Google replaces Gemini 3.6 Flash after three weeks, cuts prices through year-end

Google DeepMind released Gemini 3.7 Flash on August 13th, replacing Gemini 3.6 Flash after 23 days in production and cutting prices through year-end. The model, led by Tulsee Doshi, Google's senior director and head of product for Gemini, improves agent performance with higher scores on coding benchmarks like FrontierCode 1.1 Main (43.6% vs 34.4%) and DeepSWE v1.1 (65.3% vs 48.6%).

Google DeepMind https://deepmind.google/about/?ref=runtimewire , the AI lab led by co-founder Demis Hassabis https://deepmind.google/about/?ref=runtimewire , released Gemini 3.7 Flash on August 13th, replacing a workhorse model that had been in production for 23 days and putting a temporary price cut behind its push into coding agents. The compressed release cycle belongs to Tulsee Doshi https://blog.google/authors/tulsee-doshi/?ref=runtimewire , Google's senior director and head of product for the Gemini model. Doshi described 3.7 Flash https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-3-7-flash/?ref=runtimewire as the result of developer feedback and algorithmic changes that Google expects to carry into later models. Ars Technica reported https://arstechnica.com/ai/2026/08/google-announces-gemini-3-7-flash-just-three-weeks-after-previous-release/?ref=runtimewire that the rollout comes as developers continue waiting for Gemini 3.5 Pro, which Google previously said was being tested with partners. Doshi has spent years at the point where model research becomes a product decision. Before taking responsibility for Gemini's roadmap, she led Google's Responsible AI product work for five years. She studied Symbolic Systems and computer science at Stanford, an academic path that combined the technical development of AI systems with questions about how people interact with them. Her latest release treats a frontier model less like an occasional research milestone and closer to production software under continuous revision. Google launched Gemini 3.6 Flash on July 21st https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-3-6-flash-3-5-flash-lite-3-5-flash-cyber/?ref=runtimewire , calling it a cheaper and more efficient replacement for 3.5 Flash. RuntimeWire reported at the time /article/google-launches-cheaper-gemini-3-6-flash-and-gates-cyber-model-access that Google was holding 3.5 Pro in partner testing while already beginning its Gemini 4 pre-training run. Three weeks later, 3.7 Flash has taken over the workhorse role. A model update built around agent failures Google says Gemini 3.7 Flash adapts better when an agent hits a roadblock, asks for clarification more effectively and follows multi-step instructions with fewer retries. Those improvements target a costly weakness in coding agents: a model can be inexpensive per token and still become expensive when it loops, edits the wrong files or needs repeated human correction. The developer guide https://ai.google.dev/gemini-api/docs/latest-model?ref=runtimewire lists a 1 million-token context window, a maximum output of 64,000 tokens and adjustable low, medium and high thinking levels. Gemini 3.7 Flash uses the same built-in tools as 3.6 Flash, limiting the migration work for developers already using the older model. The new version is based directly on 3.6 Flash, according to Google's model card https://deepmind.google/models/model-cards/gemini-3-7-flash/?ref=runtimewire . Google describes the changes as algorithmic improvements to the model's core reasoning foundation, rather than a new architecture or a fresh base model trained from scratch. That helps explain how Doshi's team moved from one numbered release to the next in 23 days. Google's reported coding gains are substantial on several tests. Gemini 3.7 Flash scored 43.6% on FrontierCode 1.1 Main, up from 34.4% for 3.6 Flash. Its DeepSWE v1.1 score rose to 65.3% from 48.6% in the model card, while its WebDev Arena Elo score increased to 1,588 from 1,538. The enterprise workflow results moved further. Google reported a 30.4% score on AutomationBench, compared with 17% for 3.6 Flash, and a 34% result on the GDP.pdf document-comprehension test, up from 22%. These remain Google-reported evaluations. The launch materials do not establish how the scores will translate into latency, reliability or total task cost inside production agent systems. A higher completion rate can reduce retries, though developers will need to test that effect against their own repositories, tools and approval processes. Google's full table is less tidy than the launch pitch The detailed evaluation table shows that 3.7 Flash did not improve every measured capability. Its no-tools CharXiv score, which tests reasoning over complex charts, slipped to 84.5% from 85.2%. With tools, it scored 88.7%, compared with 89.4% for 3.6 Flash. Other gains were narrow. The long-video LVBench result rose to 85.4% from 84.2%, while the Agent's Last Exam pass rate moved to 26.3% from 24.2%. Google's strongest evidence sits in coding, terminal work, document processing and business automation, the workloads Doshi's team is explicitly targeting. The model card also lists hallucinations, occasional slowness and timeouts among the model's known limitations. It gives Gemini 3.7 Flash a March 2026 knowledge cutoff, with some domains potentially limited to information from January 2025. Applications that depend on current information will still need search, retrieval or another source of live data. Google says its safety results were broadly similar to 3.6 Flash. The company added safeguards for chemical, biological, radiological, nuclear and cyber misuse, and said 3.7 Flash remained below the tracked critical capability thresholds in its frontier safety framework. The price cut expires Gemini 3.7 Flash costs $0.75 per million input tokens and $3.75 per million output tokens through December 31st. Google is applying the same promotional rates to 3.6 Flash, cutting both models to half of 3.6 Flash's original launch price. On January 1st, 2027, the rates are scheduled to rise to $1.50 for input and $7.50 for output. Developers choosing 3.7 Flash for its launch economics therefore face a straightforward planning problem: production workloads need to work at the January price, unless Google extends the promotion or changes the rate again. The temporary discount gives Doshi's team four and a half months to push developers onto the model, collect production feedback and demonstrate that fewer failed agent loops offset the eventual price increase. It also lets Google lower the immediate cost of testing 3.7 without permanently resetting the listed value of its main Flash tier. Gemini 3.7 Flash is generally available through the Gemini API, Google AI Studio, Android Studio, Google's Antigravity development platform and Gemini Enterprise. In the consumer app, Google is initially using it to power Gemini Spark for AI Pro and Ultra subscribers in supported countries. The standard Gemini chatbot remains on 3.6 Flash, according to Ars. That distribution choice puts the model first in places where repeated tool calls and long-running workflows can generate meaningful usage. Gemini Spark can consolidate files, draft emails and update status documents across Google Workspace, giving Google a controlled setting to measure whether 3.7's benchmark gains produce agents that require less supervision. Hassabis built DeepMind around a long-term effort to understand and recreate intelligence. Doshi's current task is shorter-cycle and operational: turning that research into a model developers will keep running after the introductory price disappears. Shipping 3.7 Flash three weeks after 3.6 shows how quickly Google can revise the product. Production results will determine whether the faster cadence saves developers the same amount of time.