On August 13, 2026, Google released Gemini 3.7 Flash, aimed at coding and AI agent workloads—only 23 days after announcing 3.6 Flash. According to Google's own evaluations, the new model shows substantial gains in code generation and long-horizon software development, and it's priced at half of 3.6 Flash's launch-time rate through the end of 2026. However, the current 3.6 Flash is also subject to the same limited-time pricing. For developers, the real change isn't an immediate drop in billed unit price—it's how much higher a success rate they can achieve at the same unit price.
Algorithmic refinements stacked over 23 days
Gemini 3.7 Flash is a model built on 3.6 Flash with algorithmic refinements applied to its reasoning component. Google explains that developer feedback also contributed to this rapid turnaround update. According to the model card, the balance among quality, cost, and latency can be adjusted via thinking settings. Inputs support text and images as well as audio and video. It handles up to 1 million tokens of context and 64,000 tokens of text output.
The model's API name is gemini-3.7-flash, and it launched as Stable rather than a preview. Developers can access it through the Gemini API, Google AI Studio, Android Studio, and Google Antigravity. For enterprises, it's rolled out to the Gemini Enterprise Agent Platform and the Gemini Enterprise app; for individual users, it's available via Gemini Spark for Google AI Pro and Ultra subscribers in supported countries.
What Google emphasized most in this update is the algorithmic refinement that sharpened reasoning and execution accuracy on top of 3.6. The existing model was refined in just 23 days and shipped as a Stable API.
Substantial gains concentrated in code evaluation
On FrontierCode 1.1 Main, which Google published, 3.7 Flash scored 43.6%, 9.2 points above 3.6 Flash's 34.4%. DeepSWE v1.1, which measures long-horizon software development, rose from roughly 49% to 65.3%. Note that the figure for 3.6 Flash is listed slightly differently across sources—49.0% in Google's announcement post and 48.6% in the model card.
| Benchmark | 3.7 Flash | 3.6 Flash | Claude Sonnet 5 | GPT-5.6 Terra |
|---|---|---|---|---|
| FrontierCode 1.1 Main | 43.6% | 34.4% | 42.7% | 41.3% |
| DeepSWE v1.1 | 65.3% | ~49% | 53.8% | 69.6% |
| AutomationBench | 30.4% | 17.0% | 10.7% | 23.6% |
| Intelligence Index | 56 | 52 | 55 | 57 |
Comparing the numbers, the shape of the improvement becomes clear. 3.7 Flash outperforms both rivals on FrontierCode and on AutomationBench, which measures enterprise task automation, but falls 4.3 points short of GPT-5.6 Terra on DeepSWE. The Artificial Analysis Intelligence Index, which aggregates overall model capability, also rose to 56 for 3.7 Flash, surpassing 3.6 Flash's 52 and Claude Sonnet 5's 55, but still trailing GPT-5.6 Terra's 57. The performance gains are concentrated especially in code generation and agentic execution.
The improvements aren't limited to code. Code Arena, which measures web development, rose from Elo 1538 to 1588, and GDP.pdf, which tests reading complex PDFs, climbed from 22.0% to 34.0%. However, CharXiv, which measures integrating complex charts and figures, actually declined—from 85.2% to 84.5% without tools, and from 89.4% to 88.7% with tools. Agent's Last Exam, which measures desktop operation, also came in at 26.3%, below Claude Sonnet 5's 33.3%. Even in Google's own table, there remain areas where the model trails competitors depending on the use case.
$0.75 through year-end—not a permanent price cut
Through December 31, 2026, the standard API pricing is $0.75 per million input tokens and $3.75 per million output tokens (including thinking). Since 3.6 Flash launched on July 21 at $1.50 for input and $7.50 for output, Google's claim of a "half price" reduction is a comparison against that launch-time rate. Under the current price list, 3.6 Flash is also set at $0.75 and $3.75, meaning that switching from 3.6 to 3.7 before year-end won't change the per-token rate at all.
The limited-time pricing will rise on January 1, 2027, to $1.50 for input and $7.50 for output. Take a simple example using 1 million tokens each of input and output: the standard rate would go from $4.50 to $9.00. So if you build your adoption plan for 3.7 Flash based solely on today's rates, API costs will double for the same usage volume once the new year arrives.
On the other hand, if task success rates improve at the same per-token rate—reducing the need for retries and rework—the cost per completed deliverable could actually fall. Google states that 3.7 Flash adapts better to obstacles, follows instructions more faithfully, and executes multi-step planning and tool calls with greater discipline. However, Google hasn't disclosed by what percentage output tokens or retry counts decrease compared to 3.6 Flash. To verify the price-performance tradeoff, you'll need to measure, for your own workloads, the total tokens and retry counts required to fully complete a task.
Can benchmark gains translate into real-world operating costs?
Support for up to 1 million input tokens and 64,000 output tokens gives you room to handle long codebases or multiple documents at once. On the flip side, deeper thinking increases billable output and changes latency. The model card lists hallucination, along with occasional slowdowns and timeouts, among its known limitations. The knowledge cutoff varies by domain—either March 2026 or January 2025—so workloads dealing with the latest information will also need to be supplemented with search or internal data.
Google's published evaluation figures show that 3.7 Flash outperforms 3.6 Flash on code fixes and business process automation. However, benchmark rankings can't substitute for measuring actual success rates, latency, and tool costs on your own repositories. During the same-price window through year-end, you can run 3.6 and 3.7 side by side on the same tasks and compare the total cost per completed job. Teams that verify whether that gap holds up even at double the unit price before the January 1, 2027 price change will be better positioned to move 3.7 Flash into production.