AI Price War: Google Halves Rates, DeepSeek Quadruples Its Own

Gemini 3.7 Flash arrived on Thursday, August 13, twenty-three days after Gemini 3.6 Flash. That pace is no longer a product calendar — it's a rate of fire, and it says something about Google's position: the Flash line groups together fast, cheap models built for high volume and automated agents. What's still missing, though, is the flagship model. Gemini 3.5 Pro, promised at May's developer conference and tested with partners in July, remains unreleased.

Plenty of Flash, still no Pro

On paper, the new version makes clear gains in coding: Google claims a jump from 49% to 65.3% on the DeepSWE benchmark, from 34.4% to 43.6% on another coding test, and a task-automation score that has nearly doubled. These are in-house measurements that haven't been independently validated, so they deserve a grain of salt. The internal context doesn't exactly inspire confidence either. DeepMind underwent a full reshuffle the week before, with its longtime head stepping aside for his deputy while two technical leads behind Gemini left the company to start their own.

The real announcement was in the price list

Gemini 3.7 Flash is priced at $0.75 per million input tokens and $3.75 per million output tokens — half the launch price of the previous model. It's a promotional rate good through the end of the year, before returning to full price. The strategic choice is clear: rather than chase the world's best model, sell mass compute at the most aggressive price possible.

Except that ground is already occupied. DeepSeek has been charging around $0.14 per million input tokens for its fast version — five times cheaper than Google — and releases its models under an open license, meaning any company can run them on its own servers without paying anyone. Qwen, from Alibaba, plays the same game with a permissive license and hundreds of downloadable variants. GLM, Kimi and MiniMax round out a Chinese lineup that has turned price-cutting into an industrial strategy.

DeepSeek's reversal

That's exactly where the week gets interesting. On August 13, DeepSeek announced a new price list on its own site, effective August 16. DeepSeek-V4-Flash now costs $1.32 per million output tokens during peak hours, up from $0.28 — more than a fourfold increase. V4-Pro follows the same pattern, at $3.96 versus $0.87. Off-peak hours remain half price, with the company saying it wants to smooth demand through pricing rather than through a queue.

• Gemini 3.7 Flash: $0.75 per million input tokens, $3.75 per million output tokens — promotional rate through the end of 2026.
• DeepSeek-V4-Flash: $1.32 per million output tokens at peak hours, up from $0.28 before August 16.
• DeepSeek-V4-Pro: $3.96 per million output tokens at peak hours, up from $0.87 previously.
• For comparison, Kimi K3 is priced at $15 and Fable 5 at roughly $50 per million output tokens.

Even after this increase, DeepSeek remains far cheaper than most of its rivals: Moonshot charges $15 per million output tokens for Kimi K3, and Anthropic around $50 for Fable 5, according to comparisons cited by Bloomberg. The move has less to do with costs than with the financial calendar: the Hangzhou-based company is raising funds and eyeing an IPO as early as 2026. In other words, the market's rock-bottom price was never an economic equilibrium — it was a land-grab position that its own author is now starting to abandon.

What it means for English-speaking businesses

Two practical consequences stand out. The first is about budgeting: the prices used today to build an automation budget are, on both sides, political prices — a limited-time promotion at Google, a land-grab position being wound down at DeepSeek. A workload plan built on August 2026 prices will need to be recalculated in January.

The second is geographic. The new model powers Spark, Google's personal agent designed to carry out tasks on a user's behalf, limited to paying subscribers and rolled out in more than 160 countries. The European Economic Area, Switzerland and the United Kingdom are excluded, with no official explanation — the same carve-out Apple applied to its new generation of Siri. Developers, including in the UK, can use Gemini 3.7 Flash through dedicated tools, but the consumer agent built on top of it doesn't cross that border. For businesses on either side of the Atlantic, the question isn't only which model costs less — it's which one will actually be available to their teams.