Gemini Flash: 3.8 Flash and Flash-Lite Specs and Prices

Gemini 3.8 Flash is Google's newest Flash model: $0.75 per million input tokens and $3.75 per million output tokens through 31 December 2026, then $1.50 and $7.50. It reads text, images, video, audio and PDFs, has a 1,048,576-token input limit and writes up to 65,536 tokens. Its cheaper sibling, 3.5 Flash-Lite, costs $0.30 and $2.50. Checked on 31 October 2026.

The Flash line at a glance

Flash is the middle of the Gemini range. Above it sits Gemini 4 Argon and the Pro line. Below it sits Flash-Lite. Google keeps several generations on its models page at once, so the first job is knowing which row you are looking at.

All five are text-output models. None generates audio or images; Google ships separate TTS, Live and Nano Banana models for that, and we list them on the Gemini models page.

What each one is for

Google's own one-line descriptions are the best summary, because they are the vendor's stated intent.

• 3.8 Flash: "our most intelligent Flash model, engineered for long-horizon software engineering, autonomous agents, and complex enterprise workflows," with "the speed and cost efficiency of Flash."
• 3.6 Flash: a previous-generation model that "remains available for production use," described as strong at code generation, agentic execution and spatial reasoning, and "particularly effective for rapid agentic loops involving complex coding cycles."
• 3.5 Flash-Lite: a low-latency, cost-effective multimodal model "for subagent tasks and document parsing."
• 3.1 Flash-Lite: high-frequency, lightweight tasks, with translation named as a lead use case.

Notice the pattern. Google no longer positions Flash as the budget choice. It positions Flash as the model for long agent work and Flash-Lite as the budget choice. That is a change from how many people still think of the name.

Maker, licence and open weights

Made by Google and closed-weight. You cannot download Flash or Flash-Lite. Access is through the Gemini API, Google AI Studio and Google Cloud, under Google's terms. If you need open weights from Google, the answer is Gemma, a different family. For open weights from anyone, start at our open source LLM overview.

Prices, tiers and the date that matters

The date is 1 January 2027. On that day 3.8 Flash and 3.6 Flash both double, from $0.75 and $3.75 to $1.50 and $7.50. Flash-Lite prices are not scheduled to change on the page we read.

After the change, 3.8 Flash Standard is $1.50 / $7.50, Batch and Flex are $0.75 / $3.75, and Priority is $2.70 / $13.50. Cached input on 3.8 Flash is $0.075 per million until 31 December and $0.15 afterwards. On 3.5 Flash-Lite it is $0.03. Explicit caches on Flash models also carry a storage fee per million tokens per hour, which is $0.50 until the end of 2026 and $1.00 after that on 3.8 Flash, and $1.00 on 3.5 Flash-Lite.

There is also a Contributor tier, a private preview offered by invitation only, listed at $0.375 input and $1.875 output for 3.8 Flash, with speed and reliability equal to Standard. We cannot tell you how to get an invitation.

What an agent loop costs

Suppose a subagent makes 200 calls in a session, each with 20,000 input tokens and 1,000 output tokens. That is 4 million input tokens and 0.2 million output tokens. These sums use list prices only.

If caching works well and 80 percent of those input tokens hit the cache, 3.8 Flash input falls to 0.8M at $0.75 plus 3.2M at $0.075, which is $0.84. The session then costs about $1.59 with the output. That ignores any storage fee. Implicit caching is what you get for free on Flash, but how often it hits depends on how stable the start of your prompt is. Keep system prompts and tool definitions at the front and unchanged.

Thinking tokens count as output on every tier, according to the pricing page. So the output column above is a floor for a model set to think a lot.

Thinking and tools

The 3.8 Flash page lists thinking at low, medium and high, and says minimal is not supported and returns an error. This surprises people who set a minimal level on an older Flash model and carry the setting across. The 3.6 Flash and 3.5 Flash-Lite pages list thinking as supported without naming levels.

Supported on 3.8 Flash, per its page: code execution, computer use, file search, function calling, Maps and Search grounding, structured outputs and URL context. Not supported: audio generation, image generation, and the Live API. 3.1 Flash-Lite lacks computer use, so it is a poor fit for browser-driving agents. 3.5 Flash-Lite has it.

Grounding is billed on top. Google gives 5,000 free Search grounding requests a month across Gemini 3 and newer models, then charges $14 per 1,000 requests, capped at 1,500 a day.

