Gemini 3.5 Flash Capabilities: Limits, API Access, and the 3.6 Replacement
Gemini 3.5 Flash supports coding, document analysis, and tool workflows with text output. In Google's Developer API, its old model ID now automatically serves Gemini 3.6 Flash; Cloud has a separate lifecycle.
On this page

Gemini 3.5 Flash can work with text, images, video, audio, and PDFs, then produce text answers, code, or structured data. Its documented strengths make coding assistants, document analysis, and multi-step tool workflows sensible uses to evaluate. It does not generate images or audio, and it does not support the Live API. Those distinctions remain important even when a model can understand a screenshot or recording. Google's model specifications list the supported inputs, outputs, and tools.
There is also a change that affects every existing Developer API integration: since Google's October 8, 2026 announcement, requests for gemini-3.5-flash automatically route to gemini-3.6-flash. The endpoint is not shutting down, but the old string no longer selects the original 3.5 model there. Use gemini-3.6-flash explicitly for that replacement, or evaluate 3.8 Flash for a new demanding workflow. Google Cloud's 3.5 Flash offering has its own support schedule. Developer API deprecations and the Cloud model page describe these separate cases.
This guide uses Google's documentation checked on October 14, 2026. Capability and migration guidance below comes from those documents; we have not run model calls, latency comparisons, or billing tests.
Which model does a Gemini 3.5 Flash configuration use now?
Check the provider and endpoint before changing the model name. The same product label can appear in configurations with different lifecycles.
| Where you use it | Current documented status | What to do |
|---|---|---|
| Gemini Developer API | gemini-3.5-flash is deprecated and automatically routes to gemini-3.6-flash; Google says the endpoint is not shutting down. | Make the replacement explicit and review 3.6's parameter changes. |
| Google AI Studio | 3.5 Flash is no longer available to select. | Select an active model such as 3.6 or 3.8 Flash. |
| Gemini Enterprise Agent Platform in Google Cloud | The Cloud card still lists gemini-3.5-flash as GA, with retirement on May 19, 2027 or later. | Follow that Cloud model's region, authentication, API support, and lifecycle documentation. |
| A third-party provider | Google's direct API policy does not establish what that provider serves. | Check the provider's mapping and returned model information. |
The first two rows follow the October 8 changelog and deprecation table. The Cloud date is specifically a retirement date for Cloud support, as the Cloud card's footnote explains. It is not an extension of the Developer API's original 3.5 runtime.
For native Developer API responses, record the requested model alongside modelVersion, which Google defines as the version used to generate the response. Also keep the response ID, date, consumption mode, and usage metadata. A configuration string alone cannot show what generated an answer. These fields are documented in the GenerateContent response reference.
If you need Cloud identity, IAM, regional controls, or centralized governance, the Gemini API versus Vertex AI API guide explains the operational choice. Moving to that route involves more than replacing a model string.
What Gemini 3.5 Flash can do
The archived 3.5 card and current 3.6 replacement card both list a 1,048,576-token input limit and a 65,536-token output limit. A large input window lets you provide more material; it does not guarantee that every detail will be recalled correctly or allow a million-token answer. Both cards specify text output. 3.5 specifications, 3.6 specifications.

May 2026 illustration retained for context. Current capabilities, including Computer Use Preview, are described below.
Coding and debugging
Coding is a reasonable first evaluation when your assistant must inspect several files, explain a failure, propose a change, and work with tools. The documented combination of thinking, function calling, code execution, and structured output supports that kind of application. Google describes the original 3.5 release as optimized for iterative coding and agentic loops in its 3.5 developer guide.
For the currently served 3.6 replacement, Google reports fewer unnecessary edits, debugging loops, tool calls, and conversational turns than 3.5 in its own assessment. That is a reason to evaluate it on repository tasks, rather than proof that it will fix every bug faster. The same 3.6 guide says human evaluators sometimes preferred earlier models' visual styling. Give a frontend task explicit layout and design requirements, then judge the working result as well as its appearance.
A practical trial is a bug with a known reproduction and an existing regression check. Ask for the cause and a bounded change, then assess whether the proposed code resolves the failure without touching unrelated behavior. The model's coding ability and your application's permission to edit or execute files are separate decisions.
