Gemini 3.6 Flash (2026): Pricing, Benchmarks, API, and What Changed

Gemini 3.6 Flash is Google's July 2026 Flash model for developers who need strong reasoning without moving to a heavier Pro model. It is aimed at coding, tool use, long context, and mixed media inputs.

Google says it uses fewer output tokens than Gemini 3.5 Flash and improves several coding and computer-use benchmarks.

Google released Gemini 3.6 Flash on July 21, 2026. The stable API model ID is gemini-3.6-flash.

It supports up to 1,048,576 input tokens and 65,536 output tokens. Google's Gemini 3.6 Flash model page lists text, images, video, audio, and PDF as inputs, with text as the output.

Current status: Google released Gemini 3.7 Flash on August 13, 2026. Gemini 3.6 Flash is no longer the newest Flash model, but Google still lists it as a stable model in the Gemini API release notes.

What is Gemini 3.6 Flash?

Gemini 3.6 Flash is built for developers who want one model to handle code, documents, images, audio, video, and tool-driven workflows. Google built it on Gemini 3.5 Flash and focused on better coding, knowledge work, visual tasks, and token use.

Google's Gemini 3.6 Flash model card calls it a workhorse model. In practice, that means jobs such as reviewing a large codebase, extracting details from long PDFs, reading charts, or running an agent that has to call several tools before it can answer.

The best model still depends on the job. A high-volume agent may care most about speed and cost.

A harder reasoning task may need a different balance. Chatbase's overview of choosing the right AI model explains that broader choice.

Gemini 3.6 Flash release date and availability

Gemini 3.6 Flash launched on July 21, 2026 as a stable, production-ready model.

Google currently lists access through:

• Gemini API: use the stable model ID gemini-3.6-flash.
• Google AI Studio: test prompts and build with the API.
• Gemini app: listed by Google DeepMind as a model distribution channel.
• Google Cloud: generally available in Gemini Enterprise Agent Platform, with global, US, and EU regions.
• Google Antigravity: listed as a distribution channel for agent development.

Google's Gemini 3.6 Flash launch post covers the original rollout and the gains Google reported at launch.

Gemini 3.6 Flash specs at a glance

Also supported: code execution, Google Search grounding, Google Maps grounding, file search, URL context, and context caching.

Not supported: Live API, native image generation, or native audio generation.

The 1M-token context window can hold long files, code, transcripts, and mixed media. You still should not send everything by default.

For repeated production work, a smaller set of relevant context is often cheaper and easier for the model to use well.

Gemini 3.6 Flash pricing

Gemini 3.6 Flash is cheaper than its original launch price for the rest of 2026. Google's Gemini API pricing page lists these temporary Standard rates through December 31, 2026:

• Input: $0.75 per 1 million tokens
• Output: $3.75 per 1 million tokens, including thinking tokens
• Context caching: $0.075 per 1 million tokens
• Cache storage: $0.50 per 1 million tokens per hour
• Free tier: available for Standard usage

On January 1, 2027, Google says Standard pricing will rise to $1.50 per 1 million input tokens and $7.50 per 1 million output tokens.

Google also lists lower temporary rates for Batch and Flex. Through December 31, 2026, both cost $0.375 per 1 million input tokens and $1.875 per 1 million output tokens.

What thinking tokens cost

Gemini 3.6 Flash uses dynamic thinking by default. The default level is medium. Developers can choose minimal, low, medium, or high.

Google's thinking documentation says thinking tokens count toward output billing. More thinking can help on hard tasks, but it can also raise cost and make the user wait longer.

For a fast user-facing task, minimal, low, or medium may be a better place to start. High thinking is more useful when the task is hard enough to need the extra work.

Gemini 3.6 Flash benchmarks

Google reports gains over Gemini 3.5 Flash on coding, computer use, knowledge work, and long-context tests.

SWE-Bench Pro, DeepSWE, and MLE-Bench focus on coding or ML work. OSWorld-Verified tests computer use. GDPVal-AA looks at knowledge work, while GDM-MRCR tests long-context retrieval and reasoning.

These are Google-reported results, so treat them as signals rather than a guarantee for your own app. The biggest gains here matter most if your workload looks like the benchmark being tested.

Google also reported 17% fewer output tokens than Gemini 3.5 Flash on the Artificial Analysis Index, which may help in long or tool-heavy runs where output grows quickly.

How fast is Gemini 3.6 Flash?

Artificial Analysis reports an Intelligence Index score of 52 for the high-thinking version it tested. It currently reports about 225 output tokens per second for the high-thinking configuration it tested.

That speed number starts after output begins. Thinking time and tool calls can add delay before the answer appears. If latency matters, test the full response time with the same prompts, tools, and thinking level used in production.

