What is Gemini 3.8 Flash?
Gemini 3.8 Flash is a Google text model with multimodal input, released on 2 September 2026 — a launch its own release blog describes as Google’s third Flash release in only six weeks. Its API model code is gemini-3.8-flash, the model page carries a single Versions line reading Stable with no preview row, and the official “what’s new” page states that the model is generally available and ready for production use. The documented limits are a 1,048,576-token input and a 65,536-token output, the documented inputs are text, images, video, audio and PDF, and the documented output is text. Google states that the model is based on Gemini 3.7 Flash, and the model card’s architecture, training-dataset, training-data-processing, hardware and software sections all refer readers to the 3.7 Flash card rather than restating anything, so this is a refinement of the previous Flash rather than a new line. What places the model more precisely than its marketing line does is a single sentence on Google’s deprecations page, where the schedule for older generations is introduced with the recommendation “For any new projects, use our latest models: 3.5 Flash-Lite or 3.8 Flash.” Two Flash names appear on that list and this is one of them. The same table carries a row for gemini-3.8-flash with a release date of 2 September 2026, a shutdown-date column reading “No shutdown date announced”, and an empty recommended-replacement column. The second thing worth knowing is that Google’s own description of where this model sits runs to three sentences rather than one, and they are not identical. The Gemini API model page and the pricing page both open this model’s section with “our most intelligent Flash model”; the release blog calls it “our best reasoning and coding model yet”; and the DeepMind product page uses “our most intelligent workhorse model yet for coding and agents”. This page reports each by source rather than settling on one, because that language is the provider’s wording rather than a measurement. What can be measured sits in the model card, whose single evaluation table is reproduced below with both the rows it wins and the rows where other columns lead. The knowledge cutoff is given as March 2026, with the qualification that some domains may be more current while others remain limited by January 2025, and the model is listed as reachable through six distribution channels.
What Gemini 3.8 Flash is worth using for
Read the model-card table with all four gaps 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 this model’s favour against the model this release was built on, Gemini 3.7 Flash: DeepSWE v1.1 at 73.7% against 65.3%, Terminal-bench 4.0 at 19.1% against 11.2%, OSWorld-2.0 at 59.0% against 50.6%, GDPVal-AA v2 at 1545 against 1482 on the Elo scale, and LABBench2 at 86.2% against 82.1%. That is exactly the shape that gets written up as a clean sweep, and the same card rules it out without any outside source. On Terminal-bench 4.0 it prints Claude Opus 5 at 51.8% and GPT-5.6 Sol at 37.3%, both far above this model’s 19.1%, with GPT-5.6 Terra at 23.6% above it as well. On GDPVal-AA v2 Claude Opus 5 reads 1824 and GPT-5.6 Sol 1710 against 1545. On OSWorld-2.0 Claude Opus 5 reads 75.4% and GPT-5.6 Sol 62.6% against 59.0%. On GDP.PDF GPT-5.6 Sol reads 40.0% and Claude Opus 5 37.0% against 35.0%. Two narrower margins sit alongside those: on DeepSWE v1.1 Claude Opus 5 reads 74.0% against 73.7%, and on BioMysteryBench (Human Solvable) it reads 90.1% against 88.8%. Counted across all fourteen rows, this model holds first place on eight, second on two, and third or lower on four. One more caution belongs to the same table: LVBench carries two readings in this model’s column, one agentic and one static, and BioMysteryBench is printed as two separate rows, Human Solvable and Human Difficult, so neither is a single measurement and neither pair should be averaged.
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, then list the rows where the first column is not the highest. For each of those rows tell me which model leads and by how much, and flag any benchmark that appears twice on different scales.
Budget for a model that says it uses more tokens by design
This is the one thing about Gemini 3.8 Flash that the provider documents in its own words rather than leaving to reviewers, and the wording is unusually direct. The official “what’s new” page reads: “Gemini 3.8 Flash can use more tokens on longer running and complex tasks, by design. To deliver higher-quality results on difficult, multi-step goals, the model takes smaller reasoning steps, calls tools iteratively, and verifies its work along the way.” The release blog makes the same point from the other side, explaining that the performance gains “stem from a core design choice: 3.8 Flash works harder”, and adds “At times, the model might use more tokens to maximize performance, especially at higher effort levels.” The model card goes further and files that sentence under Known Limitations rather than under benefits. What matters for a budget is that Google publishes the remedy in the same breath, and the remedy is not a discount. The “what’s new” page continues that “not every workflow needs this level of verification. For everyday tasks, you can lower the reasoning effort to reduce token consumption. Alternatively, Gemini 3.7 Flash remains fully supported”, and the blog repeats the second half for applications where compute efficiency is the primary constraint. The official position, then, is that extra token consumption is a deliberate trade and that the correct response is to pick an effort level per task or to stay on 3.7 Flash. No percentage is published anywhere on those pages, which is worth knowing, because the figure most often quoted alongside this claim does not come from Google.
