Scale · founder · 7 min read

Google Halved Its Coding Model Price. DeepSeek Raised Its 11x. Same Week.

In three days, the model layer under your vibe coding tool moved hard in both directions. Here's what actually reaches your bill, and what to do.

Published August 15, 2026 ·

pricing models gemini deepseek model routing budget

Between August 13 and August 14, two things happened that point in opposite directions.

Google launched Gemini 3.7 Flash and priced it at $0.75 per million input tokens and $3.75 per million output tokens — half what 3.6 Flash cost at launch, three weeks earlier. On the same schedule, DeepSeek announced new API pricing for its V4 family, taking some rates up by more than 1,100% and switching to a peak/off-peak clock. Those increases take effect August 16.

You don’t buy tokens. You buy a $25/month Lovable seat or a $20 Cursor subscription. So why does this matter?

Because the tool you pay a flat fee to is buying tokens, and the gap between what it pays and what it charges you is the entire business. When that gap moves, your plan moves — as a credit limit, a “fair use” cap, a slower default model, or a price change three months later. Every billing change this site has covered in 2026 started as a model-layer move first.

What Google actually shipped

Gemini 3.7 Flash is aimed squarely at coding and agent work. Google calls it “our most intelligent workhorse model yet for coding and agents,” which is marketing, but the benchmark deltas over 3.6 Flash are steep enough to note:

• DeepSWE v1.1 (issue resolution): 49.0% → 65.3%
• FrontierCode 1.1 Main: 34.4% → 43.6%
• WebDev Arena Elo: 1538 → 1588
• AutomationBench (multi-step business workflows): 17.0% → 30.4%

These are vendor numbers on vendor evals, and they should be treated as a claim until independent production evidence catches up. Greyhound Research’s Sanchit Gogia made exactly that point when the model landed, and he’s right.

The number that isn’t a claim is the price. $0.75/$3.75 is an introductory rate running through the end of 2026 — which means it’s a promotional floor, not a permanent one. Diarise that.

The model is live in Google Antigravity, AI Studio, Android Studio, the Gemini API, and the Gemini Enterprise surfaces. If you use Antigravity, it’s already there — you just need to upgrade.

What DeepSeek actually did

DeepSeek’s V4-Pro went generally available and V4-Flash went to beta, and the pricing came with them:

• V4-Flash: $0.22/M input (cache miss) and $0.66/M output off-peak; $0.44 and $1.32 at peak. Previously a flat $0.14 and $0.28.
• V4-Pro: $0.66/M input and $1.98/M output off-peak; $1.32 and $3.96 at peak. Previously $0.435 and $0.87.

The eye-watering “1,100%” figure comes from cached input tokens, where the old discount was extraordinary — roughly 98% against an industry norm nearer 90%. That cache discount was the real mechanism behind DeepSeek’s cost advantage, and repricing it is the substantive change here, not the headline percentage.

The nuance that matters: 17 of every 24 hours stay at the half-price off-peak rate, and the peak window falls hardest on DeepSeek’s home market. Most Western buyers will mostly pay off-peak. DeepSeek is still cheaper than the alternatives for most work. It just isn’t dramatically, structurally cheaper anymore.

Info-Tech’s Mark Tauschek called it plainly: supply and demand. Compute is constrained, demand isn’t, and Anthropic raised prices for the same reason in April. Expect more of this, not less.

The thing to actually take from this

The instinct is to read one of these as good news and the other as bad news. That’s not the useful frame.

The useful frame is that model pricing is now a variable, not a constant — it moves in both directions, on a few days’ notice, sometimes with no changelog at all. OpenAI quietly cut Codex’s configured context window from 372K to 272K tokens in July and labelled it a “correction” four days later. Anthropic raised prices in April. Google cut prices twice in six weeks. DeepSeek is now literally pricing by the hour.

If you’re building a product on top of any of this, the design principle is the one enterprise architects have already landed on: don’t hard-code the assumption. Concretely, for a founder or PM:

Don’t pick a tool because of this week’s model price. The model underneath your builder can change without you being told. We covered how to find out which model your tool is actually running in Which Model Is Your Vibe Coding Tool Actually Running? — the answer changes more often than the marketing page does.

Assume the introductory rate expires. Gemini 3.7 Flash’s price is explicitly promotional through December. Budget on the post-promo number, not the launch number.

Watch for the pass-through, not the announcement. A model price cut rarely shows up as a lower subscription. It shows up as more generous credits, a better default model on the free tier, or longer agent runs before you hit a wall. A price rise shows up as the reverse. If your builder quietly downgrades its default model, that’s the model layer reaching your account.

Know where your work can move. The cheapest insurance against any of this is portability: can you export the codebase and run it somewhere else? For most builders the answer is yes, and it’s worth checking before you need it rather than after.

The routing layer is getting crowded, and that’s the real signal

The more interesting development sits one level up. On August 14, Nvidia launched NeMo Switchyard, a library for routing requests across models — picking the cheapest one that can do the job. Cloudflare shipped a model router earlier this summer. Stripe has been in talks since late July to buy OpenRouter, the marketplace that fronts 400+ models from ~70 providers, at a reported ~$10B against a $1.3B valuation in May.

Three infrastructure companies, none of them model vendors, all building the same thing: the toll booth between your app and whichever model is cheapest today. That only makes sense as a business if everyone expects model prices to keep moving unpredictably. The routing layer is a hedge, and serious money is buying it.

For you, the practical version of the same hedge is much smaller: your coding tool probably already routes automatically, and you should know what it’s routing to. We wrote that up in Your AI coding tool now picks the model for you.

Bottom line

Nothing about this week changes what you should build or which tool you should start with. Gemini 3.7 Flash getting better and cheaper is straightforwardly good for anyone using Antigravity or AI Studio. DeepSeek getting more expensive matters mainly if you were using it directly through the API, which most non-technical founders aren’t.

What it changes is your assumption. The price of the intelligence under your product is not a fixed input. Build and budget accordingly — and if you want the full picture on what building this way actually costs, start with The Real Cost of Vibe Coding.