Gemini 3.8 Flash vs Gemini 3.1 Pro Preview: When Flash beats Pro

Within a model family, the usual rule is simple: the bigger tier is smarter, and the smaller tier is cheaper. For Google's Gemini models, that has meant reserving Pro for complex reasoning and nuanced analysis, and using Flash models for routine tasks where cost or high volume matters more. The Flash line goes back several versions, including Gemini 2.5 Flash, and it has been Google's lower-cost tier.

The current versions don't follow that rule. Gemini 3.8 Flash ranks #5 in our Ready for Enterprise (RFE) LLM Benchmark, while Gemini 3.1 Pro Preview ranks #20. Flash scores higher on capability and costs less per token, which makes this comparison especially relevant for teams choosing a Gemini model for production AI workloads. Pro Preview still leads in two areas that matter for some teams: factual accuracy and output speed.

Here's how the two models compare on our Gemini Flash vs Pro data, including benchmark performance and the price difference, and how to decide which model choice fits each task type.

Gemini 3.8 Flash vs Gemini 3.1 Pro Preview: Benchmark comparison

Source: FullStack RFE LLM Benchmark. Rankings can change as new models ship and results are added. Benchmark listing prices come from a provider listing and can differ slightly from Google's list price.

Our RFE score combines six independent benchmarks with price and speed, and the capability score reflects benchmark performance alone. Gemini 3.1 Pro Preview hasn't been tested on Terminal-Bench 4.0 yet, so its scores are based on five of the six benchmarks. It also falls below our capability threshold, so it ranks after every model that reaches it. We show SWE-Bench Pro for reference only and don't build any comparisons on it.

Where does Gemini 3.8 Flash beat Pro?

Flash leads on four of the five benchmarks both models were tested on. The gaps are uneven, so it's worth looking at them one at a time.

Web development and math show the biggest gaps

The largest lead is on Arena WebDev, where people compare models on building web apps. Flash rates 1585 against 1447 for Pro Preview. Flash also leads on FrontierMath, an advanced math test, at 68.4% versus 59.6%.

That fits how Google describes the model. Its changelog calls Gemini 3.8 Flash "our most intelligent Flash model, engineered for long-horizon software engineering." Flash also scores 19.1% on Terminal-Bench 4.0, which tests agentic tasks like writing code and using tools from the command line. Pro Preview hasn't been tested there yet, so we can't compare the two on agentic workflows from our data.

Reasoning and general chat are close

On GPQA Diamond, a graduate-level science test that calls for deep reasoning, the two models are within one point (95.4% versus 94.4%). On Arena Text, which measures human preference in general conversation, they're within nine points (1496 versus 1487). For general reasoning and writing, you're unlikely to see much quality difference, so reasoning depth alone isn't a strong reason to pay for Pro.

Where does Gemini 3.1 Pro Preview still lead?

Pro Preview's clear lead is factual accuracy. On SimpleQA Verified, which tests whether a model answers short factual questions correctly, it scores 73.5% against 69.7% for Flash.

That gap matters most when the model answers from its own knowledge, such as a support assistant asked about product details or an internal tool asked for dates and names. If your workload depends on the model recalling facts correctly, test both models on your own questions before switching. Where accuracy is non-negotiable, such as research summaries that users will act on, Pro Preview's edge on factual recall may matter more than the lower cost of Flash.

A common way to reduce that risk is to stop relying on the model's memory. Retrieval-augmented generation (RAG) gives the model the right documents at the time of the question, so the answer comes from your sources rather than from what the model learned in training. With good retrieval, the gap between the two models on recall matters less.

^{For more on grounding answers in your own data, see our Enterprise Knowledge Management services.}

Which model is better for coding work?

For most engineering work, start with Flash. It leads on Arena WebDev, and Google built it for long-horizon software engineering. For writing code and code generation inside an existing system, from a first implementation to a bug fix, Flash is a reasonable default.

Google positions Gemini 3.1 Pro Preview for software engineering too. Its model page says Pro Preview is built for ‘precise tool usage and reliable multi-step execution across real-world domains. If Pro handles a specific job better in your tests, like complex debugging across a large codebase, it's fine to keep it there.

