What is Gemini 3.5 Flash?
Gemini 3.5 Flash is a Google text model with multimodal input, released on 19 May 2026 as the first member of the Gemini 3.5 family. Its API model code is gemini-3.5-flash, it is listed as Stable rather than preview, and Google documents a 1,048,576-token input limit with a 65,536-token output limit. The documented inputs are text, images, video, audio and PDF, and the documented output is text. Two things about this model are easy to get wrong, and both are worth settling before you decide whether to use it. The first is its position in the family. Google has since released four more Flash models: Gemini 3.6 Flash and Gemini 3.5 Flash-Lite on 21 July 2026, Gemini 3.7 Flash on 13 August 2026, and Gemini 3.8 Flash on 2 September 2026. That places Gemini 3.5 Flash four releases behind on the Flash line, and Google’s own pricing page describes it, in the sentence that names the model, as “our earlier Flash model, built for speed and foundational performance”. A different Google page, the what’s-new page written for this model, still describes it as “our most intelligent Flash model”. Both pages are official and both are still published, so this page reports each by source rather than choosing between them. The second mistake is the opposite of the first. Being superseded is not the same as being withdrawn. Google’s deprecation page carries a row for gemini-3.5-flash with a release date of 19 May 2026 and an empty shutdown-date column, and it states no shutdown date for the model. Compare that with Gemini 3.1 Flash-Lite two rows below on the same table, which carries an actual shutdown date of 7 May 2027 and a named replacement: the page distinguishes the two states, so you can too. The accurate sentence is therefore that Gemini 3.5 Flash is a stable, still-served model that four newer releases have overtaken, and this page writes both halves of that sentence wherever it matters. In this workspace the model is available in a shared chat box, and the specification cards further down describe the model Google documents rather than the controls in the box above.
What Gemini 3.5 Flash is worth using for
Check whether a retirement claim is still current
Most of what is written about Gemini 3.5 Flash was written in the weeks after its launch, when it was the newest Flash model, and the version of the model those articles describe has been overtaken four times since. The useful question now is not what any one article says but what Google currently shows, and three checks answer it mechanically. Open the deprecations page and find the row for gemini-3.5-flash: its release date is 19 May 2026 and its shutdown-date column is empty, which that page renders as no shutdown date announced. Open the pricing page and confirm the model still holds a section of its own with four rate tables under it. Open the API model page and confirm the version line still reads Stable. If all three hold, the model is current, whatever an older page says. The same method also protects you from the error in the other direction, because Gemini 3.1 Flash-Lite sits two rows above this one on that same table with a real shutdown date of 7 May 2027 and a named replacement. The deprecations page is the place those two states are told apart, and it is the only page whose whole purpose is to do so.
I read an article saying Gemini 3.5 Flash has been retired and that requests now go to a newer model. Walk me through how to check that claim myself: which official pages to open, what to look for on each, and what result on each would change the conclusion. Then tell me exactly what would have to be true for the article to be correct.
Plan around the documented context, not around this chat box
The documented context is 1,048,576 tokens and the documented maximum output is 65,536 tokens, and both numbers describe Google’s API rather than one conversation in this box. The gap matters. This chat box publishes no context length at all, so the interface shows no context figure, and it requests at most 4,096 output tokens per reply. That figure is a workspace setting far below the official ceiling, which means a long deliverable has to be produced in parts with the shared context carried forward between them. The context number also deserves one caveat that comes from Google’s own card rather than from commentary. The card measures long context twice on the same benchmark, MRCR v2 with eight needles: at 128k tokens the model reads 77.3%, and at the full 1M tokens, scored pointwise, it reads 26.6%. That is not a statement that the model fails at a million tokens. It is the card showing that the published ceiling and the published accuracy at that ceiling are two different claims, and it is the reason to plan a long job around what the model reads at the length you are actually sending rather than around the largest number on the page.
I need to work through a long specification and produce a long written deliverable from it. Ask me how long the source is and how long each finished section needs to be. Then tell me how to split the job, what to carry forward between the parts, and where the output ceiling of this chat box will force a change of plan.