I want to estimate token consumption for a multi-step agent task on this model. Ask me what the task involves and how often it runs. Then help me choose an effort level for it, explain what the official documentation says about the model spending more tokens on longer tasks, and tell me which parts of my estimate are not supported by any published number.
Separate “no shutdown date” from “recommended for new projects”
Two different questions get asked about a model that is no longer the newest in its family, and one official page answers both of them separately. The first is whether it is going away, and Google’s deprecations page answers it in a single row: gemini-3.8-flash carries a release date of 2 September 2026, the shutdown-date column reads “No shutdown date announced”, and the recommended-replacement column is empty. That table does distinguish the two states rather than leaving the column blank throughout, because further down gemini-3.1-flash-lite carries a real shutdown date of 7 May 2027 together with a named replacement. The second question is whether the model is still the right thing to build on, and the same page answers that one too, in the sentence that introduces the schedule: “For any new projects, use our latest models: 3.5 Flash-Lite or 3.8 Flash.” Of the Flash variants Google has shipped, two names appear on that list and this is one of them. A third state is worth separating out as well, because the table shows it: the preview model gemini-3-flash-preview has this generation pointed at gemini-3.6-flash as its replacement, which makes that older model a migration destination rather than a waypoint. The recommendation sentence is a live value that will change as Google ships more models, so this page reports it as at the information-check date rather than as a standing promise.
I found that Gemini 3.8 Flash is not the newest model in its family and I need to decide whether to build on it. Walk me through which official pages to open and what to look for on each, and tell me which results would change the answer. Then tell me what conditions would all have to hold before calling it retired.
Keep the official capabilities apart from this site’s switches
Two sets of statements 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, and a capability list that includes caching, code execution, file search, function calling, structured outputs, URL context, search grounding, Google Maps grounding, thinking, and computer use marked as a preview, with audio generation, image generation and the Live API listed as not supported. Google also publishes a default thinking level of medium for this model, with low, medium and high accepted and minimal returning an error. On this site’s side the catalog entry behind the page declares vision as unsupported, so the attach button takes PDF files only; it declares web search as unsupported, so the search switch is disabled; and it declares reasoning as unsupported, so the composer offers no way to choose a thinking level even though Google lists thinking among the model’s capabilities. That reasoning gap is worth stating plainly, because the three controls a visitor is most likely to look for here are exactly the three that are not switched on. Attachments accept 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 limit of the model, and because the catalog entry publishes no context length the interface displays no context figure at all. The catalog entry also publishes no price for a web search, though that is moot here since search is unavailable. Read the specification cards as Google describing Gemini 3.8 Flash, and the controls in the chat box as what this site actually switches on.
I want to use this chat box on a task that involves reading source material. Ask me what the material is before answering. Then tell me which parts of the task are limited by this interface rather than by the model, which official capabilities have no equivalent control here, and how to split the work around the per-reply output ceiling.
Official Gemini 3.8 Flash specifications
- Developer
- Model ID
- gemini-3.8-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 3.8 Flash and Gemini 3.7 Flash side by side
Every value below comes from one table, the single evaluation table inside Google’s Gemini 3.8 Flash model card, checked on 25 September 2026, so a single official page carries both columns. The left column is this page’s model and the right column is Gemini 3.7 Flash, which the card itself names as the model this release is based on, and the card’s own two price rows already show the pair at the same figures. Rows are matched by benchmark name, never by position. That table also prints four further columns — Claude Opus 5, Claude Sonnet 5, GPT-5.6 Sol and GPT-5.6 Terra — and reading only the first two would make this comparison look like a clean sweep, which is why the rows those columns lead are named in the feature section above instead of being dropped: on Terminal-bench 4.0 Claude Opus 5 reads 51.8% against 19.1% here, on GDPVal-AA v2 it reads 1824 against 1545, on OSWorld-2.0 it reads 75.4% against 59.0%, and on GDP.PDF GPT-5.6 Sol reads 40.0% against 35.0%. Three things have to travel with the numbers. All fourteen rows move in this model’s favour against 3.7 Flash, so the table supports a claim about that pair and nothing wider; across the same fourteen rows this model holds first place on eight, second on two and third or lower on four. LVBench carries two readings in the official column, one agentic and one static, and BioMysteryBench is printed as two separate rows, Human Solvable and Human Difficult, so the row labels carry the scale and a pair must not be averaged into one number. Finally, the card links its evaluation methodology out to a separate page rather than stating it inline, and its two price rows are deliberately left out of this comparison and reported in the pricing section instead. Read this as a provider comparing its own releases and its own peers, not as a promise about the quality or speed of a conversation on this site.