PDFs, screenshots, recordings, and other documents
Multimodal input is useful when the evidence is not all plain text: a PDF with tables, a screenshot of a broken layout, or an audio recording that needs a written summary. The result still comes back as text. For extraction, ask for source locations and an explicit representation of missing information, then validate required fields in your application. For a document question, preserve enough original context to check the answer against the material.
The input modalities are documented capabilities, not evidence of perfect OCR, transcription, or table extraction. A long context window also does not resolve contradictions between documents. Test the documents that matter to your product, especially cases with unclear scans, inconsistent values, or facts spread across multiple files. Google's 3.6 model card provides the relevant input and token limits.
Structured answers, search, and tool workflows
The model cards list structured outputs, function calling, code execution, file search, URL context, caching, and grounding with Google Search and Google Maps. These let a backend combine model reasoning with retrieved information and constrained output. A structured response still needs business validation: a valid JSON object can contain the wrong amount, date, or classification. Supported tools and features.
For custom functions, your application receives a requested call, checks it, executes the permitted operation, and returns the result. In a multi-step conversation, preserve the full history and thought signatures. Google's 3.5 guide requires matching each function response to the preceding call's ID and name, with exactly one response per call. Put returned images or other multimodal assets inside the function response. Dropping signatures or mismatching results can break a workflow even when a simple chat request works.
Can it control a browser or desktop?
The current 3.5 and 3.6 model cards list Computer Use as supported in Preview. Google's current Computer Use guide recommends 3.8 Flash and lists 3.8 Flash, 4 Argon, 3.5 Flash-Lite, and 3 Flash Preview in its model-version section; it does not list 3.6 there. The cards alone do not establish that a particular 3.6 tool request will work unchanged.
For a new UI automation implementation, start with the guide's documented 3.8 path. You must supply an execution environment and a loop: send the current screenshot and tool configuration, receive a proposed action, execute it only when allowed or confirmed, and return the updated state. Halt blocked actions. A text request without that handler does not click anything. Computer Use remains a Preview capability, and the guide calls for supervision and a restricted execution environment.
A first native API request for the current replacement
Use gemini-3.6-flash explicitly to try the current replacement through the Gemini Developer API. The native endpoint is https://generativelanguage.googleapis.com/v1beta/models/{model}:generateContent, with contents holding the user message. It is distinct from the Interactions API and from Cloud's project/location endpoints. GenerateContent reference.
The example below assumes a terminal with curl and jq, and an existing permitted API credential already supplied as GEMINI_API_KEY. Google now creates auth keys by default in AI Studio; unrestricted standard keys are rejected, while explicitly restricted standard keys continue to work. Access also requires the applicable account, age, and supported-region conditions. See Google's API key guidance.
This is a documentation-based example, not a recorded successful call:
curl --fail-with-body --silent --show-error \
"https://generativelanguage.googleapis.com/v1beta/models/gemini-3.6-flash:generateContent" \
-H "x-goog-api-key: ${GEMINI_API_KEY:?Supply an existing permitted API key}" \
-H "Content-Type: application/json" \
--data-binary @- \
--output flash-response.json <<'JSON'
{
"contents": [{
"role": "user",
"parts": [{
"text": "A JavaScript function uses items.sort() and unexpectedly changes the caller's array. Explain why, show a version that returns a sorted copy, and give a test that checks the original array remains unchanged."
}]
}],
"generationConfig": {
"maxOutputTokens": 4096,
"thinkingConfig": {"thinkingLevel": "MEDIUM"}
}
}
JSONAfter a successful HTTP response, inspect the result:
jq '{modelVersion, responseId, usageMetadata,
finishReason: .candidates[0].finishReason,
promptFeedback}' flash-response.json
jq -r '.candidates[]?.content.parts[]? | select(.text != null) | .text' \
flash-response.jsonSuccess for this task means a response with usable text explaining mutation, a copy-based implementation, and a relevant test. It also means your logs preserve the returned model version and usage. If the request fails, read its error body before parsing it as a model answer. If there are no candidates, inspect promptFeedback; for a candidate that stops unexpectedly, inspect finishReason. These response fields are documented in the API reference.
Native REST uses generationConfig.thinkingConfig.thinkingLevel with uppercase enum values such as MEDIUM. The Interactions API uses /v1beta/interactions, model, input, and lower-case generation_config.thinking_level values such as medium; its examples expose output_text. Do not paste fields from one API into the other. Native thinking configuration, Interactions examples in the 3.6 guide.
