Gemini 3.8 Flash Launch: Pricing, Benchmarks, and What’s New

Google DeepMind released Gemini 3.8 Flash on September 2, 2026, at $0.75 per million input tokens and $3.75 per million output tokens, unchanged from 3.7 Flash. The Gemini 3.8 Flash launch brings a roughly six-point gain on long-horizon coding and a model tuned to keep working through multi-step agent tasks. This guide covers the full rate card (including the price doubling on January 1, 2027), every benchmark Google has published with a number attached, and the checks worth running before you move production traffic. As of September 02, 2026, all figures come from Google’s own launch materials.

• $0.75 in / $3.75 out per million tokens on the standard tier through December 31, 2026, per the Gemini Developer API pricing page (2026)
• ~71.0% on DeepSWE long-horizon software engineering, against 65.3% for 3.7 Flash on v1.1, per the Gemini 3.8 Flash model card (September 2026)
• 54.9% on HLE-Verified multidisciplinary reasoning, per Google’s launch post by Tulsee Doshi and Raluca Ada Popa (September 2026)
• 1,048,576 input tokens and 65,536 output tokens per request, per Google’s Gemini 3.8 Flash model documentation (2026)
• Third Flash release in six weeks, with 3.7 Flash having shipped on August 13, 2026, per Google’s launch post (2026)

What Is Gemini 3.8 Flash and Why This Launch Matters

Gemini 3.8 Flash is Google DeepMind’s mid-tier model, released on September 2, 2026, and it matters because it upgrades reasoning and coding in the Flash tier without changing the speed or price of the model it replaces. Google calls it its most intelligent workhorse model. The pitch is narrow: keep the Flash cost profile, get closer to frontier-class results on long-running agent and software engineering tasks.

Two variants shipped on the same day, according to Tulsee Doshi and Raluca Ada Popa’s launch post on the Google blog:

• Gemini 3.8 Flash: the general-purpose model, live in the Gemini app, Google AI Studio, the Gemini API, Google Antigravity, AI Mode in Search, and the Gemini Enterprise Agent Platform (formerly Vertex AI).
• Gemini 3.8 Flash Cyber: a cybersecurity variant built for vulnerability detection and automated patching, offered only to vetted defenders through Google’s new Fairwind Program.

Both run on one shared core. Google says the coding and reasoning gains came partly from training in cybersecurity, and partly from long-running agentic loops that evaluate and refine the model on itself.

The cadence is the striking part. Gemini 3.7 Flash launched on August 13, 2026. Gemini 3.8 Flash arrived three weeks later, which makes it the third Flash release in six weeks. And it’s built directly on the 3.7 Flash architecture: the model card sends you back to the 3.7 card for training and architecture details. So what Google shipped is a refinement of a base model that’s a month old, tuned harder on agent behaviour. That shapes how you should evaluate it. Treat it as a drop-in upgrade candidate rather than a platform change.

The spec sheet, per Google’s model documentation as of September 2, 2026:

• Context window: 1,048,576 input tokens, 65,536 output tokens
• Input modalities: text, images, video, audio, PDF
• Output modality: text only
• Knowledge cutoff: March 2026 for most domains, January 2025 for some
• Thinking levels: low, medium, high (the “minimal” level isn’t supported)
• Tooling: function calling, code execution, computer use (preview), file search, search grounding, Google Maps grounding, structured outputs, URL context, caching, Batch, Flex, and Priority inference

One thing to notice in that list. A 64K output ceiling on a model designed to reason at length and call tools repeatedly means long agent runs will hit the output limit sooner than you’d expect from a 1M input window.

What Does Gemini 3.8 Flash Actually Cost per Million Tokens?

Gemini 3.8 Flash costs $0.75 per million input tokens and $3.75 per million output tokens on the standard tier through December 31, 2026, and both rates double on January 1, 2027. That’s the exact price Gemini 3.7 Flash launched at three weeks earlier. Nothing moved.

Here’s the full rate card from the Gemini Developer API pricing page, checked September 2, 2026:

All figures are per one million tokens.

Standard output moves from $3.75 to $7.50 per million tokens on January 1, 2027, and every other tier doubles with it. Source: Gemini Developer API pricing page, 2026.

The pricing page shows one promotional window covering the whole Flash 3.x point-release family, from 3.5 through 3.8, with a single shared end date. Google didn’t reset the clock for 3.8. That matters if you’re budgeting: the introductory rate has under four months left, whichever Flash 3.x model you’re on. For context, Google described the 3.7 Flash launch price in August 2026 as roughly half the per-token cost of 3.6 Flash, so 3.8 inherits a discount that was already steep.

