Prompt caching, explained
DeepSeek automatically caches repeated leading prompt prefixes. The cache is prefix-based, not semantic: changes early in a prompt can turn a matching prefix into a cache miss. Keep stable system prompts, few-shot examples, deterministic tool definitions, and ordered message history first; place dynamic user input last. Only the matching input tokens receive the cache-hit rate. At peak rates, Flash cached input costs $0.006 per 1M tokens instead of $0.30 uncached — a 98% reduction. V4 Pro cached input costs $0.044 instead of $1.32 — about 96.67% less. Off-peak, those hit/miss rates are $0.003/$0.15 for Flash and $0.022/$0.66 for V4 Pro. Output is billed separately and receives no prompt-cache discount. Measure the actual hit rate rather than assuming every input token is cached.
Peak vs off-peak pricing
Off-peak input cache-hit, input cache-miss, and output prices are exactly 50% of peak prices. Peak hours are 01:00–04:00 and 06:00–10:00 UTC, Monday through Friday, excluding Chinese public holidays. All other hours are off-peak, including weekends and Chinese public holidays in full. For Flash, uncached input/output falls from $0.30/$1.20 to $0.15/$0.60 per 1M tokens; for V4 Pro it falls from $1.32/$3.96 to $0.66/$1.98. Schedule flexible batch work off-peak and check the official docs for schedule changes. The calculator's toggle selects a rate scenario; it does not detect whether a live request qualifies. Mixed traffic must be budgeted separately for each schedule.
Free tier: the API is not free
The DeepSeek API has no free tier. Usage is billed per token from a topped-up or granted balance. A granted balance is credit toward billed usage, not an ongoing free allowance, and this guide does not promise any signup credit. Consumer chat access is separate from API billing. Check your available balance before starting a job and include retries and long outputs in your budget.
DeepSeek API cost vs GPT-4o
Compare like-for-like text token rates, then measure total cost per solved task. A more expensive call that succeeds immediately may cost less than several cheaper calls plus retries, review, or extra tool steps. Flash has lower uncached input and output prices than GPT-4o, but GPT-4o-mini's $0.15 input and $0.60 output match Flash's off-peak uncached rates and are half Flash's peak uncached rates. Cache reuse changes the comparison again. Do not infer quality, latency, or success rate from a price table; run representative support, coding, and translation tasks and count all billed tokens.
| Model | Input / 1M | Cached input / 1M | Output / 1M |
|---|---|---|---|
| deepseek-flash (peak) | $0.30 | $0.006 | $1.20 |
| deepseek-v4-pro (peak) | $1.32 | $0.044 | $3.96 |
| GPT-4o | $2.50 | $1.25 | $10.00 |
| GPT-4o-mini | $0.15 | — | $0.60 |
GPT-4o-family standard text rates verified October 25, 2026 against OpenAI's official API pricing. Mini cached pricing is not included in this comparison.
Budgeting in practice
The formula is (input tokens × cache-hit share × hit rate + input tokens × cache-miss share × miss rate + output tokens × output rate) ÷ 1,000,000. Multiply by request count for the workload total. On Flash at peak prices, a 2K-token input and 500-token output costs $0.000612 with fully cached input, or $0.0012 with no cached input — $0.612 versus $1.20 per 1,000 requests. At a 50% cache-hit ratio it costs $0.906 per 1,000 requests. Off-peak halves each figure. Monthly examples below assume 30 days, identical requests, no retries, and no third-party hosting fees.
Real-world workload math
- Support bot: 10,000 requests/day, 1,000 input tokens and 300 output tokens each. Flash peak output costs $3.60/day. Fully cached input adds $0.06/day: $3.66/day or $109.80/month. At 50% cache hits, input is $1.53/day: $5.13/day or $153.90/month. No caching yields $6.60/day or $198/month. Off-peak totals are $54.90, $76.95, and $99/month respectively.
- Code-generation agent: 50 files/day, one request per file, 5,000 input and 2,000 output tokens per request. Flash peak costs $0.195/day ($5.85/month) with no caching, or $0.1215/day ($3.645/month) with fully cached input. V4 Pro peak costs $0.726/day ($21.78/month) uncached, or $0.407/day ($12.21/month) fully cached. Off-peak halves these costs. Agent loops, retries, and additional thinking/output tokens increase the bill.
- Batch translation: assume 1M input tokens and 1M output tokens per batch. Flash peak costs $1.206 fully cached, $1.50 uncached, or $1.353 at 50% cache hits. Off-peak costs $0.603, $0.75, or $0.6765 respectively. V4 Pro peak costs $4.004 cached or $5.28 uncached; off-peak costs $2.002 or $2.64. A single batch's cost is not a per-day estimate: multiply by the actual batches processed.
Forecasting and tracking spend
Use actual token counts and cache-hit ratios from your request usage records. Track daily request volume, retries, and long completions; update your forecast after the first week. Separate interactive peak traffic from scheduled off-peak batches. A 20% contingency for retries and cache misses is a budgeting choice, not an extra API fee. Track cost per solved task alongside raw token spend, and set your own spending alerts before a runaway job drains your balance.
Price history
- June 15, 2026: this page's previous update listed V3/R1-era model names and prices. Those figures are historical and are not current billing rates.
- October 25, 2026: fully updated to V4 pricing, including DeepSeek-V4.1-Flash-1017 (deepseek-flash), DeepSeek-V4-Pro-0813 (deepseek-v4-pro), peak/off-peak rates, 1M-token context, and maximum 384K-token output. The retired legacy names deepseek-v4-flash and deepseek-v4-flash-vision-exp remain accepted, but their requests are served by V4.1-Flash at Flash pricing.