Why DeepSeek Raised API Prices 1,100% (2026)

This analysis was written by the aifreetool Editorial Team — a group of full-time AI-industry researchers and writers who verify every claim against primary sources. Last updated August 17, 2026. We keep no affiliate relationship with the companies covered here.

Quick answer: DeepSeek raised API prices for its V4 model family by roughly 50% to more than 1,100% starting August 16, 2026, while introducing peak and off-peak pricing where off-peak costs half of peak. The jump lands alongside the general release of DeepSeek V4-Pro, a 1.6-trillion-parameter mixture-of-experts model. The real story is not the sticker shock — it is that the two-year AI price war is ending, and labs are now charging for agent productivity instead of raw tokens.

For two years, DeepSeek's entire brand was undercutting American rivals on price. That ended this week, when the lab raised API prices by as much as 1,100%. On August 13 the Chinese lab announced new V4 pricing effective August 16, with increases from about 50% to over 1,100% depending on the model, the tier, and the time of day. DeepSeek's API is now split into peak and off-peak windows for the first time, and the headline jump landed the same week V4-Pro went from beta to general availability. The two events are the same story: a cheap model just got good enough to charge for.

Key takeaways:

• V4-Pro output tokens hit $3.96 per million at peak, up roughly 355% from the previous $0.87 flat rate.
• Cache-hit inputs see the steepest increases, up to 1,100%, eroding DeepSeek's biggest structural cost advantage.
• V4-Pro ships as a 1.6-trillion-parameter MoE model with a one-million-token context and three reasoning-effort levels.
• Morgan Stanley frames it as an intelligence war over a price war as Chinese model prices rise across the board.

The Price Shock, by the Numbers

The new schedule is genuinely complicated, which is part of the point. V4-Pro now costs $1.32 per million input tokens (cache miss) and $3.96 per million output tokens at peak, falling to $0.66 and $1.98 off-peak. That is a 51% to 203% increase on inputs and 127% to 355% on outputs versus the old flat rate. The cheaper V4-Flash tier moves from a flat $0.28 per million output to $1.32 at peak, roughly a 371% jump.

The sharpest sting is on cache hits — reused prompts — which climb as much as 1,100%. A detailed breakdown of the new schedule shows DeepSeek's roughly 98% cache-hit discount has been the quiet engine of its cost leadership, well below the industry norm of around 90%. By re-pricing the cache, the company is re-pricing the one place where its advantage was genuinely structural.

What V4-Pro Actually Is

To understand why DeepSeek can charge more, you have to look at what it is now selling. V4-Pro is a 1.6-trillion-parameter mixture-of-experts model with a one-million-token context window and a reasoning-effort ladder — low, high, and max — that lets developers dial thinking up for hard tasks and down for cheap ones. It adds native support for the OpenAI Responses API and ships a one-click setup script for Codex, cutting migration cost to near zero.

On DeepSeek's own benchmark sweep, V4-Pro posts 87.9 on Terminal-Bench 2.1, 83.3 on CyberGym, 62.7 on DeepSWE, and 31.8 on AutomationBench. Those are agentic-coding numbers, not chat numbers. Every one of them measures whether a model can autonomously finish multi-step coding and data work. The positioning is explicit: DeepSeek is no longer selling tokens, it is selling a model that does work, and work is worth more than words.

Peak, Off-Peak, and the Cache Game

The time-of-day structure is the most revealing design choice. Peak hours run from 01:00 to 04:00 and 06:00 to 10:00 UTC, matching DeepSeek's heaviest demand windows. Off-peak — 17 of every 24 hours — stays at half price. The company says the change is meant to allocate resources more reasonably and nudge customers to schedule batch work in quiet periods.

Analysts are already reading the fine print. Sanchit Vir Gogia of Greyhound Research notes that on paper, at peak, DeepSeek's price advantage "disappears, and in places inverts" against OpenAI's GPT-5.6 Luna — but the schedule's own clock and the cache hand most of it back to any buyer paying attention. Mark Tauschek of Info-Tech Research Group adds that V4-Pro still beats OpenAI's mid-tier Terra on price even at peak. The conclusion for developers: treat the schedule as a routing variable, not a cost line. Batch jobs, data labeling, and overnight code review belong in the off-peak window; interactive products that must answer in real time have far less room to move.

The End of the Price War

DeepSeek did not raise prices in a vacuum. A Morgan Stanley report dated August 9 — titled "Intelligence War Over Price War" — found that average Chinese LLM API input prices rose to 4.9 yuan per million tokens in Q2 2026 and output prices to 21.9 yuan, up about 48% and 80% respectively from the first quarter of 2025. ByteDance, Alibaba, Baidu, Tencent, MiniMax, Zhipu, Moonshot, and DeepSeek all moved the same direction. OpenAI cut Luna's price 80% for off-peak use in the same window, and Anthropic raised prices back in April.

The underlying driver is supply and demand, not margin expansion. When demand for agentic workloads surges, compute gets scarce, and even the cheapest provider can no longer absorb the cost. DeepSeek is reportedly seeking a valuation near $74 billion ahead of a possible mainland listing — and a company preparing to sell shares has to prove it can charge for what it builds. You can browse the coding-agent landscape driving this demand in our AI coding tools directory.

My Take / The Bottom Line

The 1,100% headline is real but slightly misleading. Most developers will not pay the full peak rate: 17 of 24 hours stay at half price, and cache hits still beat the market. What actually changed is the signal. The lab that built its reputation on giving tokens away just told the market it can charge for agent productivity — and that it is willing to segment its own demand to do it.

For teams that leaned on DeepSeek as a cheap way to run big workloads, the playbook changes: profile your spend by hour, move batch jobs off-peak, and re-run the comparison against GPT-5.6 Luna and Anthropic's tiered pricing before assuming DeepSeek is still the cheapest option. The era of Chinese models as a loss-leading undercut is over. It is an intelligence war now, and prices are only going up from here.

FAQ

How much did DeepSeek raise its API prices? Increases range from roughly 50% to more than 1,100% depending on the model, tier, and time of day. V4-Pro output tokens hit $3.96 per million at peak, and cache-hit inputs see the steepest jumps.

When does the new pricing take effect? The new peak/off-peak schedule took effect August 16, 2026, with off-peak hours priced at half the peak rate.

What are DeepSeek's peak hours? Peak windows run 01:00 to 04:00 and 06:00 to 10:00 UTC. Off-peak covers 17 of every 24 hours.

Is DeepSeek still cheaper than OpenAI? At peak, V4-Flash loses its edge over GPT-5.6 Luna, but V4-Pro still beats OpenAI's mid-tier Terra on price even at peak. Off-peak, DeepSeek keeps a clear cost advantage.

What is DeepSeek V4-Pro? A 1.6-trillion-parameter mixture-of-experts model with a one-million-token context and three reasoning-effort levels, released to general availability alongside the price change.