Gemini 3.7 Flash: What's New, Pricing, and Alternatives (2026)
Google announced Gemini 3.7 Flash on August 13, 2026, calling it "our most intelligent workhorse model yet for coding and agents." It costs half what 3.6 Flash cost, arriving three weeks after it. The listed price is introductory and expires at the end of the year.
This piece covers what changed, what the price actually is, and what the benchmarks measure. It also covers what to reach for when your deliverable is a report or a deck. Facts are as of August 17, 2026.
What's new in Gemini 3.7 Flash
The headline is the jump in agentic and coding work. Google lists the target areas as software engineering, knowledge work, web development, code generation, debugging, issue resolution, document comprehension and enterprise workflow automation.
Two of those matter to anyone working with files rather than code. Document comprehension and enterprise workflow automation are the ones that touch spreadsheets, statements and reports.
The release cadence is worth noting on its own. Gemini 3.6 Flash arrived roughly three weeks earlier, so this is a fast iteration rather than a generational change.
That pace has a practical consequence. Anything you build around a specific model version needs to tolerate the version moving underneath it.
Google also frames the tier explicitly as a workhorse rather than a frontier model. It is meant for volume and cost, not for the hardest single problem you have.
Availability splits by audience. Developers get it in Google AI Studio, Android Studio and Google Antigravity. Enterprises get it through the Gemini Enterprise Agent Platform. Individuals reach it through Gemini Spark, for AI Pro and Ultra subscribers across 160+ countries.
Those three surfaces are not the same product. A developer testing in AI Studio and a colleague using Spark are working with different interfaces and different limits.
Pricing, and the date that matters
| Now | From January 1, 2027 | |
|---|---|---|
| Input | $0.75 per 1M tokens | $1.50 per 1M tokens |
| Output | $3.75 per 1M tokens | $7.50 per 1M tokens |
Google states plainly that "Introductory pricing expires on December 31, 2026." The doubling is published, not speculated.
Against 3.6 Flash, Google describes the current rate as "half the original 3.6 Flash cost per million tokens." So the discount is real today and temporary by design.
If you are modelling a budget past this year, model the January numbers. A cost case built on the introductory rate will be wrong within months.
The reverse is also worth saying. For a project finishing this year, the current rate is real money saved rather than a marketing number.
One more detail affects planning. The doubling applies to both input and output, so the ratio between them does not shift.
What the benchmarks do and do not measure
Google publishes five comparisons against 3.6 Flash. The gains are large in every one.
| Benchmark | 3.7 Flash | 3.6 Flash |
|---|---|---|
| FrontierCode 1.1 Main | 43.6% | 34.4% |
| DeepSWE v1.1 | 65.3% | 49.0% |
| WebDev Arena Elo | 1588 | 1538 |
| GDP.pdf | 34.0% | 22.0% |
| AutomationBench | 30.4% | 17.0% |
Three of the five are coding evaluations. They tell you very little about whether a model reads your quarterly export correctly.
The one that comes closest to that work is GDP.pdf, a document-comprehension evaluation. Read its absolute number rather than its improvement: 34.0% means roughly two thirds of items are still missed.
AutomationBench behaves the same way. A jump from 17.0% to 30.4% is a large relative gain and a low absolute score.
That combination is the honest summary. This is a meaningfully better model on document work, and document work is still the weakest column on the board.
What this changes if your deliverable is a report or a deck
Cheaper and faster inference changes economics before it changes output quality. Halving the token cost makes it reasonable to run a step you previously skipped.
The obvious candidate is verification. Asking a second pass to check a figure against its source costs little when tokens are cheap.
The second candidate is volume. Processing every file in a folder becomes affordable rather than something you sample.
The third is iteration. When a first attempt is cheap, you can ask the same question three ways and compare the answers.
What does not change is the last mile. A model returning correct numbers still leaves you assembling the chart, the written summary and the slides. Our AI report generator page covers where that assembly step goes.
This is why cost per token is a weak proxy for cost per deliverable. The tokens are rarely the expensive part of producing a report.
Judge the change by what you can now afford to do twice. Checking your own work is the step most people skip on price grounds.
Three things the announcement does not say
The context window. No figure appears in the post. If your files are long, that omission is the first thing to test.
How it handles messy tabular input. Document comprehension is named as a strength without describing what happens to merged cells, multi-row headers or scanned tables.
Whether Spark availability equals API parity. Gemini Spark and the API are listed as separate surfaces, and the post does not say the behaviour is identical.
What the rate limits are. No per-minute or per-day quota appears in the announcement, which matters more at a low price than at a high one.
There is a fourth gap that matters for planning. Nothing is said about how long 3.7 Flash stays current, and 3.6 Flash lasted about three weeks.
Where a faster model stops helping
A model is a component. The work you are judged on is a finished artifact, and the gap between the two is where most of the time actually goes.
That gap has three parts. Getting the file into a usable shape, deciding which numbers matter, and producing something a colleague can read without you narrating it.
Notice that only the middle part is really about intelligence. The first and last are about handling files and producing formats.
None of those three gets solved by a cheaper token. They get solved by tooling that holds the file, the analysis and the output on one surface.
That is the layer Powerdrill Bloom works in. You upload the Excel, CSV, TSV or PDF file and explore it on a canvas. What you keep becomes a chart, a report or slides in one step. Its honest limitation sits in the other direction. It is not a raw model endpoint, so you do not control tokens, temperature or model choice.
Alternatives worth knowing
| If you need | Consider |
|---|---|
| Cheap, fast inference for agentic and coding work | Gemini 3.7 Flash |
| Deep reasoning on long, messy documents | A frontier model rather than a workhorse tier |
| A finished chart, report or deck from a data file | An action agent with file upload and export |
| An interactive view over a Google Sheet | Sheets canvas, covered in our Sheets canvas guide |
Pick by which part of the chain you are missing. If the missing piece is inference cost, a workhorse model is the answer.
Note that these are not competing purchases. A model tier and a delivery tool sit at different points, and most teams end up paying for both.
If the missing piece is the artifact at the end, a cheaper model does not produce it. That is a different tool category, not a weaker one.
Conclusion
Gemini 3.7 Flash is a genuine step up on agentic and document work, at half the previous cost, and the cost half is temporary. Those are the three facts worth carrying.
The number to keep in view is GDP.pdf at 34.0%. Improvement and sufficiency are different things, and document comprehension is still the column to verify yourself.
If your work ends in a chart, a written summary or a deck, try Powerdrill Bloom on one of your own files. See also our guides to agent plugins and what MCP is.
Frequently asked questions
What is Gemini 3.7 Flash?
It is Google's workhorse-tier model announced on August 13, 2026, described as its most intelligent workhorse model yet for coding and agents. Google lists coding, knowledge work, document comprehension and enterprise workflow automation among its target areas.
How much does Gemini 3.7 Flash cost?
Currently $0.75 per million input tokens and $3.75 per million output tokens. Google states that this introductory pricing expires on December 31, 2026, after which $1.50 and $7.50 apply.
Is it better than Gemini 3.6 Flash?
Google publishes gains on all five benchmarks it compares, including FrontierCode 1.1 Main at 43.6% against 34.4%. It also lists the current price as half the original 3.6 Flash cost per million tokens.
How large is its context window?
The announcement does not state a context window figure. If long files matter to your work, test that yourself rather than assuming parity with another tier.
Can it turn a spreadsheet into a presentation?
A model returns text and reasoning, not a formatted deck. Producing slides from a data file needs tooling that handles the file and the export, which is a separate layer from the model.