Read the model-card table with its caveats attached
Google’s model card for this release publishes one evaluation table, and the column describing this model is the first one. Fourteen rows of it are reproduced further down this page, and every one of them moves in the same direction against the model this release was built on, Gemini 3 Flash: Terminal-bench 2.1 reads 76.2% against 58.0%, MCP Atlas 83.6% against 62.0%, GDPval-AA 1656 against 1204 on the Elo scale. That much is unambiguous. What the table does not support is a claim that this model leads everything on the page, and the same card makes that clear without any outside source. On Humanity’s Last Exam it reads 40.2% and sits behind Gemini 3.1 Pro at 44.4%; on ARC-AGI-2 it reads 72.1% and sits behind Gemini 3.1 Pro at 77.1% and GPT-5.5 at 84.6%; on MRCR v2 at 128k it reads 77.3% and sits behind GPT-5.5 at 94.8%. The card also links its methodology out to a separate page rather than stating it inline, and it attaches a matching evaluation condition to one column rather than to both. Read the table as a provider comparing two of its own releases, which is genuinely useful, rather than as a ranking of the field. The launch announcement repeats four of these figures word for word, which is the cross-check that makes them safe to quote here.
Here is an official benchmark table with several model columns. Compare the first column with the column I name, give me the size of every gap, and then point out the rows where the first column is not the highest. For each of those rows, tell me which model leads and by how much.
Keep the official capabilities apart from this site’s switches
Two sets of numbers describe this model and they answer different questions. On the official side there is a 1,048,576-token input limit, a 65,536-token output limit, thinking in four levels from minimal to high with a default of medium, and a list of API features that includes code execution, function calling, structured outputs, file search, URL context and search grounding. On this site’s side the catalog entry behind the page declares thinking as unsupported, so the composer offers no reasoning-level control, and it declares web search as supported, so the search switch is there. That combination is worth stating plainly, because it runs opposite to the pattern on most model pages, where search is the missing control and reasoning is the one that is present. Attachments accept JPEG, PNG, WebP and PDF files up to 24 MB, and a PDF is converted to text before the model sees it, so a scan the extractor cannot read comes back as an error instead of an empty answer. Each reply is requested with at most 4,096 output tokens, which is a setting of this site and not a property of the model. Read the specification cards as Google describing Gemini 3.5 Flash, and the controls in the chat box as what this site actually switches on.
I want to use this chat box for a task that needs a lot of source material. Ask me what the task is before answering. Then tell me which parts of it are limited by this interface rather than by the model, which control I do have here that I might not expect, and how to split the work around the per-reply ceiling.
Official Gemini 3.5 Flash specifications
- Developer
- Model ID
- gemini-3.5-flash
- Official input limit
- 1,048,576 tokens
- Official output limit
- 65,536 tokens
- Official inputs
- Text, images, video, audio and PDF
- Output
- Text
Gemini 3.5 Flash and Gemini 3 Flash side by side
Every value below comes from one table, the evaluation table inside Google’s Gemini 3.5 Flash model card, checked on 23 September 2026, so a single official page carries both columns. The left column is this page’s model and the right column is Gemini 3 Flash, the model the card names as the one this release was built on; it is also the comparison most often searched for, and the only one where both columns sit in the same official table. Rows are matched by benchmark name, never by position. Three things have to travel with the numbers. The table is not a clean sweep, and this page would be misleading if it were presented as one: on Humanity’s Last Exam, ARC-AGI-2 and the 128k reading of MRCR v2 the first column sits behind models that appear on the same card and are named in the guide above. The card links its methodology out to a separate page rather than stating it inline, and it does not attach a matching evaluation condition to both columns. MRCR v2 appears twice, once as a 128k average and once as a 1M pointwise reading of the same benchmark; those are two scales of one measurement and must not be read together. Read this as a provider comparing two of its own releases, not as a promise about the quality or speed of a conversation on this site.