For code review on pull requests, the choice depends on how careful the review needs to be. Flash can handle a first pass on high-volume pull requests at a lower token cost. For changes where edge cases matter, like payment logic or security-sensitive code, test whether Pro Preview catches more issues before you switch. Either way, a person should still do the careful review on high-risk changes.

How much cheaper is Gemini 3.8 Flash?

The answer depends on when you're reading this, because Google has announced a change to Flash pricing.

On Google's Gemini API pricing page, Gemini 3.8 Flash costs $0.75 per million input tokens and $3.75 per million output tokens through December 31, 2026. From January 1, 2027, the listed rates double to $1.50 input and $7.50 output. For Pro pricing, Gemini 3.1 Pro Preview costs $2 input and $12 output per million tokens for prompts up to 200,000 tokens, and $4 input and $18 output above that.

Pro Preview prices shown are for prompts up to 200,000 tokens. Source: Google Gemini API pricing page.

Using our 3:1 input-to-output blend, Flash costs about a third of Pro Preview's price today. After the January change, it costs about two-thirds. Flash is still cheaper and still scores higher on our benchmark, but if you're building a budget for 2027, use the new rates.

Both models bill thinking tokens at the output token rate, so the price per token is only part of the cost. A model that thinks longer on a task can cost more per completed task even at a lower rate. For real cost efficiency, compare the cost of a full run on your own workload, not just the rate card. Budget efficiency comes from the total, not the list price.

^{For help bringing model costs down, see our AI Cost Optimization services.}

Is Gemini 3.8 Flash faster than Pro?

Not in our measurements, so Flash's speed isn't an advantage here. Gemini 3.1 Pro Preview generates 90 tokens per second, compared with 76 for Gemini 3.8 Flash. If raw output speed is your main concern, Flash isn't the automatic pick, and you shouldn't assume Flash means lower latency.

The difference is modest. For a 1,000-token answer, it works out to about 11 seconds of output time on Pro Preview versus about 13 on Flash, before any thinking time. Real response times also depend on how long each model thinks and on server load. We don't score time to first token, because it's unreliable for reasoning models, so neither model has a clear speed advantage you can count on. Pro Preview isn't slow, and Flash isn't always faster.

What else is different between the two models?

The biggest practical difference is release status. Google released Gemini 3.1 Pro Preview on February 19, 2026, and it's still labeled a preview. Gemini 3.8 Flash became generally available on September 2, 2026. Preview models can change before a stable release. For production work, a generally available model is usually the safer default.

Both models accept up to about 1 million input tokens, so neither has a larger context window. Context length isn't a reason to choose one over the other. Flash is also available on the Gemini API free tier, while Pro Preview is not, which makes Flash easier to test in Google AI Studio before you commit.

Should you switch from Gemini Pro to Flash?

For most workloads, our data says yes. Gemini 3.8 Flash scores higher overall and costs less, and it's generally available. Here's a simple way to pick the right model, with the trade-offs to weigh:

• Move to Flash for web development and math, where Flash leads, and for general reasoning, where the two are close.
• Use Flash for high-volume work, like pulling structured data out of documents or answering support questions with retrieval.
• Test both first for fact-retrieval tasks where the model answers from memory, since Pro Preview leads on SimpleQA Verified.
• Add retrieval if factual accuracy is the only thing holding you on Pro. Grounding answers in your own documents can make Flash a safe choice for those tasks too.

You don't have to use one model for everything. Running two models, with Flash as the default and Pro Preview for the few complex tasks where it tests better, can be the better choice. Whichever you choose, keep the model name in configuration rather than hard-coded. New Flash versions can arrive within weeks of each other, so you'll want to be able to retest on your real work without rewriting code.

How FullStack can help

We help teams test models like Gemini Flash and Pro against their real workloads, route each task to the model that fits it, and keep AI spending under control as prices change. When factual accuracy matters, we build retrieval systems that ground answers in your own data.

^{Compare these models in the RFE benchmark.}