In brief: According to Google, Gemini 3.7 Flash delivers noticeable coding and agentic improvements over its predecessor, but does not consistently outperform GPT-5.6 Terra and Claude Sonnet 5 in competitive comparisons, and costs only half its launch price until the end of 2026.
Google released Gemini 3.7 Flash just three weeks after its predecessor and is promoting it as its most capable model to date for coding and agentic tasks. A launch price reduced by half remains in effect until the end of 2026.
Google describes Gemini 3.7 Flash as its most intelligent “workhorse model” yet for coding and agents. Across several software-engineering benchmarks, the company reports significant progress over the predecessor Gemini 3.6 Flash: on FrontierCode 1.1 Main, which measures the quality of production-ready code, the score rises from 34.4 to 43.6 percent according to Google, and on the long-form test DeepSWE v1.1 from 49.0 to 65.3 percent. For web layout generation, the Elo rating in the WebDev Arena improves from 1538 to 1588, while the AutomationBench benchmark for enterprise automation jumps from 17.0 to 30.4 percent. Google also states that the model responds more flexibly to obstacles in agentic workflows, clarifies user intent more effectively, and follows instructions more reliably. In addition, safeguards against misuse in the areas of CBRN risks and cyberattacks were strengthened as part of the Frontier Safety Framework.
In direct comparison with competing models, Google’s own published benchmark tables show no consistent lead. On Terminal-bench 2.1, Gemini 3.7 Flash achieves 85.8 percent, while OpenAI’s GPT-5.6 Terra leads with 87.4 percent and also leads on Terminal-bench 3.0 and OSWorld-2.0. On the Agent’s Last Exam test, which evaluates multimodal desktop and operating-system tasks, Anthropic’s Claude Sonnet 5 achieves a success rate of 33.3 percent compared to 26.3 percent for Google’s model. For engineering teams, this means: Gemini 3.7 Flash positions itself less as a model superior across every discipline and more as a competitive option in a lower price bracket – the decision depends on the specific use case.
Until December 31, 2026, the model costs 0.75 US dollars per million input tokens and 3.75 US dollars per million output tokens – half the original price of Gemini 3.6 Flash in each case. From January 1, 2027, both prices will double to 1.50 and 7.50 US dollars respectively. Anyone planning to deploy the model at scale for coding or business agents therefore has a limited window to assess whether the stated improvements in error correction and manual oversight actually translate into lower total operating costs.
Gemini 3.7 Flash is available via Google AI Studio, Android Studio, the Gemini Enterprise Agent Platform, the Gemini Enterprise app, and Google Antigravity in more than 160 countries. In the consumer Gemini app, it also powers the always-on agentic feature Spark, which requires an AI Pro or Ultra subscription. The release comes as Google’s next flagship model, Gemini 3.5 Pro, still has no confirmed release date – even though CEO Sundar Pichai had announced it for the following month back in May. Google also confirmed that it is already working on the successor model, Gemini 4.
Source: www.it-daily.net · Published August 15, 2026
Lumi AI News — AI-assisted curation in accordance with Art. 50 EU AI Act. Paraphrasing and classification by Lumi News Pipeline v1.8.3.