admin The model, released on September 2, is the latest version in Google’s Gemini 3 Flash family and is designed to combine stronger reasoning and coding performance with the lower latency and cost profile associated with the Flash tier. Google said Gemini 3.8 Flash is ready for production use through its developer and enterprise platforms.
Gemini 3.8 Flash supports a context window of about one million input tokens and up to 65,536 output tokens. It accepts text, images, video, audio and PDF files as inputs and produces text output. Developers can use functions including code execution, file search, function calling, structured outputs, search grounding, URL context and computer use, the latter still offered in preview.
A central change is greater emphasis on sustained reasoning across extended tasks. Google said the model is intended to handle real-world software engineering jobs involving multiple files, repeated tool calls and longer sequences of decisions, rather than only short coding prompts. The company said its evaluations show stronger performance than Gemini 3.7 Flash on software engineering and agentic knowledge workflows.
On the DeepSWE v1.1 benchmark for long-horizon software engineering, Google said Gemini 3.8 Flash outperformed most larger frontier models while operating at a lower cost. The company also reported a 61.4 per cent score on Vals Finance Agent v2 and 54.9 per cent on HLE-Verified, a benchmark covering multi-step reasoning across scientific, humanities and professional subjects.
Google said Gemini 3.8 Flash also improves deterministic tool execution, an important requirement for agents used in production environments where the same instruction must reliably trigger the intended action. Its tool set includes caching and support for batch, flex and priority inference. The model does not support image or audio generation, and the Live API is unavailable. Google lists search grounding, Google Maps grounding and multimodal understanding among supported capabilities for applications that combine external information with model reasoning.
The model is also built for autonomous agents that must plan, call tools, inspect results and recover from failures over several steps. Google said the upgrade is intended to reduce failed loops and execution errors in workflows that rely on repeated reasoning and tool orchestration.
For enterprise users, Google is positioning Gemini 3.8 Flash for large-scale knowledge work, specialised analysis and data-processing pipelines where factual accuracy and consistent tool execution are important. The company said the model can sustain longer document-heavy tasks and perform multi-stage work across specialised domains.
Developers can adjust the model’s thinking level to low, medium or high, allowing them to trade greater reasoning effort for lower latency and token consumption. Medium is the default setting. Google cautioned that the model can use more tokens on difficult tasks, particularly at higher effort levels, because it may take additional reasoning steps and call tools repeatedly.
That behaviour could make some jobs more expensive even though the headline token prices are unchanged from Gemini 3.7 Flash during the introductory period. Google is charging $0.75 per million input tokens and $3.75 per million output tokens through December 31. From January 1, 2027, the rates are scheduled to rise to $1.50 and $7.50 respectively.
Gemini 3.8 Flash is available through the Gemini API, Google AI Studio, Gemini Enterprise and Google Antigravity. Google has also made it available to Google AI Pro and Ultra subscribers through the Gemini app, AI Mode in Search and Gemini in Google Sheets.
The release came three weeks after Gemini 3.7 Flash and marked Google’s third Flash model release in six weeks, reflecting the rapid pace of upgrades across its generative AI portfolio. Google said Gemini 3.8 Flash is based on Gemini 3.7 Flash but incorporates further advances in reasoning, coding and agentic execution.
