Gemini 3.8 Flash: Pricing, Benchmarks and Google’s Fairwind Cyber Rollout

Google released Gemini 3.8 Flash as a general-purpose model for software engineering, agentic tasks and specialized multi-step reasoning, alongside a separately gated cyber-defense model. The central pitch is not simply faster, cheaper inference: Google is positioning Flash as a capable model for longer-running developer and agent workflows while keeping its entry price well below several frontier competitors. 8

What Google announced

Gemini 3.8 Flash arrived three weeks after Gemini 3.7 Flash, making it the third Flash-family release in six weeks. Google calls it its most intelligent “workhorse” model and says it improves on 3.7 Flash in software engineering, agentic tasks and critical multi-step reasoning. 38

For developers evaluating the release, the practical distinction is that Google is aiming Flash at work that can involve planning, iterative execution and tool use—not only short, low-latency prompts. That can make it a more attractive default for coding assistants and autonomous workflows, but it also makes usage patterns important to the real cost.

Gemini 3.8 Flash pricing: introductory versus standard rates

Google set introductory API pricing at $0.75 per million input tokens and $3.75 per million output tokens, the same rate offered for Gemini 3.7 Flash. The introductory rate lasts through December 31, 2026. Starting January 1, 2027, the reported standard price is $1.50 per million input tokens and $7.50 per million output tokens. 38

That gives teams a clear budgeting caveat: an evaluation performed at launch pricing will not reflect the standard 2027 list rate.

Why token price is not the whole cost story

A model designed to reason for longer, check its own work and call tools repeatedly may consume more output tokens and generate more tool traffic per completed task. Artificial Analysis reported that Gemini 3.8 Flash’s measured cost was about $0.58 per Intelligence Index task, versus $0.40 for Gemini 3.7 Flash, alongside a higher average output-token count. 2532

In other words, low per-token pricing can still produce a meaningful bill on complex agent loops. Teams should measure cost per completed task on their own repositories and workflows—not just input and output token rates. For simpler work where 3.7 Flash already meets quality requirements, keeping that model in the routing mix can be a sensible cost-control choice while the two models share introductory pricing. 3

How Gemini 3.8 Flash compares with GPT-5.6 Sol, Grok 4.6 and Claude Fable 5.1

On Artificial Analysis’s composite Intelligence Index, Gemini 3.8 Flash scored 59 at its high reasoning setting, three points above Gemini 3.7 Flash. That matched GPT-5.6 Sol at extra-high reasoning and Grok 4.6 at medium reasoning. At their higher settings in the same reporting, GPT-5.6 Sol and Grok 4.6 reached 61, while Claude Fable 5.1 reached 66. 21

The comparison is useful, but it is not a universal ranking. Reasoning settings, task selection, latency limits, tool-use policies and token budgets can all change outcomes.

Model	Reported Intelligence Index result	Headline API price per 1M input/output tokens
Gemini 3.8 Flash	59 at high reasoning	$0.75 / $3.75 introductory; $1.50 / $7.50 standard 321
GPT-5.6 Sol	59 at extra-high; 61 at maximum	$5 / $30 2021
Grok 4.6	59 at medium; 61 at high	$2 / $6 2021
Claude Fable 5.1	66 at maximum	$10 / $50 2126

Gemini’s launch token rate is therefore materially lower than the listed rates cited for those competitors. But a lower list price is not proof that it is cheaper for every workload. For example, CursorBench results included in the available sources show a different task-specific picture: Fable 5.1 Max posted a higher success rate than Gemini 3.8 Flash High, while Gemini had a lower measured run cost in that benchmark. 30

The appropriate decision is workload-specific: test models at the reasoning levels, context sizes and tool permissions that production actually requires.

Fairwind and Gemini 3.8 Flash Cyber

Alongside the general model, Google introduced the Fairwind Program, a limited-access initiative for governments, trusted partners and priority defenders. Google identifies examples including healthcare providers and telecommunications services. 514

Fairwind participants receive early access to Gemini 3.8 Flash Cyber, a cybersecurity-focused variant that Google says is intended for vulnerability detection and automated patching. The program pairs the model with CodeMender, Google’s code-security agent, to support vulnerability research and software fixes. 814

This is not the same as making a cyber model broadly available to every API customer. The restricted rollout is significant because vulnerability discovery and remediation capabilities are dual-use: they can strengthen defenders, but sophisticated capabilities can also create misuse risks. Google’s choice to limit access to trusted defenders is therefore part of the product design, not merely an availability detail. 514

What the release says about Google’s AI strategy

The launch reflects a two-track approach:

• General availability: rapidly release an economical model for coding, reasoning and agentic development.
• Controlled access: put the more sensitive cyber capability behind a vetted defender program. 58

It also arrives amid uncertainty about Google’s next flagship model. Demis Hassabis moved from day-to-day leadership of Google DeepMind into the roles of DeepMind chair and Alphabet chief scientist; he did not leave Alphabet. Separately, reporting said Gemini 3.5 Pro, expected in June, had been delayed. 33

That context does not establish why Google is accelerating Flash releases. It does, however, make the product strategy easier to read: Gemini 3.8 Flash is a concrete, currently available bet on capable and cost-conscious agents, while Google’s flagship-model roadmap remains less clear.

Bottom line

Gemini 3.8 Flash is most compelling for teams that want stronger coding and agentic reasoning than a typical lightweight model at a low launch token price. Its benchmark position suggests competitive capability at selected reasoning settings, but it does not make it the blanket leader over GPT-5.6 Sol, Grok 4.6 or Claude Fable 5.1. 21

For buyers, the key test is total cost and reliability on real tasks—especially when long reasoning traces and repeated tool calls are involved. For security organizations eligible for Fairwind, the more consequential announcement may be the controlled access to Flash Cyber and CodeMender for vulnerability detection and remediation. 14