DeepSeek V4 API Price Hike (Aug 2026): Peak/Off-Peak Pricing & What It Costs BYOK Coding Agents

TL;DR

On August 17, 2026 (00:00 Beijing time), DeepSeek's new V4 API pricing took effect. V4-Pro moved from a flat rate to peak/off-peak pricing, and the headline increases are steep: peak-hour output is ¥27/M tokens (+350%), uncached input ¥9/M (+200%), and cached input ¥0.30/M (+1100%). Off-peak hours cost exactly half.

V4-Flash also rose (peak output ¥9/M), the end of DeepSeek's "extreme value" era. One developer told Caixin his monthly bill would jump from about ¥3,000 to nearly ¥10,000 for the same workload.

If you run a Cline, Aider, or Codex CLI workflow with DeepSeek as your model backend, this is the most important pricing change of the quarter. The good news: off-peak scheduling, cache discipline, and model substitution can claw back most of the increase.

What changed

DeepSeek announced the adjustment on August 13 (the same day V4-Pro went fully GA, ending its test period) and the new prices went live four days later. This is DeepSeek's fourth pricing adjustment of 2026, and by far the largest single jump.

Three structural changes, not just a price bump:

• Peak/off-peak tiers. Prices now differ by time of day. Peak hours are Beijing time 9:00–12:00 and 14:00–18:00; everything else (nights, early mornings, lunch break) is off-peak at half the peak rate.
• V4-Pro went GA. The flagship model is no longer in test pricing — its commercial rate is the new baseline, and it now emphasizes agentic workloads.
• The cache discount widened. Cached input stays cheap (¥0.30/M peak on Pro), which makes prompt-caching strategy matter more than ever.

New V4 price table

All prices in RMB per million tokens, effective August 17, 2026:

The previous V4-Pro rate was ¥3 input (uncached) / ¥6 output / ¥0.025 cached input. Under the new peak schedule, cached-input is up 12x (still cheap in absolute terms), uncached input 3x, and output 4.5x. Off-peak pricing roughly returns output to ¥13.50 — still more than double the old flat rate, but much friendlier than peak.

Peak vs off-peak: the schedule

Peak hours are defined in Beijing time: 9:00–12:00 and 14:00–18:00. Everything outside those windows — including the 12:00–14:00 lunch window, evenings after 18:00, and all of the night — is off-peak at 50%.

This is a market mechanism, not a discount gimmick. DeepSeek's stated goal is to smooth out daytime compute congestion: over the past year the platform suffered repeated outages (May 8, 21, 24, and 28 all crashed as user volume grew 66.7% while compute reserves grew only 8.3%). Time-based pricing pushes batch workloads and agent pipelines to off-peak hours, flattening demand.

For coding agents, the practical implication: a V4-Pro agent run scheduled after 18:00 Beijing time costs half of the same run at 10:00. For China-based teams that's a huge lever. For teams in other time zones, the "peak window" maps to a different part of your day — check what the Beijing hours correspond to locally.

Why now

DeepSeek is not alone. The August 2026 round is the visible peak of a broader repricing wave across Chinese model vendors, driven by compute costs:

• Zhipu (GLM): three API price hikes in 2026 — GLM Coding Plan restructure in February, GLM-5-Turbo +20% in March, GLM-5.1 +10% in April. CEO Zhang Peng says Q1 pricing rose 83% year-over-year while demand still grew 400%.
• MiniMax: doubled API prices with the M3 launch on June 1.
• Moonshot (Kimi): K3 (July) raised input to ¥20/M and output to ¥100/M — over 3x the previous generation.
• Cloud giants: Tencent Cloud, Alibaba Cloud, and Baidu AI Cloud have all raised compute/API prices this year.

The 2025–2026 price war — where Chinese models sold tokens at or below marginal cost to grab share — is over. Analysts describe this as a correction of long-subsidized inference pricing rather than a one-off. Expect further adjustments, not reversals.

Impact on BYOK coding agents

DeepSeek's V4 models are the default cheap backend for most bring-your-own-key workflows — Cline, Aider, Codex CLI with an API key, and open-source harnesses like the DeepSeek harness (dsh) itself. That audience is hit hardest:

• Always-on agents (24-hour monitors, CI-integrated review bots) that previously ran at ¥6/M output on Pro now pay up to ¥27/M during peak windows — a 4.5x swing before off-peak optimization.
• Translation, summarization, and refactoring pipelines with low cache-hit rates see the full uncached-input increase (¥3 → ¥9 peak).
• Heavy parallel runs (multiple sub-agents) are disproportionately exposed because they rarely respect time-of-day scheduling by default.

Caixin's example is representative: a developer whose software burns tokens continuously estimated monthly API spend rising from ~¥3,000 to ~¥10,000 — 3–4x — for the same workload, and is now considering passing the cost to end users.

The competitive context matters. On a pure per-token basis, DeepSeek is no longer the automatic cheapest choice for every workload: Muse Code's Contributor tier ($0.20/M output) undercuts it for open-source work, and GPT-5.6 Luna ($1.20/M output after the July 30 cut) is competitive at scale. The calculus shifted from "DeepSeek is cheap, done" to "match model to task."

How to cut your bill

Four levers, roughly in order of impact:

1. Schedule agent runs off-peak. Move batch jobs, migrations, and test-generation pipelines to after 18:00 Beijing time (or the lunch window 12:00–14:00). A cron wrapper around cline or aider is a ten-minute change that halves output cost.
2. Raise cache-hit rates. Cached input is ¥0.30/M on Pro — 30x cheaper than uncached. Keep system prompts and project context stable across runs, use the same session prefix, and structure prompts so the first N tokens repeat. Agents that re-send identical context on every turn benefit the most.
3. Drop to V4-Flash for routine work. Flash at peak (¥9 output) is cheaper than Pro at off-peak (¥13.50). Route simple edits, autocomplete, and boilerplate to Flash; reserve Pro for architecture and multi-file refactors.
4. Evaluate substitutes per task. For token-heavy batch work on non-proprietary code, compare Muse Code Contributor ($0.10/$0.20 per M) and GPT-5.6 Luna ($0.20/$1.20 per M, API key). For frontier reasoning, Kimi K3 and Qwen3.8-Max are the domestic alternatives worth benchmarking on your actual workload — not on benchmarks.

Bottom line

DeepSeek's August 17 repricing ends its extreme-value positioning and replaces a flat price with a time-sensitive one. The absolute numbers are still competitive — off-peak Pro output at ¥13.50/M remains well below most frontier Western models — but the old assumption "DeepSeek is the cheap default, no thought required" no longer holds.

For BYOK users, this is a workflow-design problem, not a crisis: off-peak scheduling plus cache discipline restores most of the cost advantage. For teams running 24/7 agents, budget for the peak-rate worst case and treat off-peak as the saving, not the baseline. And if you're shopping for a new default model backend, the decision is now genuinely task-dependent — which is exactly how the market is supposed to work.

How we wrote this

Compiled August 21, 2026, four days after the new pricing went live. Sources cross-referenced:

• DeepSeek's official announcement (August 13) and the pricing change effective August 17.
• Coverage from Nanfang News, China Economic Net, and Caixin via NetEase.
• 21st Century Business Herald analysis of the industry-wide repricing wave (Zhipu / MiniMax / Kimi / cloud vendors).
• Our August 2026 model pricing tracker for cross-vendor comparison.

We have not independently verified DeepSeek's operational cost claims. All price figures are from official announcements and major outlets; verify against DeepSeek's pricing docs before budgeting, as rates may shift again. Corrections welcome via our about page.