| Provider benchmark | Gemini 3.5 Flash | Gemini 3 Flash |
|---|---|---|
| Terminal-bench 2.1 | 76.2% | 58.0% |
| SWE-Bench Pro (Public) | 55.1% | 49.6% |
| MCP Atlas | 83.6% | 62.0% |
| Toolathlon | 56.5% | 49.4% |
| OSWorld-Verified | 78.4% | 65.1% |
| Finance Agent v2 | 57.9% | 42.6% |
| GDPval-AA (Elo) | 1656 | 1204 |
| CharXiv Reasoning | 84.2% | 80.3% |
| MMMU-Pro | 83.6% | 81.2% |
| Blueprint-Bench 2 | 33.6% | 0.0% |
| MRCR v2 (8-needle) · 128k avg | 77.3% | 67.2% |
| MRCR v2 (8-needle) · 1M pointwise | 26.6% | 22.1% |
| Humanity’s Last Exam | 40.2% | 33.7% |
| ARC-AGI-2 | 72.1% | 33.6% |
How to use Gemini 3.5 Flash online
Choose the model
This page opens with Gemini 3.5 Flash selected. The model navigation on the left and the selector in the chat box stay in sync, and switching models keeps the current conversation. Start a new chat when you want a clean context.
Write the task and the output format
State the goal, the constraints and the shape of the answer you want. This box exposes no reasoning-level control for this model, so a tightly scoped instruction is the main lever you have; the search switch, by contrast, is available here. Any prompt in this guide can be sent as it stands or edited first.
Sign in, send and check the cost
Sending a message needs an account with available points. The composer reads a reference price from the live model catalog, and a finished request reports what it actually cost. Follow up with corrections rather than starting over.
What Gemini 3.5 Flash costs
Google publishes four pricing tiers for this model, all quoted per million tokens in US dollars. On the Standard tier an input costs $1.50 and an output costs $9.00, with the output figure stated to include thinking tokens; a cache read costs $0.15, and cached storage is billed separately at $1.00 per million tokens per hour. Batch and Flex both sit at $0.75 for input and $4.50 for output, with cache reads of $0.075 and $0.08 respectively, and Priority is the dearest at $2.70 for input and $16.20 for output with a cache read of $0.27. Two rows are shared across the tiers. Grounding with Google Search is unavailable on the free tier, and on paid tiers it includes 5,000 free search requests per month shared across the Gemini 3.x family before being billed at $14 per 1,000 requests; Grounding with Google Maps includes 5,000 prompts per month shared across Gemini 3 at the same $14 rate thereafter. The same tables state that free-tier traffic may be used to improve Google’s products and paid-tier traffic may not. Nothing in any of the four tables carries a discount, a promotional rate or an end date for this model, so none of the figures should be read as a limited-time price; they are a snapshot dated by the information-check date below. None of that is this site’s checkout price. Here you pay in points, and as of the information-check date the catalog entry lists 360 points per million input tokens, 2,200 points per million output tokens and 36 points per million cached tokens, with a web search costing 1 point per query.
View points and plans ↗Gemini 3.5 Flash FAQ
Is Gemini 3.5 Flash deprecated or scheduled to shut down?
Google has published no shutdown date for it. The official deprecations page carries a row for gemini-3.5-flash with a release date of 19 May 2026 and an empty shutdown-date column, and its recommended-replacement column is empty as well. That is a different arrangement from Gemini 3.1 Flash-Lite, two rows away on the same table, which carries a shutdown date of 7 May 2027 and names a replacement. Being superseded and being withdrawn are separate things, and for this model only the first is true: Google has since released Gemini 3.6 Flash, Gemini 3.5 Flash-Lite, Gemini 3.7 Flash and Gemini 3.8 Flash, and the model is still served.
Which models replaced Gemini 3.5 Flash, and when were they released?
Google’s deprecations table gives the dates directly. Gemini 3.5 Flash was released on 19 May 2026. Gemini 3.6 Flash and Gemini 3.5 Flash-Lite both followed on 21 July 2026, Gemini 3.7 Flash on 13 August 2026, and Gemini 3.8 Flash on 2 September 2026. The 3.6 announcement adds a measurable claim: on the Artificial Analysis Index, 3.6 Flash reduces output token usage by 17% compared with 3.5 Flash, rising to as much as 65% on DeepSWE, and it is priced at $1.50 and $7.50 per million input and output tokens, which the announcement itself describes as lower than 3.5 Flash against the $1.50 and $9.00 published here. None of those rows carries a shutdown date.
