that the cheapest competent model wins on cost without giving up quality. In mid-2026 that model is DeepSeek V4-Flash.

What it is

V4-Flash is the lightweight half of the DeepSeek V4 family, released April 2026: 284B total parameters, 13B active. That small active footprint is what makes it fast and cheap. It scores 56.9% on Terminal-Bench 2.0, below V4-Pro's 67.9% but more than enough for the bulk of everyday tasks.

The price is the headline

Per million tokens:

  • Input: ~$0.14
  • Output: ~$0.28
  • Cache hit: ~$0.014

That's on the order of 100× cheaper than closed flagships at $25–30 output. On Turiloop it costs near-zero per token, and every model is discounted right now.

Where V4-Flash shines

  • Classification and routing: sentiment, intent, content moderation, deciding which queue something goes to.
  • Extraction: pulling fields out of text, parsing receipts, structuring messy input.
  • Summarization and short drafts: TL;DRs, email first drafts, changelog notes.
  • High-volume pipelines: anything you run millions of times, where unit cost dominates.

When to step up to V4-Pro

Move to V4-Pro (or a closed frontier model) when the task needs multi-step reasoning or planning, correctness on hard code, or long-horizon agent reliability.

The smart pattern is two tiers: default to V4-Flash, and escalate the hard prompts to V4-Pro. Both sit behind one OpenAI-compatible key, so it's a one-line switch:

def ask(prompt: str, hard: bool = False):
    model = "deepseek-v4-pro" if hard else "deepseek-v4-flash"
    r = client.chat.completions.create(
        model=model,
        messages=[{"role": "user", "content": prompt}],
    )
    return r.choices[0].message.content

A quick cost example

Say you run 100M tokens a month of classification on V4-Flash, split 80M in and 20M out. That's 80 × $0.14 + 20 × $0.28, or about $16.8 a month. The same volume on a $5/$30 frontier model would run $1,000 a month. Add cache hits on a stable system prompt and the Flash number drops further.

V4-Flash is the model you route most of your traffic to and barely notice on the bill. Get a key on Turiloop and try it: international card, pay-as-you-go, no Chinese phone number.