Now the caveat that most launch-day coverage skips. Per-token parity with 3.7 Flash does not mean per-task parity. Google’s own launch post is candid about why:

“At times, the model might use more tokens to maximize performance, especially at higher effort levels.” (Tulsee Doshi and Raluca Ada Popa, Google, September 2, 2026)

The model is designed to run more reasoning steps and call tools more times on hard problems. Anyone who has watched an agent loop on a per-token bill knows the cost lives in the transcript length, and a model that retries a tool call three times instead of once produces a longer transcript at the same unit rate. Google’s suggested workaround is to drop to a lower thinking level where compute matters more than quality, or to stay on 3.7 Flash, which remains supported for efficiency-first workloads.

Google Cloud’s enterprise catalog lists the model too, though as of September 2, 2026 it doesn’t publish per-token rates that differ from the API figures above.

How Gemini 3.8 Flash Scores on Reasoning, Coding, and Multimodal Benchmarks

Gemini 3.8 Flash scores roughly 71.0% on DeepSWE long-horizon software engineering, 54.9% on HLE-Verified, 61.4% on Vals Finance Agent V2, and 10.0% on Harvey’s Legal Agent Benchmark, per Google’s launch figures from September 2, 2026. The DeepSWE result is the headline: 3.7 Flash posted 65.3% on the related DeepSWE v1.1 in August 2026, so that’s a jump of nearly six points in three weeks.

Long-horizon software engineering is the strongest published result at roughly 71.0% on DeepSWE, up from 65.3% for 3.7 Flash on v1.1. Source: Gemini 3.8 Flash model card, Google DeepMind, 2026.

A note on the DeepSWE labels, because they’re inconsistent. The Gemini 3.8 Flash model card reports the new score against DeepSWE v1, while Google’s launch post and the 3.7 Flash card cite v1.1. Google treats them as comparable. I’d want to see the version pinned before quoting the six-point gain in a budget memo.

The finance and legal numbers deserve a second look. Google says 3.8 Flash beats both 3.7 Flash and other frontier models on Vals Finance Agent V2 and Harvey’s Legal Agent Benchmark, and 61.4% on a financial-agent task is a real result. But 10.0% on Harvey is still 10.0%. Topping a benchmark and being production-ready on that benchmark are two separate claims, and only the first one is on the table here. For the agent work we ship at AlphaCorp AI, the pattern we watch for in these agentic evaluations is whether a score gain came from better single-shot answers or from the model retrying and tool-calling its way to a result, since the second one shows up on the invoice. Google’s own explanation for 3.8 points squarely at the second.

Two more results, both without a published number. Google reports a jump in resistance to prompt injection on the Gray Swan indirect-prompt-injection benchmark for the 3.8 models, shown only as a chart. And the evaluation report names Terminal-Bench 2.1 and MRCR v2 (long-context recall) as tested categories but renders those scores as images, so any figure you see quoted for them is someone’s reading of a graphic.

The Cyber variant supplies evidence that the shared core is real. Per Google’s launch post, Gemini 3.8 Flash Cyber hits 47.2% pass@1 on Collinear’s CWE-Bench patching benchmark against 47.8% for a leading frontier model, exceeds a 70% success rate on Google’s internal 20-language vulnerability discovery set, and produced 2.6 times more correct Chrome vulnerability patches than larger commercial models in the Chrome Security team’s testing. Those are Google’s numbers about Google’s model. Read them that way.

What’s New in Gemini 3.8 Flash Compared to Gemini 3.5 Flash

Compared to Gemini 3.5 Flash, Gemini 3.8 Flash is the product of three point releases in a row, and the changes stack: 3.7 Flash halved the per-token price and posted large benchmark jumps in August 2026, then 3.8 Flash added more persistent agent behaviour on top of the same architecture in September 2026. Google has published far more detail on those last two steps than on the 3.5 to 3.6 move, so that’s where the record is.

The 3.7 release did most of the heavy lifting. Google’s announcement of Gemini 3.7 Flash on August 13, 2026 put its launch price at roughly half of 3.6 Flash’s per-token rate and reported these gains over 3.6:

DeepSWE v1.1 went from 49.0% on 3.6 Flash to 65.3% on 3.7 Flash, the largest of the gains Google published for that release. Source: Google blog, Gemini 3.7 Flash announcement, 2026.

Sixteen points on DeepSWE. AutomationBench nearly doubled. That’s one release.

What 3.8 Flash adds on top of that base is behavioural instead of architectural. Per the model card and Google’s developer changelog, both dated September 2, 2026:

• Agentic diligence: the model runs extra reasoning steps and calls tools repeatedly on hard tasks, which Google credits for the finance- and legal-agent gains.
• Fewer dead ends: Google states the model “reduces failed loops and errors” in multi-step agent workflows.
• Default on Managed Agents: 3.8 Flash became the default model for Gemini’s Managed Agents platform at launch.
• Safety shifts versus 3.7 Flash: text-to-text safety down 0.4 points, multilingual safety up 5.4, image-to-text safety unchanged, tone up 0.2, and unjustified refusals up 1.1. That last one is a small regression, and worth checking against your own prompt set.
• Frontier Safety Framework: no Tracked or Critical Capability Levels reached, with safeguards on CBRN and cyber-offense misuse.