| Provider benchmark | Gemini 3.8 Flash | Gemini 3.7 Flash |
|---|---|---|
| DeepSWE v1.1 | 73.7% | 65.3% |
| GDPVal-AA v2 (Elo) | 1545 | 1482 |
| Vals Finance Agent v2 | 61.4% | 59.0% |
| Harvey’s Legal Agent Benchmark | 10.0% | 8.8% |
| Terminal-bench 2.1 | 89.4% | 85.8% |
| Terminal-bench 4.0 | 19.1% | 11.2% |
| GDP.PDF | 35.0% | 34.0% |
| CharXiv Reasoning (no tools) | 86.2% | 84.5% |
| LVBench · agentic | 87.8% | 85.4% |
| HLE-Verified | 54.9% | 53.6% |
| OSWorld-2.0 | 59.0% | 50.6% |
| BioMysteryBench · Human Solvable | 88.8% | 87.1% |
| BioMysteryBench · Human Difficult | 56.5% | 43.5% |
| LABBench2 | 86.2% | 82.1% |
How to use Gemini 3.8 Flash online
Choose the model
This page opens with Gemini 3.8 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 thinking-level control and no web search switch for this model, so a tightly scoped instruction is the main lever you have. 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 3.8 Flash costs
Google publishes four rate tables for this model, all quoted per million tokens in US dollars, and all four attach two dates to the same two figures. On the Standard tier an input costs $0.75 through December 31, 2026 and $1.50 starting January 1, 2027, and an output costs $3.75 and then $7.50, with the output figure stated to include thinking tokens; a cache read costs $0.075 and then $0.15, and cached storage is billed separately at $0.50 per million tokens per hour through the same date, rising to $1.00. Batch and Flex both sit at $0.375 for input and $1.875 for output, with cache reads of $0.0375 rising to $0.075, and Priority is the dearest tier at $1.35 for input and $6.75 for output with a cache read of $0.135, each rising to $2.70, $13.50 and $0.27 on the same date. The window is not this model’s alone. The official “what’s new” page states that the introductory pricing runs across Google AI Studio and Gemini Enterprise Agent Platform through December 31, 2026 for Gemini 3.8 Flash, Gemini 3.7 Flash and Gemini 3.6 Flash together, with standard pricing taking effect on January 1, 2027, and the pricing page quotes all three at identical figures tier for tier. Two rows are shared across the tiers: Grounding with Google Search is unavailable on the free tier and on paid tiers includes 5,000 free search requests per month shared across the Gemini 3.x family before being billed at $14 per 1,000 requests, and Grounding with Google Maps includes 5,000 prompts per month at the same rate thereafter. The free tier is worth reading per tier rather than as a whole, because it reads “free of charge” on Standard and Priority, “not available” on Batch and Flex, and the page notes that free-tier traffic may be used to improve Google’s products while paid-tier traffic is not. This dated pair is not the only official statement of price: the model card prints the same two figures as $0.75 ($1.50 regular) and $3.75 ($7.50 regular) without dates, and the release blog states $0.75 and $3.75 without dates either, so the two readings are reported here by source — the pricing page carries the dates, the other pages do not. 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 101.3514 points per million input tokens, 506.7568 points per million output tokens and 10.1351 points per million cached tokens. Those are this site’s own figures rather than a conversion of any official tier, so no dollar figure above should be read as their source; the catalog also publishes no separate price for a web search.
View points and plans ↗Gemini 3.8 Flash FAQ
Which rows of the official table does Gemini 3.8 Flash not lead?
Four of the fourteen benchmark rows in the model card, and the leading figures are printed by the card itself. On Terminal-bench 4.0 Claude Opus 5 reads 51.8% against this model’s 19.1% — a gap of 32.7 percentage points — with GPT-5.6 Sol at 37.3% and GPT-5.6 Terra at 23.6% also ahead. On GDPVal-AA v2 Claude Opus 5 reads 1824 on the Elo scale against 1545, with GPT-5.6 Sol at 1710 and Claude Sonnet 5 at 1584 also above. On OSWorld-2.0 Claude Opus 5 reads 75.4% and GPT-5.6 Sol 62.6% against 59.0%. On GDP.PDF GPT-5.6 Sol reads 40.0% and Claude Opus 5 37.0% against 35.0%. Across all fourteen rows this model is first on eight, second on two and third or lower on four. Those four rows are exactly why the comparison above is written as a comparison against Gemini 3.7 Flash rather than as a claim about the whole table.