Thinking levels and migration changes that affect results
Start a 3.6 evaluation at its default medium thinking level. Try low for simpler jobs where latency matters, and high when a difficult reasoning or coding task needs more effort. Minimal is available on 3.6, but does not guarantee that thinking is completely off. If you move to 3.8, use low, medium, or high: minimal is unsupported and returns an error. Thinking-level table, 3.8 model specifications.
Before promoting an existing integration, review these documented changes:
- Remove
temperature,top_p, andtop_koverrides (temperature,topP, andtopKin native REST). The 3.6 guide says these are deprecated and ignored for these models, and supplying them can produce errors in future model generations. - Remove
candidate_count(candidateCountin native REST); it is unsupported in Gemini 3 and later. Do not copy it simply because the general API reference still exposes a candidate-count field. - Replace numeric
thinking_budgetsettings with a supported thinking level. Do not send both. - Remove a prefilled final model turn. A request ending with a non-empty model turn returns a 400 error; express formatting requirements in instructions or structured-output configuration instead.
- Preserve conversation history, thought signatures, and correctly matched function responses. In stateful Interactions requests,
store: trueplusprevious_interaction_idlets the server manage state. In stateless operation, resend the thought blocks unchanged.
These rules come from the 3.6 migration guide, 3.5 function-response guidance, and thought-signature documentation.
Keep output limits separate from thinking effort. The thinking guide states that Interactions max_output_tokens covers both thoughts and output; a small cap can stop reasoning with incomplete or empty output while still charging for generated thoughts. Lowering thinking effort is a different control from truncating generation. The Flash model cards still specify a 65,536-token output limit.
Current API cost and capacity
The current Developer API pricing page lists 3.6 and 3.8 Flash at the same token rates below. These are USD per million tokens, with thinking tokens included in output charges. The temporary rates apply through December 31, 2026; the published rates double on January 1, 2027. They replace the original 3.5 price assumptions for a new direct-API budget. They are not an observed invoice for an old alias or a third-party provider. Google's current pricing.
| Consumption mode | Input / output through Dec. 31, 2026 | Input / output from Jan. 1, 2027 |
|---|---|---|
| Standard | $0.75 / $3.75 | $1.50 / $7.50 |
| Batch | $0.375 / $1.875 | $0.75 / $3.75 |
| Flex | $0.375 / $1.875 | $0.75 / $3.75 |
| Priority | $1.35 / $6.75 | $2.70 / $13.50 |
The 3.6 guide still quotes the undated $1.50 input and $7.50 output base rates. For the period above, use the dated pricing tables, rather than interpreting that guide as evidence that the temporary rate has ended.
As a hypothetical Standard budget, suppose 1,000 tasks consume 10 million uncached input tokens, 800,000 answer tokens, and 1.2 million thinking tokens. The token charge is 10 × $0.75 + (0.8 + 1.2) × $3.75 = $15 through December 31. At the January rates, the same usage would be $30. This excludes tools, retries, caching, storage, and taxes; it is a calculation, not a measured bill.
For Standard, cached input is currently $0.075 per million tokens and cache storage is $0.50 per million token-hours, rising to $0.15 and $1.00 respectively on January 1. Paid Search and Maps grounding have shared allowances and separate charges; one user request can generate several billable Search queries. Add those costs when your workflow uses them. Caching and grounding prices.
The Standard pricing table lists a free tier for 3.6 and 3.8, but there is no universal request allowance to copy into your product. Limits apply per project, not per API key, and cover RPM, input TPM, and RPD. RPD resets at midnight Pacific time. Read your project's active limits in AI Studio; Google's rate-limit guide says actual capacity can vary. A 429 can also reflect a conditional rolling spend limit, so creating another key in the same project does not solve the capacity problem.
For sensitive code or documents, pricing is only part of the decision. Google's terms allow unpaid-service inputs and responses to be used for product improvement and human review. The terms apply paid-service data-use conditions to users in the EEA, Switzerland, and the UK even for unpaid quota, and require Paid Services when making API clients available to users in those markets. Paid-service content is not used to improve products, but limited safety logging still applies. Check the applicable conditions before supplying private material.
Keep the current workflow, or move to another model?
Keep a working integration on its documented 3.6 replacement when its accepted results and operational behavior still meet your requirements. Updating the string from 3.5 to 3.6 makes the configuration honest; it does not itself demonstrate a quality improvement.