Gemini 3.6 Flash vs Gemini 3.5 Flash

Gemini 3.6 Flash builds on Gemini 3.5 Flash. That makes 3.5 the most useful comparison for teams planning an upgrade.

The main changes are:

• Coding: 3.6 scores higher on Google's DeepSWE, SWE-Bench Pro, and MLE-Bench results.
• Agent workflows: Google reports fewer reasoning steps and tool calls on some multi-step tasks.
• Token use: Google measured 17% fewer output tokens on the Artificial Analysis Index.
• Computer use: OSWorld-Verified rose from 78.4% to 83.0% in Google's tests.
• Long context: the 1M-token GDM-MRCR result rose from 26.6% to 54.0%.
• Price: 3.6 is $0.75 input and $3.75 output per million tokens through December 31, 2026. Gemini 3.5 Flash is $1.50 input and $9 output.
• Context size: both support very large contexts, so the upgrade is more about behavior, cost, tools, and quality.

Should you upgrade from Gemini 3.5 Flash?

Gemini 3.6 Flash is worth testing if your app spends time editing code, calling several tools, reading mixed media, or working with long context. The current lower price can also matter when those requests run at scale.

Do not swap models just because the benchmark table looks better. Run the same real prompts, tool calls, edge cases, and latency checks you use in production before moving traffic.

What developers need to change when moving from Gemini 3.5 Flash

Some Gemini 3.x settings can break older request setups. Google's Gemini 3.x migration guidance lists several changes that apply across the 3.x family.

Check these before switching:

Change the model ID: use gemini-3.6-flash.

Remove custom sampling values: Google says custom temperature, top_p, and top_k values are ignored on Gemini 3.6 Flash, so remove them from the request. Check penalties: Custom frequency and presence penalty values are not supported. Setting them throws an error.

Replace thinking_budget: current Gemini 3.x guidance uses thinking_level.

Remove candidate_count: Gemini 3.x does not support it.

Remove prefilled model turns: requests cannot end with a model role.

Retest function calls: check schemas, multi-turn state, and function responses with real app flows.

Google's Cloud documentation for Gemini 3.6 Flash explains the unsupported settings and turn rules in more detail.

Gemini 3.6 Flash API: model ID, setup, and tools

The API model ID is:

gemini-3.6-flash

Gemini 3.6 Flash works with Google's Interactions API and the generateContent API. A basic Python call with Google's Gen AI SDK looks like this:

from google import genai

client = genai.Client()

response = client.models.generate_content(

model="gemini-3.6-flash",

contents="Summarize the main risks in this deployment log."

)

print(response.text)

Google's GenerateContent API reference uses the same stable model ID in current examples.

Tools and structured outputs

Gemini 3.6 Flash supports:

• function calling
• structured outputs
• code execution
• Google Search grounding
• Google Maps grounding
• file search
• URL context
• context caching
• computer use in preview

Function calling is useful when the model needs to do something instead of only answer. An app can expose approved actions, while structured output keeps the result in a format downstream code can use.

Search and file tools can add fresh or private context. Computer use needs more care because Google still marks it as preview, so permissions, logs, and approval steps matter for risky changes.

Gemini 3.6 Flash rate limits

Gemini 3.6 Flash does not have one rate limit for every project. Google's Gemini API rate-limit documentation ties limits to the model and the project's usage tier.

Batch work has separate queue limits. Interactive traffic uses different limits. Check the live limits for the project in AI Studio before planning capacity.

What Gemini 3.6 Flash is built for

Coding and software engineering

Coding is one of the clearer reasons to test 3.6 against 3.5 Flash. Google reports better results on long-horizon software work and ML engineering.

For a developer, the useful question is not only whether it can write a function. It can also help review a repo, migrate code, refactor a feature, debug failures, and work through repeated edit-test loops. Frontend work still needs visual QA because valid code can still miss the intended design.

Agentic workflows and tool use

Gemini 3.6 Flash can reason, call a tool, read the result, and continue the task. That is useful for agents that need to look something up, take an action, and then decide what to do next.

Google says it uses fewer reasoning steps and tool calls than 3.5 Flash on some multi-step tasks. You still need to limit the tools it can reach, validate arguments, set stop rules, and log actions.

Payments, large refunds, account deletion, and other hard-to-reverse actions should use extra checks or human approval.

Multimodal and spatial reasoning

Gemini 3.6 Flash accepts text, images, video, audio, and PDFs. That matters when the useful information is spread across formats.

A developer could use it to read a PDF with charts, compare a screenshot with written rules, or summarize a video alongside text notes.

Google also highlights spatial reasoning and visual blueprint conversion. The model's output is still text. It does not provide native image or audio generation.

