How to install Google says its new Gemini 3.8 Flash model ‘works

What happened

Google launched Gemini 3.8 Flash just a few weeks after its predecessor, and the company is positioning the new model as one that works harder than Gemini 3.7 Flash. That harder work is described as performing more reasoning steps on complex tasks and calling tools iteratively, which is the practical change builders need to plan around when they install or upgrade. The Verge reports the same introductory pricing as Gemini 3.7 Flash, $0.75 per million input tokens and $3.75 per million, while the headline still warns that the model might cost more once those extra steps start running.

This guide is a how-to for moving from Gemini 3.7 Flash to Gemini 3.8 Flash without inventing a new stack. Treat the model name as the switch, treat the published rates as the billing floor, and treat the extra reasoning and iterative tool calling as the load that can raise the bill even when the per-token numbers stay the same. Confirm what you run today, change the model you call, then measure tokens, tool rounds, and cost on the same tasks you already trust.

Google shipped Gemini 3.8 Flash as the successor to Gemini 3.7 Flash. The launch is close in time to that predecessor, close enough that The Verge notes it arrived just a few weeks later. The product claim is not a new price list. It is a behavior claim: Gemini 3.8 Flash works harder than Gemini 3.7 Flash.

Harder, in the language Google used, means more reasoning steps on complex tasks. It also means calling tools iteratively. Those two behaviors are the whole of what shipped beyond the model name itself. If a task is simple, you may not see the extra work. If a task is complex, or if it already depends on tools, you should expect more steps and more tool rounds than you saw on Gemini 3.7 Flash.

The commercial wrapper around that launch is introductory pricing that matches Gemini 3.7 Flash. Input is listed at $0.75 per million tokens, with $3.75 per million as the companion figure The Verge published next to it. That is the rate card you should copy into your upgrade notes. Do not assume a discount. Do not assume a surcharge on the rate card. The surcharge risk is usage, not the published introductory numbers.

How it works

Keep the source of those facts in view while you write runbooks. The Verge is the report this guide follows. Google is the company that launched the model and made the works-harder claim. Gemini 3.8 Flash is the model you install or upgrade to. Gemini 3.7 Flash is the model you are leaving, comparing against, or keeping as a control.

The builder-facing change is the work the model does after you send a request, not a new install ritual with a new version number attached. If you already call Gemini 3.7 Flash, the upgrade is a model swap to Gemini 3.8 Flash. If you do not yet call either Flash model, the install is to start sending work to Gemini 3.8 Flash under the same introductory pricing Google listed for Gemini 3.7 Flash.

What you must redesign is how you budget reasoning and tools. More reasoning steps on complex tasks means a complex prompt can burn more tokens than the same prompt did on Gemini 3.7 Flash. Calling tools iteratively means a tool-using agent can loop more times before it returns. Both of those loops are billable work if you pay per token. Both can also change latency and the shape of traces you already collect.

Builders who ship agents should assume tool calling is no longer a single pass. Iterative calling is the stated behavior. That affects retries, timeouts, and the caps you put on tool use. Builders who ship chat or analysis features should assume complex tasks may take more reasoning steps. That affects max-output settings, stop conditions, and how you explain cost to a team that still sees $0.75 per million input tokens and $3.75 per million on the rate card.

The title of the report is the other builder change. Google says Gemini 3.8 Flash works harder, but it might cost more. Harder work at the same introductory price still costs more when the model spends more tokens. Plan for that in budgets, alerts, and customer quotes before you flip production.

Why it matters

Start by listing every surface that currently names Gemini 3.7 Flash. Include playgrounds, batch jobs, eval harnesses, and production agents. Write down whether each surface sends complex tasks, calls tools, or does both. That list is the install map. You cannot upgrade what you have not named.

Next, keep Gemini 3.7 Flash as a control on a copy of the same tasks. Do not drop the predecessor until you have a side-by-side. Google launched Gemini 3.8 Flash only a few weeks after Gemini 3.7 Flash, so both models are the pair you should compare. Run the same complex tasks through both. Count reasoning-heavy turns. Count how often tools are called more than once. Record tokens in and tokens out against $0.75 per million input tokens and $3.75 per million.

