DeepSeek V4 Peak/Off-Peak Pricing: The End of Flat-Rate LLM APIs
DeepSeek introduces time-of-day billing
On August 16, 2026 at 16:00 UTC (19:00 EAT), DeepSeek switches its V4 API from flat-rate pricing to peak / off-peak billing — a first for a major Chinese model provider at this scale. Off-peak rates are exactly half of peak rates, and the peak windows are narrow: 01:00–04:00 and 06:00–10:00 UTC, which translates to 04:00–07:00 and 09:00–13:00 EAT. Everything else is off-peak, which for Kenyan users is roughly 75% of the day.
The headline: even the new off-peak rates are more expensive than today’s prices, and peak rates are 2–4× current levels depending on token type. The era of flat, ever-declining API prices is over for DeepSeek.
What actually changes
| Model | Period | Input (cache hit) | Input (cache miss) | Output |
|---|---|---|---|---|
| v4-flash | Today | $0.0028 | $0.14 | $0.28 |
| Off-peak | $0.007 (↑150%) | $0.22 (↑57%) | $0.66 (↑136%) | |
| Peak | $0.014 (↑400%) | $0.44 (↑214%) | $1.32 (↑371%) | |
| v4-pro | Today | $0.003625 | $0.435 | $0.87 |
| Off-peak | $0.022 (↑507%) | $0.66 (↑52%) | $1.98 (↑128%) | |
| Peak | $0.044 (↑1114%) | $1.32 (↑203%) | $3.96 (↑355%) |
All prices per 1M tokens. Source: DeepSeek pricing docs.
Two patterns stand out. First, cache-hit input is the biggest jump: 150–507% off-peak, up to 1,114% at peak for v4-pro. DeepSeek is pricing the convenience of hot KV-cache reads much closer to their real serving cost. Second, output tokens are the real money: at peak, a v4-pro output token costs $3.96/1M, over 4.5× today’s price. Anyone running agent loops or long generation pipelines just saw their bill’s dominant line item inflate.
Why now: inference is a capacity business
This is textbook demand-shaping, straight out of the electricity market playbook. When a model is as popular as V4, the marginal cost of serving a token at 3 a.m. is far lower than at 9 a.m. V4 released April 24, 2026 with 87.5 MMLU-Pro, 80.6% SWE-bench Verified and 93.5 LiveCodeBench for v4-pro (official numbers). Peak/off-peak pricing nudges batch jobs and cron workloads into the cheap hours, smoothing the load curve without building idle capacity.
DeepSeek’s own efficiency claims make the economics visible: the V4 technical report says v4-pro needs only 27% of the single-token inference FLOPs and 10% of the KV cache of V3.2 in the 1M-token setting. That efficiency is why V4 could undercut Western labs for months. But capacity is still finite, and time-of-day pricing is the lever that keeps utilization high without rationing.
DeepSeek is not alone — the whole market is repricing
The flat-rate-to-dynamic shift is happening while frontier labs are raising prices across the board:
- Anthropic: Claude Sonnet 5 launched at intro pricing of $2/$10 per 1M input/output tokens. That intro pricing ends August 31, 2026, with an increase effective September 1.
- OpenAI: GPT-5.5 launched at roughly 2× GPT-5.4’s per-token price (GPT-5.4 was $2.50/$15), with the flagship tier now at $30/1M input. A 10% regional-processing uplift applies to models released on or after March 5, 2026.
- Google is the counter-trend: Gemini 3.7 and 3.6 Flash are offered at $0.75/$3.75 introductory pricing through December 31, 2026, a deliberate price cut against the hiking pack.
The market is bifurcating: frontier labs raise prices while challengers (and open weights) compress them. DeepSeek’s move is best read as capacity management rather than a pure hike — they keep the absolute floor low (cache-hit at $0.007 off-peak is still cheaper than any Western frontier model) while making the premium tiers behave like a real utility.
What this means for how we use LLMs
1. Time-shift everything that can wait. Anything without a hard latency requirement should run in off-peak hours: batch inference, nightly indexing, evals, report generation, training-data synthesis. For Kenya, that means scheduling heavy workloads after 13:00 EAT (and avoiding the 09:00–13:00 morning peak entirely).
2. Cache hits are now the single biggest lever. A cache-hit input token at off-peak is ~30× cheaper than a cache-miss ($0.007 vs $0.22 for v4-flash). Designing prompts with stable system prefixes, shared context blocks, and conversation-level caching has stopped being an optimization. It’s now the difference between a $10/month bill and a $300/month bill on high-volume workloads.
3. Judge models by price-per-benchmark-point, not sticker price. The old DeepSeek story was “near-frontier quality at 1/10th the price.” The new story is “near-frontier quality at 1/10th the price if you schedule smart.” At off-peak, v4-flash output at $0.66/1M with 79.0% SWE-bench Verified is still a phenomenal deal vs GPT-5.5 Pro at $30/1M input. At peak, the gap narrows fast.
4. Agents change the calculus. Agent loops are output-heavy and latency-sensitive — the two things peak pricing punishes hardest. Expect agent platform costs to be the first thing renegotiated, and expect providers to start offering agent-specific off-peak tiers.
Practical takeaway for Kenyan teams
With EAT peak hours at 04:00–07:00 and 09:00–13:00, the cheapest compute day looks like: schedule cron jobs and batch pipelines after 13:00, keep interactive work on cache-heavy prompts, and push any 24/7 agent workloads to the 13:00–04:00 window where possible. This is the same demand-response thinking that keeps Kenyan electricity tariffs lower at night — applied to tokens.
Bottom line: DeepSeek’s peak/off-peak model is the first visible sign that LLM pricing is maturing from “race to the bottom” to “capacity market.” The cheap era isn’t over. It has just moved to specific hours, and the teams that schedule around it will pay a fraction of what the peak-hour crowd pays.
References
- DeepSeek API pricing docs — official peak/off-peak announcement
- Macaron: DeepSeek V4 Benchmarks — official V4 benchmark summary
- Claude pricing docs — Sonnet 5 intro pricing ending Aug 31
- OpenAI API pricing — GPT-5.5 tiers and regional uplift
- Google Cloud Agent Platform pricing — Gemini 3.7/3.6 Flash intro pricing