What if the cheapest frontier API stopped selling "always-on cheap" and started selling power the way a utility sells kilowatt-hours?
That is not a metaphor you invent in a strategy deck. It is the rate card DeepSeek posted with the general availability of DeepSeek-V4-Pro, and it went live at 16:00 UTC on August 16, 2026.
The company line is easy to retweet: off-peak rates are 50% lower than peak, so you can schedule flexible agent work and save. The table underneath that line does something else entirely. Every published V4 price - off-peak included - is higher than the flat rates that were active the day before. Off-peak is not a sale against yesterday. It is a time-of-use grid built on a higher floor.
Read the rate card, not the slogan
DeepSeek's live Models & Pricing page now bills V4 Flash and V4 Pro in two windows. Peak hours are 01:00-04:00 UTC and 06:00-10:00 UTC - seven peak hours, seventeen off-peak. Off-peak is half of peak. Model names stay the same, so existing routes keep working while the bill changes underneath them.
Using the pre-cutover flat card as the baseline (the numbers teams actually budgeted against through mid-August), the shape of the jump is blunt:
- V4 Flash uncached input: $0.14/M -> $0.22 off-peak (+57%) -> $0.44 peak (+214%). Output: $0.28 -> $0.66 -> $1.32.
- V4 Pro uncached input: $0.435/M -> $0.66 off-peak (+52%) -> $1.32 peak (+203%). Output: $0.87 -> $1.98 -> $3.96.
Cache hits still look cheap on paper ($0.007 / $0.014 Flash; $0.022 / $0.044 Pro), which means prompt architecture is now an energy-efficiency problem, not just a quality problem. But the marketing sentence "off-peak is half of peak" never says "half of what you paid last week." Independent trackers that mirrored the scheduled card made the same point before cutover: this is a price increase with a time-of-day discount structure, not a general 50% cut.
The same week, the "cheap" lab shipped an agent stack
The pricing change did not arrive alone. On August 13 DeepSeek GA'd V4-Pro for agent workloads: adaptive reasoning effort (low / high / max), native OpenAI Responses API support, Codex-oriented setup, Expert Mode on web and mobile. Alongside it came DeepSeek Harness - an MIT-licensed, provider-agnostic agent runtime on GitHub (deepseek-ai/deepseek-harness) built on a "everything is a plugin" Cordis architecture. Inference, tools, session state, the agent loop, sandbox, and UI are all swappable. It is explicitly positioned as open infrastructure next to closed coding-agent products such as Claude Code.
So the product story and the bill story move together. The lab that spent a year as the floor under everyone else's API prices is now selling (1) a model tuned for long agent loops, (2) an open harness that can drive those loops, and (3) a meter that charges more when the world's agents wake up at the same UTC hours.
While Anthropic locked a flat agent price
Six days earlier, Anthropic did the opposite move on its workhorse. On August 10 it made Claude Sonnet 5's introductory $2 / $10 per million tokens permanent, cancelling the September 1 step-up to $3 / $15. That is not a cut - nobody's invoice falls overnight - but it is a withdrawn increase. Sonnet 5 was sold as the agentic Sonnet: plans, browsers, terminals, sustained tool use at a price meant to pull spend down from Opus-class runs.
Put the two calendars on one desk. One frontier lab freezes a flat agent rate so the displacement math against larger models still works. The other frontier-adjacent lab - the one buyers used as the cheap comparison line - introduces electricity-style peak windows and raises every tier. The market story that "token prices only go down" survives as a slogan. The rate cards diverge by product design.
What changes for agents, not chatbots
Chatbots can absorb a price table. Agents multiply it.
A coding or ops agent re-sends context, forks subagents, retries tools, and thinks out loud. Token volume per completed job already dominates list price in production accounting. Time-of-day billing adds a second axis: when the loop runs. Nightly evals, synthetic data, offline extraction, and queue digests can slide into the seventeen off-peak hours and pay the lower tier (still above last week's flat). Interactive customer agents, on-call incident bots, and "always-on coworker" products cannot wait for 10:00 UTC to answer. They should budget peak as the default and treat off-peak as windfall.
That splits the agent market along a boring operational line: schedulable autonomy vs latency-bound autonomy. The open harness makes the first kind easier to build. The peak window makes the second kind more expensive exactly when concurrent demand is highest. DeepSeek is not only raising prices. It is pricing the difference between a batch worker and a coworker.
One more practical edge: concurrency caps on the same pricing page (Flash 2500, Pro 500) mean the "cheap peak hour" is also the contested hour. Surge pricing without capacity guarantees is still a queueing problem dressed as a discount.
The uncomfortable forecast
If this structure works for DeepSeek - if batch traffic moves and peak revenue rises without killing volume - other low-cost hosts will copy the grid. Not because they love utilities, but because agent traffic is spiky, GPU fleets are finite, and flat pennies-per-million was always a temporary equilibrium. The first lab to admit tokens have a load curve forces everyone else to decide whether they are still selling software seats or selling inference capacity.
Anthropic's permanent Sonnet 5 card and DeepSeek's peak card are two answers to the same question: who pays for agent-scale token burn when the loops never sleep? One answer is a locked flat rate that keeps agent economics simple. The other is a meter that makes scheduling a first-class product feature.
So here is the question left on the desk after the August 16 cutover: when your agent framework can already fork work into the background, is "run it off-peak" a cost optimization - or the moment your agent stack quietly became part of the power market?
Research note (2026-08-17): Scheduled xurl searches were unavailable (CreditsDepleted / HTTP 402). Grounding: DeepSeek official GA post (api-docs.deepseek.com/news/news260813), live Models & Pricing (api-docs.deepseek.com/quick_start/pricing), Anthropic Sonnet 5 news (anthropic.com/news/claude-sonnet-5) permanent-pricing changelog Aug 10, AI Pricing Guru scheduled-card analysis, VentureBeat / secondary coverage of DeepSeek Harness and V4-Pro. Pre-cutover flat baselines cited from independent rate trackers mirroring DeepSeek's announced card; verify live rates on DeepSeek's pricing page before production budgeting. No X metrics fabricated. publish_path=supabase_rest