How does Gemini 3.5 Flash compare with Gemini 3 Flash?
The model card for this release publishes one evaluation table, Gemini 3 Flash is one of its columns, and that is why the table above uses it: it is the only comparison where a single official page proves both columns. Across the fourteen rows shown, the newer model is ahead on every one, with the largest gaps in agentic and tool-use work — MCP Atlas at 83.6% against 62.0%, Terminal-bench 2.1 at 76.2% against 58.0%, and GDPval-AA at 1656 against 1204 on the Elo scale. The honest qualification is that the same card also shows this model behind other models on Humanity’s Last Exam, ARC-AGI-2 and the 128k reading of MRCR v2, so the table supports a strong agentic profile rather than uniform leadership.
How is Gemini 3.5 Flash priced by Google, and how is it priced here?
Google publishes four tiers, all per million tokens in US dollars. Standard is $1.50 in and $9.00 out, with a $0.15 cache read plus $1.00 per million tokens per hour of cached storage; Batch and Flex are both $0.75 in and $4.50 out, with cache reads of $0.075 and $0.08; Priority is $2.70 in and $16.20 out with a $0.27 cache read. No tier carries a discount or an end date. This site does not use those dollar figures at all: it bills in points, and the catalog entry currently lists 360 points per million input tokens, 2,200 per million output tokens and 36 per million cached tokens, with 1 point per web search.
Can I use Gemini 3.5 Flash for free, and do I need an API key?
Browsing this page and preparing a prompt costs nothing, and no Google API key is required, because the conversation runs on your NanoBanana Pro account. Sending a message does need an account with available points, so this is not unlimited free chat here. Whether an account carries trial points depends on the offer shown inside that account. If you would rather call Google directly, the documentation pages linked below are the right starting points, and Google’s own free tier for this model is described on its pricing page.
Can Gemini 3.5 Flash read images and PDFs here?
Yes to both, and this is a point where the official specification and the site controls line up instead of differing. Google documents text, image, video, audio and PDF as inputs and text as the output, and the catalog entry behind this page enables vision. The composer therefore accepts JPEG, PNG, WebP and PDF files up to 24 MB. A PDF is not handed to the model as a document: this site extracts its text first, so a scanned page the extractor cannot read comes back as an error rather than an empty answer, and complex layouts may need a clearer image or manually pasted text. Audio and video uploads are not available here even though the model documents them as inputs.
Why is there no thinking-level selector in this chat box?
Because the catalog entry behind this page declares reasoning as unsupported, so the composer has no control to show. The official model differs here: Google documents four levels, minimal, low, medium and high, with medium as the default, changed from high in the preview generation, and it also documents thought preservation across turns. That is a gap between the official model and this interface rather than a property of the model, and it is stated here instead of left to be discovered. It is also worth noting that this page is the mirror image of the usual pattern in one respect: web search is switched on here, so the reasoning control is the one that is missing.
How long can an input or an answer be?
This site requests at most 4,096 output tokens per reply, which is a workspace setting and not a limit of the model. Google documents a 1,048,576-token input limit and a 65,536-token maximum output for the API. Because the catalog entry behind this page publishes no context length, the interface shows no context figure, so the official number is an API specification rather than a description of this page. Long deliverables are best asked for in parts, with the shared context carried forward between them. One caveat is worth carrying: on the model card’s own long-context benchmark, the 1M-token reading is 26.6% against 77.3% at 128k, so a larger context is not automatically a better one.
Sources and model information
Information checked
- Google AI for Developers · Gemini 3.5 Flash model page ↗
- Google DeepMind · Gemini 3.5 Flash model card ↗
- Google AI for Developers · What’s new in Gemini 3.5 Flash ↗
- Google AI for Developers · Pricing ↗
- Google AI for Developers · Gemini deprecations ↗
- Google Blog · Gemini 3.5: frontier intelligence with action ↗
- Google Blog · Introducing Gemini 3.6 Flash, 3.5 Flash-Lite and 3.5 Flash Cyber ↗
Start with the task you run most often
Give it the goal, the constraints and the format you need, and keep the inputs within what this chat box can actually take.
Chat with Gemini 3.5 Flash ↑