Gemini 3.6 Flash limitations and production trade-offs

The strong benchmark numbers do not make Gemini 3.6 Flash the right model for every production task.

• It can hallucinate: Google's model card lists hallucinations as a known foundation-model limit. Grounding can help, but important facts still need checks.
• More thinking can cost more and take longer: hard tasks may need it, while simple live tasks may not.
• A 1M-token window can be wasteful: sending too much context can raise cost and hide the useful details.
• Computer use is preview: browser and interface actions need tighter controls.
• Live API is not supported: choose another model if native real-time Live API behavior is required.
• Native image and audio output are not supported: use a model made for those output types when needed.
• Slow responses and timeouts can happen: Google's model card says these may occur, so production apps need retries, timeouts, and fallback plans.

What Gemini 3.6 Flash should not automate on its own

Do not give the model sole control over high-stakes or hard-to-reverse decisions. Examples include large refunds, payments, account deletion, and medical or legal decisions.

Use clear permissions, tool limits, logs, and approval steps. The model can prepare or route the work without being the final authority.

Where Gemini 3.6 Flash fits in the current Gemini lineup

Gemini 3.6 Flash is still stable, but it is no longer Google's newest Flash model.

Google released Gemini 3.7 Flash on August 13, 2026. Google now positions 3.7 as its newer workhorse for coding and agents. The current Gemini model update also applies the new temporary 2026 pricing to 3.6 Flash.

A simple way to think about the current options:

• Gemini 3.7 Flash: test first for new coding and agent-heavy work where you want the newest Flash model.
• Gemini 3.6 Flash: useful for stable deployments already tested on it, or while a team tests 3.7.
• Gemini 3.5 Flash: the direct predecessor and the main migration baseline.
• Gemini 3.5 Flash-Lite: a lower-cost choice for high-volume tasks that need less reasoning.

A broader ChatGPT, Claude, and Gemini comparison makes more sense when the choice is between model families.

Using Gemini 3.6 Flash in an AI agent stack

Gemini 3.6 Flash can supply the reasoning and tool-use layer of an AI agent. It can understand a request, read business context, and decide which approved tool to call.

That alone does not make a production agent. The surrounding system still has to control what data the model can use, what actions it can take, where it can talk to users, and when a person should step in.

Chatbase currently lists Gemini 3.6 Flash as a selectable model for its AI agents at two message credits per response in the Chatbase model settings documentation.

What Chatbase adds around Gemini 3.6 Flash

Using Gemini 3.6 Flash through Chatbase adds the layer around the model that a customer-facing agent needs:

• Grounded knowledge: give the agent approved website content, files, and business data.
• Actions and workflows: let the agent use approved tools when the right integrations and permissions are set.
• Multiple channels: run the same agent across website chat, email, messaging, and voice.
• Human handoff: route uncertain, sensitive, or approval-heavy cases to a person.
• Conversation oversight: review chats and use analytics to find weak answers, missing knowledge, or poor instructions.

That is the part model benchmarks cannot measure. A model may reason well in a test and still give a poor customer answer if it has stale product data, the wrong policy, or too much freedom to act.

For example, Gemini 3.6 Flash can understand a request and choose an approved function. Chatbase brings that model into AI customer service by adding the knowledge, tools, channels, and workflows needed for a real support process.

Some work should still go to a person. Large refunds, account changes, unusual billing disputes, or unclear cases can use human handoff workflows instead of letting the model make the final call.

Teams that want to test this setup can build an AI agent and compare Gemini 3.6 Flash with real questions, tool calls, handoffs, latency, and failure cases.

Is Gemini 3.6 Flash still worth using?

Yes, for the right workload.

Gemini 3.6 Flash is still a practical choice when an app already depends on it or when its mix of price, long context, mixed media input, and tool use fits the workload. It also beats 3.5 Flash on several Google-reported coding and agent benchmarks.

For a new build, test Gemini 3.7 Flash too because it is now Google's newer Flash model. For an existing 3.5 app, 3.6 is still a useful upgrade candidate when coding, tool use, long context, or token cost matters.

Test answer quality, tool accuracy, latency, token use, and failure cases together. One benchmark win does not decide the best model for a product.

Frequently asked questions

Is Gemini 3.6 Flash free?

Yes. Google lists free Standard usage for Gemini 3.6 Flash.

Paid Standard usage is $0.75 per 1 million input tokens and $3.75 per 1 million output tokens through December 31, 2026. Usage and rate limits still depend on the project and tier.

What is the Gemini 3.6 Flash context window?

Gemini 3.6 Flash supports up to 1,048,576 input tokens and 65,536 output tokens. It accepts text, images, video, audio, and PDFs.