Gemini 3.7 Flash costs half the old workhorse price — so does 3.6 Flash

Three things happened to AI model pricing this week. Two went down and one is scheduled to go up.

On August 10 Anthropic cancelled a scheduled increase. On August 13 Google halved its workhorse tier. On August 16 DeepSeek has said it will raise prices.

Start with the one in the middle.

1. What exactly got halved

Google released Gemini 3.7 Flash on August 13, 2026.

The baseline for "half" is the workhorse tier price before August 13: $1.50/$7.50. Against that, $0.75/$3.75 is exactly half.

The complication is what else happened that day. Google cut the price of the older Gemini 3.6 Flash to $0.75/$3.75 as well.

The table says one thing plainly: 3.7 Flash has never been cheaper than 3.6 Flash. They are level through year-end and level after it. What fell was the tier, not a model, and only for four and a half months.

For a developer this distinction is operational. If you are running 3.6 Flash today, your bill halved on August 13 without you touching anything. The reason to migrate to 3.7 is capability, not price.

2. Capability — 23 days later

Google's stated improvements are production-oriented: debugging and other coding work, and generating deployable code on the first attempt. In the company's phrasing, it "better adapts to roadblocks, clarifies intent when needed, and follows instructions with greater fidelity."

The striking figure is the interval. Gemini 3.6 Flash shipped 23 days earlier. And 3.7 is not a new base model — it is built on 3.6, with Google citing developer feedback and algorithmic improvements.

Whether a fast follow-up is a good sign is genuinely ambiguous. It can mean the iteration loop has tightened. It can also mean the workhorse tier is filling a gap left higher up the lineup. And there is a gap: Google's top model, Gemini 3.5 Pro, remains delayed, with no stated release date.

3. The week's three prices

Anthropic's move is the cleanest of the three. Sonnet 5 launched with $2 per million input tokens and $10 per million output tokens described as introductory pricing through August 31, 2026, with a rise to $3/$15 scheduled for September 1. On August 10 that increase was cancelled and the introductory rate became the standard rate. Subscription plan prices were unchanged.

DeepSeek moved the other way. It released the full version of V4-Pro on August 13 and flagged an API price increase for August 16, three days later. That model's architecture — 1.6tn total parameters with roughly 49bn active on any single inference — was covered in "DeepSeek V4-Pro ships."

Two companies cut, one raises. The cut came with an expiry date (Google, December 31); the cancellation removed one (Anthropic). A discount with a deadline and a price with no deadline ask a developer for different arithmetic.

4. These rates are not your bill

One thing to keep straight: what is printed above is a list price, not an invoice.

At least three other factors set the real number.

① Output costs 5× input. $0.75 against $3.75. The same million tokens costs five times more if the model wrote them. Summarising a long document and generating long text from a short prompt have completely different cost profiles.

② Conversations accumulate input. A model has no memory between calls, so the entire prior exchange is resent as input on every request. A ten-turn session does not cost ten times a one-turn session; it costs considerably more. The structure is set out in "What a context window is."

③ Caching and batch discounts exist. Repeated prefixes billed at a cache rate, and non-urgent work submitted as batches, both carry different rates. This article does not cover those.

The arithmetic that turns list rates into real spend is in "What AI token pricing is."

5. What is still open

Nobody knows when 3.5 Pro arrives. Google's top model is delayed, with neither a date nor a reason given. Whether the rapid workhorse cadence relates to that gap is inference, not fact.

Context length and maximum output were not verified. For choosing a workhorse model those matter as much as price, and this article could not confirm them.

The comparison model is unsettled. Coverage of what still leads on harder agentic tasks names different OpenAI versions. Rather than pick one, this article leaves it unnamed.

The largest unknown is January 1, 2027, when $0.75/$3.75 is scheduled to become $1.50/$7.50 again. Though as Anthropic demonstrated in the same week, a scheduled price change is not necessarily an executed one.