DeepSeek's Peak/Off-Peak Pricing Is Live: Cut Your Coding Bill in Half by the Clock

As of 16:00 UTC on August 16, 2026, DeepSeek no longer has one price. It has two — and the gap between them is exactly 2x. If you schedule your agent runs, you can cut your DeepSeek bill in half without changing a single line of code.

DeepSeek quietly moved its flagship coding models, V4 Flash and V4 Pro, from flat per-token pricing to a peak/off-peak model. The mechanic is simple: there are defined "peak" hours when demand is highest, and every token — input and output — costs double what it does the rest of the day. Off-peak is the base rate; peak is 2x that base.

The New Rate Card

All figures are per million tokens, cache-miss. Cache hits are cheaper still: V4 Flash off-peak cache-hit input is just $0.007/M, and V4 Pro off-peak cache-hit input is $0.022/M. The peak window is 01:00–04:00 and 06:00–10:00 UTC. Every hour outside those two blocks is off-peak.

This Is a Price Increase Dressed as a Schedule

Be honest about what happened here. Before this change, V4 Flash was $0.14/$0.28 and V4 Pro was roughly $0.435/$0.87. The new off-peak rates are higher than the old flat rates — off-peak Flash output ($0.66) is more than double the old $0.28. So even if you run every job off-peak, you are paying more than you did in July. Peak is simply the ceiling, not a penalty relative to the old world.

That reframes the optimization. You are not "avoiding a surcharge" — you are avoiding the top half of a rate card that shifted up on both ends. DeepSeek is still one of the cheapest frontier-adjacent coding models on the market, but the era of a single $0.14 headline number is over.

What Scheduling Actually Saves

Consider a team running batch code review and test generation on V4 Pro. Say each run consumes 8M input tokens and 2M output tokens per day.

Run entirely at peak: 8M × $1.32 + 2M × $3.96 = $10.56 + $7.92 = $18.48/day.

Run entirely off-peak: 8M × $0.66 + 2M × $1.98 = $5.28 + $3.96 = $9.24/day.

That's a flat 50% difference — $9.24/day, or about $2,770/year on this one workload — determined purely by when the job runs. For anything that isn't latency-sensitive (nightly CI, doc generation, bulk refactors, embeddings backfills), there is no reason to touch a peak hour.

Which Work Is Time-Flexible?

• Move off-peak: scheduled code review, test suite generation, documentation passes, large-scale refactors, migration scripts, anything queued rather than typed.
• Accept peak: interactive pair-programming, live debugging, anything where a developer is waiting on the response. The 2x premium on a $0.30 task is $0.30 — not worth blocking a human over.

The peak window sits in the European working morning and the tail of the US night. If your team is US-based, a surprising amount of your interactive work already lands off-peak. Check your timezone against 01:00–04:00 and 06:00–10:00 UTC before assuming you need to change anything.

The Broader Signal

Time-of-day pricing is common in electricity and cloud spot markets but new to frontier LLM APIs. DeepSeek is effectively exposing its own capacity constraints to customers and asking them to smooth demand. If it works — if enough batch load shifts to off-peak — expect other providers under GPU pressure to follow. Grok already doubles rates past a context threshold; DeepSeek now doubles them by the clock. The single-number price tag is fragmenting into conditional rates, and cost optimization increasingly means reading the fine print, not just the headline.

Want to see how DeepSeek's off-peak rates stack up against Claude, GPT, and Gemini for your project size? Run the numbers in our AI Cost Calculator.