DeepSeek V4 Flash Pricing: Token Costs & How to Save (Aug 2026)

DeepSeek moved every V4 model to peak/off-peak billing at 16:00 UTC on August 16, 2026.^[1] For DeepSeek V4 Flash, the off-peak price is $0.22 per 1M input (cache miss), $0.007 per 1M input (cache hit), and $0.66 per 1M output. Peak hours run 01:00-04:00 and 06:00-10:00 UTC and bill at double: $0.44 / $1.32 (and $0.014 cache hit).

The pre-Aug-16 price was a single flat rate: $0.14 input (cache miss), $0.0028 input (cache hit), $0.28 output. Off-peak is already above that flat rate — $0.22 versus $0.14 on input and $0.66 versus $0.28 on output — so the effective cost rose even in the cheap hours, with peak hours roughly twice off-peak again. Third-party trackers (Engadget, aiPricing.guru, chat-deep.ai) all confirm the new schedule.

Billing stays simple: expense = number of tokens x price, no monthly fee, no per-seat charge. One API key funds both V4 Flash and V4 Pro; only the model name and price change.

• 1M tokens is roughly 750,000 English words — a useful yardstick for estimating a workload's bill.
• The cache hit / cache miss split applies to input only; output has no cache tier.
• The old aliases deepseek-chat and deepseek-reasoner were permanently retired on 2026-07-24;^[3] new integrations call deepseek-v4-flash directly.

curl https://api.deepseek.com/chat/completions \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer ${DEEPSEEK_API_KEY}" \
  -d '{
        "model": "deepseek-v4-flash",
        "messages": [
          {"role": "system", "content": "You are a helpful assistant."},
          {"role": "user", "content": "Summarize the attached changelog."}
        ],
        "thinking": {"type": "enabled"},
        "reasoning_effort": "high",
        "stream": false
      }'

DeepSeek split every V4 billing line into peak and off-peak. Off-peak is 50% below the new peak rate — $0.22 versus $0.44 on Flash input — but both tiers sit above the old flat $0.14/$0.28.

Peak windows are daily 01:00-04:00 and 06:00-10:00 UTC (they replace the earlier Beijing-time proposal). All other hours are off-peak. Seventeen of every 24 hours land in off-peak, so work that can wait should move off-peak.

The 2x peak multiplier applies to every line — cache-hit input kicks up to $0.014 on Flash, cache-miss input to $0.44, output to $1.32. Scheduling non-realtime jobs (cron, retries, batch, synthetic data) into off-peak is now the single biggest structural lever on the bill.

• Peak windows: 01:00-04:00 and 06:00-10:00 UTC daily.
• Off-peak = 50% of the new peak rate (not 50% of the old flat rate).
• Move batch and retry workloads into off-peak hours to dodge the 2x multiplier.

DeepSeek's context caching is automatic. A disk-based prefix cache stores repeated prompt prefixes, and any token served from it bills at the cache-hit price instead of the cache-miss price — still roughly 97% cheaper on input, as it was before the rate change.

At off-peak, the cache-hit rate is $0.007 versus $0.22 on a miss — about a 97% cut. The rule for maximizing hits is unchanged: put the reusable prefix first. System prompts and context documents sit at the top, volatile content at the bottom.

A worked example at off-peak: 1M input on a full cache miss plus 1M output costs $0.22 + $0.66 = $0.88. With a 90% cache-hit rate, the same call lands around $0.22 x 0.1 + $0.007 x 0.9 + $0.66, or roughly $0.69.

• Cache is best-effort — DeepSeek does not guarantee a hit rate.
• Monitor usage.prompt_cache_hit_tokens and prompt_cache_miss_tokens to see your real hit rate.
• There is no Batch API discount tier for DeepSeek, so caching and off-peak scheduling are the main structural levers.

DeepSeek prices thinking and non-thinking modes identically — no surcharge per reasoning token. But reasoning generates extra tokens billed at the output price ($0.66 off-peak per 1M on Flash), so a long agentic run's spend is real.

Thinking is on by default with effort set to high; complex agent requests such as Claude Code or OpenCode auto-use max. Toggle it off per request via "thinking": {"type": "enabled"} to drop the reasoning-token stream entirely.

Reasoning also counts toward output length, which matters against the 384K max-output cap.

• Reasoning tokens count in output_tokens, billed at the output rate ($0.66 off-peak) on top of the final answer.
• Same per-token price for all effort levels — high reasoning carries no multiplier.
• Off-peak scheduling and max_tokens caps tame the worst of the reasoning spend.

Flash runs about a third of Pro per token at off-peak: $0.22 versus $0.66 on input, $0.66 versus $1.98 on output. Since both read the same 1M context, the gap is pure margin for workloads that only need Flash-level quality.

The common hybrid pattern — route everything to Flash by default, escalate to Pro on failure or low confidence — is roughly 3x cheaper end-to-end than running Pro everywhere. Flash also carries a 2500-request concurrency limit versus Pro's 500.

A baseline for full-context off-peak calls: 1M input cache miss + 384K max output ≈ $0.22 + $0.253, or about $0.47 per call at the ceiling (derived from official prices).

Even after the Aug 16 increase, Flash remains far below the frontier price floor. At off-peak ($0.22/$0.66) it still undercuts GPT-5.5 ($5/$30) and Claude Opus 4.8 ($5/$25) by roughly 20x on input and 45x on output.

The gap narrows at peak ($0.44/$1.32) but stays wide — still roughly 11x input and 22x output under GPT-5.5. Cache-hit input keeps the discount mechanism no US model matches at this level.

One third-party comparison ran the same coding-agent workload at about $12,000/month on GPT-5.5 versus roughly a third on V4 Flash at off-peak rates.

Flash still undercuts every major frontier model, and stacking a few habits pushes the effective rate lower. The biggest savings come from layering levers: schedule off-peak, route Flash, engineer cache-friendly prefixes, cap output, and shop third parties.

The scale math once at off-peak: 10M input / 5M output per day with an 80% cache-hit rate gives input about 10 x (0.2 x $0.22 + 0.8 x $0.007) = $0.50/day and output $3.30/day — roughly $3.80/day, about $114/month if run entirely off-peak, more if you spill into peak.

Remember caching is automatic and best-effort: put stable content first and the system does the rest.

• Schedule non-realtime work in off-peak hours (avoid 01:00-04:00 and 06:00-10:00 UTC) — the single biggest lever after the rate change.
• Route Flash by default; escalate to V4 Pro only on failure or low confidence. Hybrid routing keeps mixed workloads about 3x cheaper end-to-end.
• Put reusable prefixes first: at 90% cache hit, effective off-peak input drops from $0.22 to about $0.028 (-87%).
• Cap output with max_tokens — Flash is verbose; Artificial Analysis consumed 210M output tokens on its eval suite.^[5]
• Check third-party providers: OpenRouter and DeepInfra often list Flash below official rates.

from openai import OpenAI
import os

client = OpenAI(
    api_key=os.environ.get("DEEPSEEK_API_KEY"),
    base_url="https://api.deepseek.com",
)

response = client.chat.completions.create(
    model="deepseek-v4-flash",
    messages=[
        {"role": "system", "content": "You are a helpful assistant"},
        {"role": "user", "content": "Write a concise summary"},
    ],
    max_tokens=1024,  # cap output spend
    reasoning_effort="high",
    extra_body={"thinking": {"type": "enabled"}},
)
print(response.choices[0].message.content)