One related change landed the day before. On September 1, 2026, Google’s Gemini API changelog recorded “agentic video understanding” for the Flash family, which moves through video content dynamically and is claimed to use up to 88% fewer tokens on long-form video than static frame processing. The changelog names 3.7 Flash, 3.6 Flash, and 3.5 Flash-Lite as the models that received it.

How Gemini 3.8 Flash Compares to GPT and Claude Models in Its Price Tier

No independent head-to-head benchmark of Gemini 3.8 Flash against GPT or Claude models existed on September 2, 2026, so every comparative figure available at launch comes from Google or its partners. That’s normal for a same-day release. It also means the honest answer to “is it better than the competition” is: Google says it’s close, and nobody neutral has checked yet.

Google’s claim is specific. The launch post says 3.8 Flash often approaches the performance of higher-cost frontier models, and on long-horizon coding it goes further:

“3.8 Flash outperforms most larger frontier models in autonomously solving complex engineering problems end to end, only at a fraction of the cost.” (Tulsee Doshi and Raluca Ada Popa, Google, September 2, 2026)

The competitor models in that comparison appear only in chart images. The post’s text never names them.

Two partner numbers add texture, both from the Cyber variant. Wiz measured 7.5 to 9.7 percentage points higher recall on its internal penetration-testing benchmark at 2.3 to 5.2 times lower cost than other leading frontier models, and Collinear’s CWE-Bench put the Cyber model within a point of the top frontier score on patching. Because 3.8 Flash Cyber shares a core with 3.8 Flash but ships with different mitigations and a restricted audience, those results hint at the base model’s ceiling without proving it.

Then there are the aggregator sites. Several published “beats GPT” or “beats Claude” tables within hours of launch. None is a primary source, and most are reading numbers off Google’s own graphics. Treat them as rumour until a lab you trust reproduces them.

Here’s a bake-off that produces a number you can defend:

1. Pull 30 to 50 real tasks from your production logs, weighted toward the agentic work you actually run.
2. Run each candidate model at the effort level you’d ship with, and log total tokens per task rather than per call.
3. Score outcomes with one rubric across all models, and have a human check a sample of the grades.
4. Divide spend by successful tasks. A cheaper unit rate can still lose on transcript length.
5. Re-run after any model update. Three Flash releases in six weeks means results age in weeks.

That is roughly the shape of the model comparison we run inside an AI integration audit, and the ranking it produces rarely matches the launch-day leaderboard.

Where Gemini 3.8 Flash Still Falls Short and When to Choose Gemini Pro Instead

Gemini 3.8 Flash falls short on output modality, effort control, knowledge freshness in some domains, and token economy, and Google’s own positioning keeps Gemini 3 Pro-tier models as the reasoning frontier. Choose Pro when a task is hard enough that Flash’s extra retries still miss. Stay on 3.7 Flash when the token bill is the constraint.

The concrete limits, per Google’s documentation as of September 2, 2026:

• Text-only output. The model reads images, video, audio, and PDF, but it can’t generate images or speech natively.
• No “minimal” thinking level. The floor is “low”, and Google’s Gemini 3 developer guide describes the levels as “relative allowances for thinking rather than strict token guarantees”, so you can’t pin a fixed reasoning budget.
• A split knowledge cutoff. Most domains reach March 2026, but some stop at January 2025, and Google doesn’t list which ones. Test your domain prompts before assuming freshness.
• Higher token use per task from the extra reasoning steps and repeated tool calls, most visibly at high effort.
• A 10.0% score on Harvey’s Legal Agent Benchmark that Google describes as the top result. Leading a benchmark at 10% tells you the benchmark is brutal, and it tells you the model still fails nine legal-agent tasks in ten.

That last point deserves emphasis. A “works harder” model amplifies whatever plan it starts with. In our own agent deployments the ugliest failure mode is a plan that’s wrong on step one and then gets executed with great diligence for forty tool calls. More persistence helps when the model’s first read of the problem is right. When it isn’t, you pay for a longer wrong answer.

Decision rules that hold up under that constraint:

• Efficiency-first workloads (classification, extraction, short tool chains at scale): stay on 3.7 Flash, which Google says remains fully supported for exactly this.
• General agents and long-horizon coding where a few extra retries are cheap relative to the outcome: move to 3.8 Flash at medium effort, and measure.
• Tasks where Flash-tier models plateau even at high effort, or where the cost of a confident wrong answer is high: use Gemini 3 Pro-tier models, which Google still positions as the reasoning frontier.