Original launch-era migration illustration; use the current model IDs and migration conditions in this section.
For a new demanding coding or multi-step agent, evaluate 3.8 Flash as well. Google positions it for long-horizon software engineering and complex workflows in its current model card. For routine extraction, translation, classification, or short summaries, evaluate a cheaper Flash-Lite option before paying for full Flash reasoning. The Flash versus Flash-Lite comparison focuses on that cost-versus-result choice.
Compare genuinely different served models on representative work. Calling the old 3.5 Developer API alias and explicit 3.6 is not an independent original-3.5 versus new-3.6 benchmark. For each candidate, keep inputs, tools, and acceptance requirements comparable, then record accepted results, schema failures, tool errors, retries, latency, usage, and human repair time. Promote the workloads that improve, rather than assuming a single winner for every job.
Keep a fallback that you have actually verified remains available. Changing the Developer API string back to gemini-3.5-flash will not restore the original 3.5 runtime. And if the product requires generated images, speech output, or real-time voice, choose a model and API designed for that output or session type instead of treating text-output Flash as a universal upgrade.
FAQ
What is the Gemini 3.5 Flash output token limit?
The 3.5 and 3.6 Developer API cards list a maximum of 65,536 output tokens, with 1,048,576 input tokens. These are separate limits, not a promise of a million-token answer. See the 3.5 card and 3.6 replacement card.
Can Gemini 3.5 Flash generate images?
No. Its documented image support is for input; output is text. The current 3.6 replacement also lists image generation and audio generation as unsupported. Use a dedicated image-output model for drawing or editing images. Model capabilities.
Does it support the Live API?
No. Both the Developer API card and Cloud 3.5 card list Live as unsupported. Understanding an uploaded audio file is different from running a live voice session.
Is Gemini 3.5 Flash-Lite the same model as 3.5 Flash?
No. Flash-Lite is a separate lower-cost model. The Developer API's gemini-3.5-flash alias routes to 3.6 Flash, not 3.5 Flash-Lite. Google's 3.6 and 3.5 Flash-Lite guide describes their different roles.
Is 3.6 Flash faster or better than 3.5 Flash?
Google reports fewer turns, tokens, and tool calls in its 3.6 assessment. That does not establish a universal speed multiplier or a winner on your tasks. In the Developer API, the old 3.5 alias now serves 3.6, so those two request strings cannot recreate a comparison with the original 3.5 model.
Should I replace a Pro model with Flash?
Evaluate Flash where iteration speed, tool use, or cost matters, and keep Pro where it produces more reliable accepted results on difficult work. The documentation does not prove that Flash replaces every Pro workload. Compare the current served models on your own tasks, including the cost of retries and review; do not use a launch-era model name as the baseline.
Sources17
External pages this guide links to, in the order they appear. Last updated Oct 14, 2026.
Sources17
External pages this guide links to, in the order they appear. Last updated Oct 14, 2026.
- 1.Google's model specificationsai.google.dev/gemini-api/docs/models/gemini-3.5-flash
- 2.Developer API deprecationsai.google.dev/gemini-api/docs/deprecations
- 3.Cloud model pagedocs.cloud.google.com/gemini-enterprise-agent-platform/models/gemini/3-5-flash
- 4.October 8 changelogai.google.dev/gemini-api/docs/changelog
- 5.GenerateContent response referenceai.google.dev/api/generate-content
- 6.3.6 specificationsai.google.dev/gemini-api/docs/models/gemini-3.6-flash
- 7.3.5 developer guideai.google.dev/gemini-api/docs/whats-new-gemini-3.5
- 8.3.6 guideai.google.dev/gemini-api/docs/whats-new-gemini-3.6
- 9.Computer Use guideai.google.dev/gemini-api/docs/computer-use
- 10.supported-region conditionsai.google.dev/gemini-api/docs/available-regions
- 11.API key guidanceai.google.dev/gemini-api/docs/api-key
- 12.Thinking-level tableai.google.dev/gemini-api/docs/thinking
- 13.3.8 model specificationsai.google.dev/gemini-api/docs/models/gemini-3.8-flash
- 14.Google's current pricingai.google.dev/gemini-api/docs/pricing
- 15.active limits in AI Studioaistudio.google.com/rate-limit
- 16.rate-limit guideai.google.dev/gemini-api/docs/rate-limits
- 17.termsai.google.dev/gemini-api/terms