Is Gemini 3.8 Flash retired, and is it recommended for new projects?
Google has published no shutdown date for it, and the page that would carry one recommends it. The deprecations page has a row for gemini-3.8-flash with a release date of 2 September 2026, a shutdown date reading “No shutdown date announced”, and an empty replacement column. In the sentence introducing that schedule, Google writes “For any new projects, use our latest models: 3.5 Flash-Lite or 3.8 Flash.” Two Flash names appear on that list and this is one of them. The table does distinguish the states, because gemini-3.1-flash-lite further down carries a shutdown date of 7 May 2027 and a named replacement. The recommendation sentence is a live value that changes as more models ship, so this page reports it as at the information-check date.
What does Gemini 3.8 Flash cost, and when does that price change?
Google’s pricing page quotes four tiers, and every tier carries two figures with two dates. Standard input is $0.75 per million tokens through December 31, 2026 and $1.50 starting January 1, 2027; Standard output is $3.75 and then $7.50. Batch and Flex both sit at $0.375 and $1.875, and Priority, the dearest tier, at $1.35 and $6.75. The window is shared rather than personal to this model: the official “what’s new” page states that the introductory pricing runs through December 31, 2026 for Gemini 3.8 Flash, Gemini 3.7 Flash and Gemini 3.6 Flash together, with standard pricing taking effect on January 1, 2027, and the pricing page quotes all three at identical figures. Google’s model card and release blog also state $0.75 and $3.75, but without dates, so the two readings are reported here by source. This site does not use those dollar figures: it bills in points, and its catalog reference price is its own basis.
Does Gemini 3.8 Flash use more tokens than Gemini 3.7 Flash?
Google says it can, and says so by design. The official “what’s new” page reads “Gemini 3.8 Flash can use more tokens on longer running and complex tasks, by design”, explaining that the model takes smaller reasoning steps, calls tools iteratively and verifies its work. The release blog describes the same thing as a core design choice — “3.8 Flash works harder” — and adds that the model “might use more tokens to maximize performance, especially at higher effort levels”; the model card repeats that sentence under Known Limitations rather than under benefits. Google also publishes the response: for everyday tasks, lower the reasoning effort to reduce consumption, or continue to use Gemini 3.7 Flash, which remains fully supported. No percentage appears on any of those pages.
Is Gemini 3.8 Flash a new architecture?
No, and Google says so directly. The model card’s dependencies section states that Gemini 3.8 Flash is based on Gemini 3.7 Flash, and the architecture, training dataset, training-data processing, hardware and software sections all refer readers to the 3.7 Flash card rather than restating anything. What the card claims for this release is performance advancement across software engineering and agentic knowledge workflows, together with customizable effort levels, which is a refinement claim rather than an architecture claim. The documented limits do not distinguish the two models either: both are published with a 1,048,576-token input and a 65,536-token output.
Can Gemini 3.8 Flash read images and PDFs here?
PDFs, yes; images, no, and that is a difference between this site’s controls and the official specification rather than a property of the model. Google documents text, images, video, audio and PDF as inputs and text as the output. The catalog entry behind this page declares vision as unsupported, so the attach button accepts PDF files only, up to 24 MB each, and a PDF is converted to text before it is sent, which means a scan the extractor cannot read returns an error rather than an empty answer. Audio and video uploads are not available here either.
Why is there no thinking-level selector and no web search switch here?
Because the catalog entry behind this page declares reasoning and web search as unsupported, so the composer has no working control for either. The official model differs on both counts: Google lists thinking among the capabilities of Gemini 3.8 Flash, with low, medium and high levels and a published default of medium, and it also lists search grounding and Google Maps grounding. Those are gaps between the official model and this interface rather than properties of the model, and they are stated here instead of left to be discovered. The three controls a visitor is most likely to look for — image attachments, web search and a thinking level — are the three that are not switched on for this model here.
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, and it states a knowledge cutoff of March 2026 with the qualification that some domains may be more current while others remain limited by January 2025. 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.
Sources and model information
Information checked
- Google AI for Developers · Gemini 3.8 Flash model page ↗
- Google DeepMind · Gemini 3.8 Flash model card ↗
- Google AI for Developers · Pricing ↗
- Google AI for Developers · Gemini deprecations ↗
- Google AI for Developers · What’s new in Gemini 3.8 Flash ↗
- Google Blog · Introducing Gemini 3.8 Flash and 3.8 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 3.8 Flash ↑