Gemini 3.8 Flash pushes agents to work longer

Google is launching Gemini 3.8 Flash for coding, autonomous agents and complex tasks. A more permissive Cyber version is restricted to approved organizations.

A Flash model capable of pursuing a task, checking its work and making repeated tool calls instead of stopping at the first plausible answer. That is the promise of Gemini 3.8 Flash, launched by Google only three weeks after Gemini 3.7 Flash.

The new version is not solely designed to respond faster. On complex requests, it can break its reasoning into more steps, attempt several operations and check their results before reaching a conclusion. Google summarizes this change by saying that the model “works harder.”

This behavior primarily targets long-running software development, autonomous agents and professional workflows made up of several stages. Gemini 3.8 Flash can, for example, examine multiple files, modify a project, run tools, identify an error and resume its work without waiting for another instruction.

The tradeoff appears in consumption. Two requests billed at the same unit price will not necessarily cost the same if the second produces more tokens, performs more calls and remains inside a longer loop. The listed price is therefore not enough to determine the true cost of completing a task.

Google offers three reasoning levels: low, medium and high. Medium is enabled by default. Low reduces latency and consumption for routine work, while high allocates more resources to reasoning, verification and tool use. The minimal level available on some other Gemini models is not supported by 3.8 Flash.

Teams prioritizing speed or predictable spending can continue using Gemini 3.7 Flash, which remains available. This coexistence shows that version 3.8 does not automatically replace its predecessor: its value depends on what the additional steps contribute.

Early measurements published by Artificial Analysis illustrate this compromise. According to findings reported by The Verge, Gemini 3.8 Flash used about 30% more output tokens per task and cost nearly 40% more than 3.7 Flash in its evaluations, despite identical unit pricing. These figures remain preliminary and will vary according to the request and reasoning level.

Standard introductory pricing is set at $0.75 per million input tokens and $3.75 per million output tokens, including reasoning tokens. These rates remain in effect through December 31, 2026. On January 1, 2027, they will rise to $1.50 for input and $7.50 for output.

Deferred processing through Batch and Flex costs half as much: $0.375 per million input tokens and $1.875 for output during the introductory period. Priority mode, intended for requests that require more predictable capacity, costs $1.35 for input and $6.75 for output.

Context caching costs $0.075 per million tokens, plus $0.50 per million tokens for each hour of storage. These prices will also double in January 2027. Google Search and Maps grounding each include a monthly allowance before additional queries begin incurring charges.

A free tier is available through the Gemini API. The official pricing page states, however, that content submitted through this tier may be used to improve Google’s products. Content from the paid tier is not used for that purpose.

Gemini 3.8 Flash accepts text, images, video, PDF documents and audio files. Its responses remain text-only: it does not directly generate images or sound. Its input context reaches 1,048,576 tokens, while outputs can extend to 65,536 tokens.

The model supports function calling, structured responses, code execution, file search, caching, URL analysis and grounding with Google Search and Maps. Computer control is available in preview. The Live API, image generation and audio generation are not supported.

Developers can access it through Google AI Studio and the Gemini API under the identifier `gemini-3.8-flash`. The model is also available through Gemini Enterprise and becomes the default engine for the Antigravity agent. For consumers, it is joining the Gemini app, AI Mode in Google Search and Gemini in Google Sheets, with access announced for Google AI Pro and Ultra subscribers.

Demonstrations include the creation of a three-dimensional game from a looping instruction, a DOS-inspired version of Google Maps and an interactive tool for virtually taking apart hardware devices. These examples mainly demonstrate the ability to sustain a technical build across several stages. They do not independently measure the reliability of the resulting code.

Google is supporting the launch with a series of benchmark results. On DeepSWE v1.1, which focuses on autonomously solving long-running software engineering problems, Gemini 3.8 Flash scores 73.7%. It sits just behind Claude Opus 5 at 74%, but ahead of GPT-5.6 Sol at 72.7%, GPT-5.6 Terra at 69.6% and Gemini 3.7 Flash at 65.3%.

The narrow gap with Opus 5 makes it difficult to claim a clear victory. Instead, it shows that a model presented as faster and less expensive can approach a much more costly system on this particular protocol. Google’s table lists Opus 5 at $25 per million output tokens, compared with $3.75 for Gemini 3.8 Flash during its introductory period.

On Terminal-Bench 2.1, 3.8 Flash reaches 89.4%, slightly ahead of Opus 5 at 89.1% and GPT-5.6 Sol at 88.8%. The picture changes completely on Terminal-Bench 4.0, a newer and more difficult version: Gemini falls to 19.1%, compared with 51.8% for Opus 5 and 37.3% for GPT-5.6 Sol.

This difference is a reminder that sharing a benchmark name does not guarantee continuity between versions. An excellent result on Terminal-Bench 2.1 cannot be used to infer performance on more demanding tasks, environments or evaluation criteria.

Gemini 3.8 Flash leads Google’s table on Vals Finance Agent v2 with 61.4%, ahead of 3.7 Flash at 59% and Opus 5 at 58.6%. On Harvey’s legal benchmark, it achieves a 10% overall pass rate, compared with 8.8% for 3.7 Flash and 6.7% for Opus 5.

The legal result remains low in absolute terms. Finishing first with a one-in-ten success rate means that most of the workflows being tested are still not completed. The relative ranking should not obscure the limited