Then point a non-production endpoint at Gemini 3.8 Flash. Leave pricing assumptions at the introductory rates already published for Gemini 3.7 Flash. After each run, ask two questions. Did the model perform more reasoning steps on the complex tasks? Did it call tools iteratively? If yes, the install is doing what Google claimed. If your traces show more steps or more tool rounds, multiply that extra work by the same rates and decide whether the bill is still acceptable.

Promote only the surfaces that stay inside your cost and latency budget. For each promoted surface, change the model name from Gemini 3.7 Flash to Gemini 3.8 Flash, keep the rest of the request path the same, and watch the first live batch for token spikes. If a surface is simple and never calls tools, still run it once through Gemini 3.8 Flash so you know the swap is clean, then move on. If a surface is an agent that already calls tools, raise the visibility of iterative tool calling before you promote. Set a hard stop on how many tool rounds you will pay for, because the model is described as calling tools iteratively and that loop is how a same-price model might cost more.

Who is affected

Finish the upgrade by writing the rate card into the same place you store the model name. Gemini 3.8 Flash. Introductory pricing the same as Gemini 3.7 Flash. $0.75 per million input tokens and $3.75 per million. Review that card after the first production week against actual tokens, not against the claim that the numbers match.

Gotchas and compatibility The first gotcha is price confusion. Same introductory pricing as Gemini 3.7 Flash does not mean the same invoice. Gemini 3.8 Flash is said to work harder. Harder means more reasoning steps on complex tasks and calling tools iteratively. Extra steps and extra tool rounds consume extra tokens. Extra tokens at $0.75 per million input tokens and $3.75 per million still raise the bill. Compatibility with the old rate card is not compatibility with the old spend.

The second gotcha is treating Gemini 3.8 Flash as a drop-in for every Gemini 3.7 Flash call without checking task shape. Simple calls may look compatible. Complex tasks may not, because more reasoning steps can change how long a response takes and how large it becomes. Tool-using flows may not, because iterative tool calling can multiply side effects you only allowed once on Gemini 3.7 Flash.

The third gotcha is upgrading too soon after the predecessor. The new model arrived just a few weeks after Gemini 3.7 Flash. That is a short window to assume every integration, quota, and dashboard already lists Gemini 3.8 Flash. If a console, proxy, or allowlist still names only Gemini 3.7 Flash, the install will fail until that name is added. Add the new name first, then send traffic.

The fourth gotcha is skipping a control. If you replace Gemini 3.7 Flash everywhere on day one, you cannot tell whether a worse answer, a longer trace, or a higher charge came from the new model working harder or from a prompt you already had. Keep Gemini 3.7 Flash available until Gemini 3.8 Flash has proved the same tasks at a cost you accept.

What to watch next

Do not invent a compatibility matrix beyond those two model names. The shipped pair is Gemini 3.7 Flash and Gemini 3.8 Flash. The shipped behaviors are more reasoning steps on complex tasks and iterative tool calling. The shipped rates are the introductory figures above. Anything else is outside this guide.

Watch token counts on complex tasks first. If Gemini 3.8 Flash performs more reasoning steps than Gemini 3.7 Flash on the same prompt, your cost will move even though introductory pricing is the same. Watch tool traces second. If the model calls tools iteratively, count the rounds and the tokens around each round. That is the mechanism behind a model that works harder but might cost more.

Watch invoices against $0.75 per million input tokens and $3.75 per million, not against a hope that matching rates mean matching spend. Watch whether Google keeps that introductory pricing or leaves you on it only for a while. Watch whether surfaces you marked simple stay simple once they sit behind Gemini 3.8 Flash, because a task that starts to look complex will pull in more reasoning steps.

Watch the predecessor as well. Gemini 3.7 Flash is only a few weeks older in this launch story. You may still need it as a cheaper or more predictable control if Gemini 3.8 Flash spends more tokens than the work is worth. Keep the comparison honest: same tasks, same rates, different model names, and a clear record of reasoning steps and iterative tool calls. That record is the upgrade test, and it is the only next step this launch gives you without adding numbers that were not reported.