Where it is served

Google serves Flash through the Gemini API and Google Cloud. None of our open-weights hosting pages carries it: Runpod, Modal and Lambda rent GPUs for models you can download, and Hugging Face Inference and Replicate do the same for open models.

Gateways that proxy many vendors under one key are a separate route, compared in OpenRouter alternatives.

Coding agents that can use Flash

Gemini CLI and Antigravity are Google's agents. Third-party agents that take a Gemini API key can use Flash too. The Gemini docs also list an OpenAI-compatible endpoint, which helps with agents that only know how to talk to an OpenAI-shaped API.

For subagent work, the 3.5 Flash-Lite description matches what you want a worker to do: parse a document, extract fields, run a narrow check. Pair it with a stronger orchestrator. We explain the pattern in subagents versus skills.

Benchmarks

Google's model pages for Flash link to model cards and do not print scores in the text we fetched. We did not copy any leaderboard numbers. For picking between Flash and a competitor, the honest method is a 20-task run on your own repository, counting passes and total cost.

Privacy

Pricing rows labelled "Used to improve our products" say Yes on the free tier for 3.8 Flash, 3.6 Flash and 3.5 Flash-Lite, and No on the paid tier. That is the whole policy as the pricing page states it. Do not put private code through the free tier. Retention, regional processing and zero-retention terms are not on the pricing page.

Limits and gotchas

• The doubling. 3.8 Flash and 3.6 Flash both double on 1 January 2027.
• No minimal thinking on 3.8 Flash. The call errors.
• 3.1 Flash-Lite has no computer use. 3.5 Flash-Lite does.
• 2.5 Flash is gated. Google limits it to users who already used it, and points new projects to 3.5 Flash-Lite or 3.8 Flash.
• Free tier trains. Google says free-tier content improves its products.
• Output is capped at 65,536 tokens on every Flash model. Long decode, up to 1M output, is listed for Argon only.

Moving between Flash models

Switching is a model-ID change, but the surrounding settings are where problems hide. Walk this list before you flip a production agent.

1. Check the thinking setting. If your code sends a minimal thinking level, 3.8 Flash rejects it. Set low, medium or high instead.
2. Re-read the price row. 3.8 Flash and 3.6 Flash carry the same list price today, so the move from 3.6 to 3.8 is not a savings play. It is a capability play, and Google calls 3.6 Flash a previous generation that stays available for production.
3. Check tool support. If an agent relies on computer use, confirm the target model lists it. 3.8 Flash, 3.6 Flash, 3.5 Flash-Lite and 3 Flash preview do. 3.1 Flash-Lite does not.
4. Watch output limits. All Flash models cap output at 65,536 tokens. A task that used to fit may still fit, but a long generation will not grow just because the window is 1M.
5. Re-measure caching. Implicit cache hits depend on stable prompt prefixes. A model swap does not change that, but a rewritten system prompt will.
6. Pin a date for the price change. Put 1 January 2027 in the calendar and re-run your cost sums with $1.50 and $7.50.

A cheap way to compare two models fairly is to record the same agent session and replay the prompts. Count three things for each model: how many tasks ended with passing tests, how many tokens were billed in total including thinking, and how many calls were retries. The model with the lowest price per token is not always the one with the lowest price per passing task.

Picking a Flash tier for the job

The same model can cost very different amounts depending on how you call it. Match the tier to how urgent the answer is.

Batch and Flex have the same listed price on 3.8 Flash and 3.5 Flash-Lite, so the choice comes down to how you want to submit the work and what latency you can live with. Webhooks and background execution also appear in the Gemini API docs navigation for long jobs.

Rate limits and quotas

Google publishes rate limits on a separate page and ties them to your usage tier. We did not extract numbers from it, so we give none here. If you plan a high-volume run, read that page first, because a cheap model you cannot call fast enough is not cheap.

Our take

Pick 3.8 Flash as your default Google model today, and pick it knowing the price you will pay in 2027. At $1.50 and $7.50 it is still well below the Opus and Fable prices on the Claude models page and the Astra price on the GPT models page, though it is not the cheapest model in its own family. Flash-Lite at $0.30 and $2.50 is the one to evaluate for volume work. Check it against your own failures: if the cheaper model needs two retries to match one Flash call, the saving disappears.

If you are only starting, use the free tier of 3.8 Flash on public code to learn the API and the thinking settings. Move to paid before you point it at anything private.

Who should pick what

Related pages

• Gemini models for the whole lineup
• Gemini Pro for Argon and the Pro line
• Gemini API pricing for every tier
• LLM API prices for a cross-vendor table