Google releases Gemini 3.7 Flash, offering faster and cheaper AI capabilities with improved coding performance, yet facing limitations in reasoning and nuanced tasks.
Google’s latest lightweight Flash model is a sharper tool than its predecessor, but not a general answer to harder reasoning tasks. Decrypt said Gemini 3.7 Flash could produce a playable browser game from a single prompt in just over two minutes, after the earlier Flash release failed to do so at all. The newer model is also cheaper for now, with Google pricing input tokens at 75 cents per million through 31 December before the rate rises on 1 January. Decrypt reported that Google released the model on 13 August and made it available in more than 160 countries on day one.
The model’s appeal is practical rather than glamorous. It accepts up to one million input tokens, returns up to 64,000 output tokens, and can process images, video, audio and PDFs, while also calling tools and controlling a computer, according to Decrypt. Google has positioned Flash for high-volume work such as summarising documents, compressing long agent sessions and sorting text, rather than for the hardest forms of reasoning. On that use case, the model looks like a meaningful step forward.
The biggest improvement showed up in coding. Decrypt found that Gemini 3.7 Flash generated a working browser game on the first attempt, with clean syntax, stable collision handling and sensible scoring logic. That was a marked change from Gemini 3.6 Flash, which could not produce a functioning file and required outside repair. Google’s own benchmark sheet also claims strong performance across a broad range of tasks, including code and automation, though those figures remain the company’s own measurements.
The picture was less flattering in reasoning tests. Decrypt said the model gave the same wrong answer as Claude Fable 5 on a bridge puzzle by assuming a constraint that was never stated. It also set up a maths problem correctly but stopped short of finishing the calculation. In both cases, the model appeared able to organise the structure of a response without reliably checking the final step.
Writing performance was mixed too. Decrypt said the model handled a paradox-driven fiction prompt with coherent plot mechanics, but the prose was heavily overworked and unmistakably machine-like. Its associative thinking test was also underwhelming: the model named the metaphor too early, which weakened the transition the prompt was designed to probe. In each case, the output was competent at surface form but weak at subtle transformation.
Taken together, the model looks most convincing where speed, cost and reliable execution matter most. Google’s introductory price of 75 cents per million input tokens undercuts the company’s earlier Flash pricing, but the rate is scheduled to double in January. Decrypt’s conclusion was that Gemini 3.7 Flash is worth considering for users already embedded in Google’s ecosystem, especially for coding and agent workloads, yet it still trails better-performing alternatives when the task demands deeper reasoning or more original writing.
Disclaimer: This content is intended for informational purposes only. Readers are advised to exercise their own judgement, conduct due diligence, or consult a qualified expert before acting on any information provided.
