TOGETHER WITH TODAY'S PARTNER
Good morning, {{first_name|there}}. The cheapest frontier model on earth introduces surge pricing tomorrow — and almost nobody has repriced their stack for it.
Read time: 3 minutes. Same time, every weekday — rate today's issue at the bottom.
🚀 The Big Story: DeepSeek ends the price war it started
Starting tomorrow, August 16, DeepSeek splits its API into peak and off-peak billing — and peak output on V4-Pro jumps from $0.87 to $3.96 per million tokens. That's 4.6x, on the model that made $0.14 inputs normal.
The new rates: V4-Pro cache-miss input goes $0.435 to $1.32 at peak; V4-Flash output goes $0.28 to $1.32. Off-peak is exactly half of peak, so nothing goes back to yesterday's price.
The clock that matters: peak is 01:00-04:00 and 06:00-10:00 UTC — roughly 9pm-midnight and 2am-6am US Eastern. If your traffic is US business hours, you land in off-peak and eat about 2.3x, not 4.6x.
The reason given: DeepSeek says it's adjusting "to allocate resources more reasonably." Translation: inference demand outran the subsidy. It warned developers on August 6 and confirmed rates on the 13th.
Jason's take: The floor was never a technology curve — it was a subsidy, and subsidies end. Anyone who priced a product off DeepSeek's rock-bottom tokens just found out their gross margin was rented, not owned. The lesson isn't "leave DeepSeek," it's that single-provider inference is now a business risk, not an architecture preference.
⚡ Quick Hits
Gemini 3.7 Flash lands cheap and good. Google's Wednesday drop jumped DeepSWE from 49% to 65.3% and ships at $0.75 in / $3.75 out per million — with 50% off through year-end.
OpenAI's Sol hits 750 tokens per second. The Cerebras-powered GPT-5.6 Sol runs up to 14x faster than standard, which quietly makes real-time agent loops viable.
Anthropic cut Opus 5 to roughly half of Fable 5. The frontier labs are discounting the same week the discount house raises prices — read that twice.
Anthropic signed a $9.1B, 20-year compute lease. Riot Platforms' 191MW Rockdale, Texas campus delivers 96MW by December 2027 — capacity is being booked out to 2048.
Apple built a China-only model with Alibaba. It has registered on-device generative AI with Chinese regulators to fight Huawei on home turf.
Encrypted reasoning blocks turned out to be portable. Researchers recovered 315,320 reasoning blocks, 367 PII artifacts, and 182 credentials across Anthropic, OpenAI, and Google systems.
📡 Trending on X
"Rented margin" is the phrase of the week. Builders are posting their DeepSeek invoices and doing the 4.6x math in public — the replies are half panic, half I-told-you-so.
The open-weights crowd is split. V4-Pro is still MIT-licensed at 1.6T parameters, so one camp says "just self-host" while the other points out GPU rental costs more than the new API price for most teams.
Meta's Glimmer bet reopened the open-vs-closed fight. A 30B Apache 2.0 model that runs on a laptop landed the same week the cheap API stopped being cheap — the timing is doing a lot of arguing for Meta.
Multi-provider routing is having a moment. Gateway and router repos are trending as teams discover their "vendor-agnostic" architecture had exactly one vendor in it.
📺 Trending on YouTube
🛠 The Workflow: Price-proof your AI stack before Sunday
You have one day. This is the 30-minute version that protects your margin.
Pull your last 30 days of token usage and write down one number: monthly spend.
Multiply it by 4.6 for peak and 2.3 for off-peak. If either number breaks your P&L, you have a pricing problem, not an AI problem.
Check when your jobs actually run. Move every batch, cron, and backfill outside 01:00-04:00 and 06:00-10:00 UTC — that alone halves the hit.
Turn on prompt caching. Cache hits stay dramatically cheaper than cache misses, and most apps re-send the same system prompt thousands of times a day.
Put a router in front of your calls today — Gemini 3.7 Flash and Opus 5 both got cheaper this week. One config change beats one code rewrite.
Reply with the word "ROUTER" and I'll send you the one-page provider-cost comparison I built this morning.
🧰 Trending Tools
Rork Max — builds real iPhone apps from plain language. For operators who want a paid micro-app without hiring a dev.
Google Stitch 2.0 — turns a text description into a usable UI design. For founders who keep losing weeks to "what should this look like."
AdAnt AI — generates ad concepts and full campaign angles. For marketers who bill for creative volume.
Fathom 3.0 — records and summarizes every call automatically. For consultants who charge for follow-through, not note-taking.
Muse Code — writes, tests, and validates software autonomously. For small teams trying to ship like a big one.
📣 Put your brand here. The AI Innovator reaches AI-first operators, creators, and marketers every weekday. Primary sponsorships are now booking.
That's a wrap
Monday: the first real invoices from DeepSeek's new pricing land — we'll show what teams actually paid versus what they budgeted.
Enjoying the new format? Share your link — every referral keeps this free:
How was today’s email?
You're reading the 3-minute AI Innovator — same time, every weekday. Hit reply and tell me what you want more of. I read every one.
