Iro AIIro AI Blog
Google shipped Gemini 3.7 Flash three weeks after 3.6 Flash, at half the introductory price, with tunable thinking and a 1M-token context window. Here is what actually changed, who it is for, and the honest answer for people who just use the Gemini app.
Gemini 3.7 Flash is Google's workhorse model, released 13 August 2026, and its headline is price-to-performance rather than a new capability ceiling. It landed three weeks after 3.6 Flash — Google credits developer feedback and algorithmic gains for the short turnaround — with meaningfully better coding and agent performance at half the introductory cost per million tokens.
It keeps the specs that made Flash useful: a 1M-token context window, up to 64k output tokens, and the same built-in tool suite as 3.6 Flash. What is new is tunable thinking at low, medium and high.
The honest framing: this is an infrastructure release. If you build with the API, it is a straightforward win. If you type into the Gemini app, you will struggle to feel the difference, and that is not a criticism of the model.
Google leads the announcement with software engineering, and the claims are specific enough to be checkable.
Note what is not claimed: no jump in reasoning ceiling, no new modality, no context increase. The 1M window and 64k output cap are unchanged. This is a sharpening, not a leap, and Google is fairly straight about that.
Iro AI turns ideas like the ones in this post into 5-minute exercises with feedback. Free tier, Pro from $4.17 a month ($49.99 a year, 7-day free trial).
The feature most worth understanding is thinking levels: low, medium and high. You are choosing how much the model deliberates before answering.
People read this as a quality dial. It is really three dials at once — quality, latency and cost — and they pull against each other.
Turning thinking up on a simple task is the common mistake. You pay more, wait longer, and often get an answer that is no better and occasionally worse, because the model talks itself out of the obvious. Knowing which tasks deserve deliberation is a real skill, and it transfers across every model that offers a reasoning setting.
Introductory pricing is $0.75 per million input tokens and $3.75 per million output, roughly half 3.6 Flash. That is the number in the headlines.
The part to diary: introductory pricing expires on 31 December 2026, after which it becomes $1.50 input and $7.50 output — a doubling.
| Input / 1M | Output / 1M | |
|---|---|---|
| Now, to 31 Dec 2026 | $0.75 | $3.75 |
| From 1 Jan 2027 | $1.50 | $7.50 |
If you are modelling unit economics on a product, model them at the January number. Building a margin on introductory pricing is how a healthy-looking service becomes unprofitable on a date you already knew about.
Depends entirely on how you touch it.
If you build on the API: yes. Better coding at half the cost is a clear win, and the thinking levels give you a genuine lever over spend.
If you use the Gemini app: honestly, not much. You do not choose the model tier, you are not billed per token, and the improvements are concentrated in coding and agent workflows most app users never touch.
Which points at something the release cycle obscures. Models now ship every few weeks, and almost none of those releases change what a normal person can get done. The gap between a good result and a bad one is still mostly the prompt. A clear task, enough context, explicit constraints and a stated output format will out-perform a version upgrade nearly every time — and unlike the version number, that skill carries to whatever ships next month.
That is the case for practising the durable part rather than chasing the changelog. If you want to build it deliberately, Iro's Gemini path works on exactly this, and the four things every prompt needs is the free version of the idea.
Iro AI turns ideas like the ones in this post into 5-minute exercises with feedback. Free tier, Pro from $4.17 a month ($49.99 a year, 7-day free trial).
Gemini 3.7 Flash was released on 13 August 2026, about three weeks after Gemini 3.6 Flash. Google attributed the fast turnaround to developer feedback and algorithmic improvements, and the model is generally available for production use.
Introductory pricing is $0.75 per million input tokens and $3.75 per million output tokens, roughly half the cost of Gemini 3.6 Flash. That introductory rate runs until 31 December 2026, after which pricing doubles to $1.50 per million input and $7.50 per million output.
Gemini 3.7 Flash supports a 1M-token context window and up to 64k output tokens, unchanged from Gemini 3.6 Flash. The release focused on coding, agent performance and price rather than on expanding context.
Thinking levels — low, medium and high — let you set how much the model deliberates before answering. It is a cost and latency control as much as a quality one: low suits high-volume, unambiguous tasks like classification and extraction, medium is a reasonable default, and high is worth the extra tokens only on genuinely hard problems such as multi-step debugging.
For coding, agents and web development, yes — Google reports higher first-pass code accuracy and more feature-complete results in fewer prompts, at half the introductory price. For everyday chat use in the Gemini app the difference is minor, because the improvements concentrate in developer workflows most app users never touch.