What happened?

Google announced Gemini 3.7 Flash, the newest member of the Flash tier within its Gemini model family, on August 13, 2026. The model arrived just three weeks after the release of Gemini 3.6 Flash. According to the model card, this is not a new pretraining run but rather a limited update consisting of algorithmic improvements to the core reasoning infrastructure.

Gemini 3.7 Flash can process text, image, audio, and video input within a 1-million-token context window, generate up to 64,000 tokens of output, and offers customizable thinking configurations that balance quality against cost and latency. The knowledge cutoff remains fixed at March 2026. The improvements are particularly concentrated in software engineering, document-heavy knowledge work, and web development.

Why does the pricing stand out?

Google's strongest claim about the model is its price. Gemini 3.7 Flash is sold at $0.75 per 1 million input tokens and $3.75 per 1 million output tokens—half the original list price of Gemini 3.6 Flash. This introductory pricing ends on December 31, 2026; starting January 1, 2027, the price will rise to $1.50 and $7.50.

By comparison, Claude Sonnet 5 is priced at $2.00/$10.00, while GPT-5.6 Terra is priced at $2.00/$12.00. Calculated using an 80% input–20% output mix, Gemini 3.7 Flash's blended cost comes to $1.35 per 1 million tokens, compared to $3.60 for Sonnet 5 and $4.00 for GPT-5.6 Terra. For teams running heavy agent workloads, this gap is presented as a more decisive factor than individual benchmark results.

What do the test results show?

On FrontierCode 1.1 Main, which measures production-quality code output, Gemini 3.7 Flash scored 43.6%, while Gemini 3.6 Flash scored only 34.4%. On DeepSWE v1.1, a long-horizon software engineering benchmark, the new model reached 65.3%. On WebDev Arena, it posted an Elo score of 1,588, the highest in Google's comparison table.

  • On the GDP.pdf document understanding benchmark, the score rose from 22.0% to 34.0%.
  • On AutomationBench, an enterprise workflow benchmark, the score climbed from 17.0% to 30.4%, surpassing Claude Sonnet 5's 10.7% and GPT-5.6 Terra's 23.6%.
  • On GDM-MRCR v2, a long-context benchmark at 128k context, the score reached 97.0%.
  • GPT-5.6 Terra outperformed Gemini 3.7 Flash on DeepSWE, Terminal-bench 2.1, Terminal-bench 3.0, and OSWorld-2.0.
  • On CharXiv Reasoning, the no-tools score fell to 84.5%, below Gemini 3.6 Flash's 85.2% result.

On the GDPval-AA v2 knowledge work benchmark, Gemini 3.7 Flash scored 1,525 Elo, while Claude Sonnet 5 scored 1,598 and Muse Spark 1.2 scored 1,628. On the Artificial Analysis Intelligence Index, Gemini 3.7 Flash scored 56, slightly behind GPT-5.6 Terra and Muse Spark 1.2, both of which scored 57.

Who can use it?

Gemini 3.7 Flash is available only through API and enterprise access; there is no open-weight version, so self-hosting or air-gapped deployment is not possible. The model can be accessed through the Gemini API, Google AI Studio, Google Antigravity, Android Studio, the Gemini Enterprise Agent Platform, and the Gemini Enterprise app. Consumers can reach the model through Gemini Spark on Google AI Pro and Ultra plans.

Google's own evaluation sets indicate the model can be used in law, financial services, life sciences, and enterprise operations. The entry-level pricing makes always-on AI agents more accessible for startups and mid-sized teams, though enterprises requiring data locality or air-gapped infrastructure cannot take advantage of this option.

What's next?

The only known timeline point is that the introductory pricing will end on December 31, 2026, and prices will double starting January 1, 2027. Given that Google has updated previous Flash versions at roughly three-week intervals, it remains an open question whether a similar update cadence will continue, though the company has made no official statement on the matter.