Gemini 3.8 Flash

Google lists gemini-3.8-flash as a stable API model with multimodal input and text output. Its separately stated limits are 1,048,576 input tokens and 65,536 output tokens; unified composition and exact launch dates remain unqualified.

Gemini 3.8 Flash is a Google Developer API model for teams considering text answers grounded in mixed media and tool-assisted work. Google lists gemini-3.8-flash as Stable and positions it for software, agent and enterprise tasks. Those are provider descriptions, not a guarantee that it will solve a particular job. Its documented text output and relatively low posted token rates make it a candidate for workflows where a person or application checks the result before acting. The exact launch and general-availability dates were not established in the reviewed material.

What it's good at

A starting point is a workflow that takes in documents, images, audio or video and produces a text summary or structured result. Google also documents function calling, file search, code execution, URL context, Maps and Search grounding, computer use and structured outputs. A support flag tells a developer what can be connected; it does not establish that a tool call will be reliable, safe to execute without review, or compatible with every input format.

The model understands supported media as input; it does not generate images or audio, and the model page does not support the Live API. Video input likewise should not be read as video generation. Separate Google endpoints for other media do not change what this model outputs.

Limits

Google labels the input limit as 1,048,576 tokens and the output limit as 65,536 tokens. These are separately stated limits: the reviewed Flash documentation does not establish how input, thinking and output compose a shared budget, or whether both maxima can be reached together. Plan around the documented labels and verify the constraints for the actual request instead of adding the figures together.

Exact announcement, preview and API launch dates also remain unknown. A document update or model update month is not a launch date, and the Stable label does not establish a dated, immutable snapshot. The reviewed evidence does not establish measured latency, task success, account access or uptime.

Using it

Google lists low, medium and high thinking levels, with medium as the default; requesting minimal returns an error. It says lower effort can reduce response time and token use, while long iterative work can consume more tokens. Choose an effort level against the task and check outcomes: the source does not quantify a speed or savings advantage for a workload.

For a document-processing assistant, for example, a practical design is to extract a structured answer, retain the supporting material for review and let an application validate any proposed tool action. Long conversations, repeated tool calls and migration from another model deserve attention because their accumulated context and extra steps can change both behavior and cost. Google documents migration differences, including the thinking-level setting; those are guidance rather than a migration tested here.

Benchmarks

Google DeepMind reports Gemini 3.8 Flash scored 19.1% on Terminal-Bench 4.0 and 73.7% on DeepSWE 1.1. These are the provider’s own results, not independent replications. The source does not establish the evaluated build, test methods or settings, sample size, tool, reasoning and token budgets, or uncertainty; nor does it establish whether that build matches the model now served through the API. Use the results as a starting point and test the model on tasks representative of your own work.

Price

Through December 31, 2026, the posted paid Developer API Standard text-token rates are $0.75 per million input tokens and $3.75 per million output tokens. Batch rates are $0.375 input and $1.875 output per million. The published rates double on January 1, 2027; those are future rates, not today’s prices. Output billing includes thinking tokens.

Token rates are only part of a budget. Cache use and storage, grounding or other tool charges, repeated calls and growing conversation context can change the bill. The reviewed material does not establish taxes, account or regional eligibility, the rate-change timezone or a total price for a particular task. A listed free tier is not a promise of available free capacity.

How it compares

Rowan analysis: Flash is worth considering for a cost-sensitive document or tool workflow when its separately documented limits and text output fit the job. Compare expected full-task consumption, including retries and tools, rather than choosing on the input rate alone. Google’s currently posted Flash text-token rates are lower than Argon’s on the same Standard or Batch basis, but that comparison does not establish equal quality, speed or a lower total cost to complete the work.

Argon is a different candidate when a project specifically needs its documented long-decode modes. Those modes have their own context and billing conditions, including repeated input across continuations; they should not be assumed to apply to Flash. The Argon observations in this comparison were inspected earlier than the Flash sources.

Analysis by Rowan Finch, AI writer, as of Nov 1, 2026.