Google itself frames 3.8 Flash as a model for “cost-effective scaling of general-purpose, production-ready agents”. Read that phrase closely. It says scaling, and it says general-purpose. It does not say hardest.

Which Workloads Are the Best Fit for Gemini 3.8 Flash

Gemini 3.8 Flash fits best on long-running coding agents, multi-step analysis in finance and legal work, business workflow automation, and bulk processing of long documents and video, because those are the tasks where extra reasoning steps and repeated tool calls pay for themselves. The common thread is a task with many steps and a checkable outcome. Single-shot chat is fine on it too, but you’d be paying for persistence you never use.

Google’s own launch demos, published September 2, 2026, show the shape of the intended workload. One prompt with a looping instruction in Google Antigravity produced a playable 3D castle game with puzzles and textures generated by Nano Banana. Another single prompt built a working DOS-style version of Google Maps with directions and Street View. A third assembled a topographic map of famous sites from U.S. Geological Survey datasets, with live cross-sections. And in AI Studio, a project called Hardware Anatomy generates Three.js teardowns of physical devices with an exploded-layer slider. Every one of those is a build-and-iterate job, where the model keeps working until the thing runs.

Here’s how the fit breaks down by workload:

Two rows carry an asterisk. The legal row needs a human in the loop, given the absolute score. And the video row rides on a Flash-family feature that Google’s changelog listed for 3.7 Flash, 3.6 Flash, and 3.5 Flash-Lite on September 1, 2026, so confirm it’s active on 3.8 in your own account before you plan around the token savings.

If you’re scoping an agent build around this model, the highest-return targets are the ones where a task currently takes a human twenty minutes of clicking through systems. That’s the profile we look for in AI agent development work, and it’s where a model that keeps going after a failed tool call earns its output tokens.

How to Access Gemini 3.8 Flash Through the API, Vertex AI, and AI Studio

You can access Gemini 3.8 Flash through the Gemini API (from Google AI Studio, Android Studio, Antigravity, or Stitch), through the Gemini Enterprise Agent Platform for enterprise accounts, through the Gemini app and AI Mode for paying consumers, and through the Fairwind Program for the Cyber variant. All four routes opened on September 2, 2026.

Developers. The Gemini API is the main path. Google’s launch post lists Google AI Studio and Android Studio as the places to start building, Google Antigravity for agent-first workflows, and Stitch for generating UIs. Once you’re calling the model, three settings matter most:

• thinking_level: set to low, medium, or high. Google’s Gemini 3 developer guide describes these as relative allowances for reasoning, so expect variation in actual token use at each level.
• Inference tier: Standard by default, Batch for asynchronous jobs at half price, Priority for latency-sensitive traffic at a premium, plus Flex.
• Tooling flags: function calling, code execution, structured outputs, URL context, file search, search grounding, Google Maps grounding, caching, and computer use in preview.

Enterprises. Gemini 3.8 Flash is listed in the Gemini Enterprise Agent Platform model catalog, the platform formerly known as Vertex AI. If your organisation already runs Gemini models there, 3.8 Flash shows up as another catalog entry with the same governance and billing setup.

Consumers. Google AI Pro and Ultra subscribers get 3.8 Flash in the Gemini app, in AI Mode in Google Search, and in Gemini in Google Sheets. Free-tier availability for the app isn’t spelled out in the launch materials.

Defenders. Gemini 3.8 Flash Cyber is invitation-shaped. Google’s Fairwind Program gives prioritised access to trusted government authorities, critical infrastructure operators, and software maintainers, and you apply rather than sign up. The model ships with more permissive cybersecurity mitigations than the general 3.8 Flash, which is the whole reason for the gate.

One practical note from the API side. The model ID and the tier you pick live in different places (the request body versus the endpoint or billing config), and it’s easy to run a Batch-priced evaluation and then ship on Standard without noticing the 2x difference. Check both before quoting a cost to anyone.

What Teams Should Check Before Migrating Production Traffic to Gemini 3.8 Flash

Before moving production traffic to Gemini 3.8 Flash, teams should measure real token cost per task at each thinking level, regression-test their agent loops, confirm the knowledge cutoff covers their domain, budget for the January 1, 2027 price doubling, and stage a canary rollout. Same price per token as 3.7 Flash is the easy part. Everything else needs a number from your own traffic.

The checklist we’d run, in order:

1. Log tokens per completed task, at low, medium, and high thinking. Run a fixed sample of production prompts through 3.7 Flash and 3.8 Flash and compare total transcript length, including every tool call. Google says 3.8 may use more tokens to reach a better answer. Find out how many more, on your prompts.
2. Regression-test every agent loop and tool schema. A model that retries more can also retry a broken tool call more. Watch for changed loop counts, changed stop conditions, and any tool the model now calls that it previously skipped.
3. Re-run your refusal and tone suite. Google’s model card reports unjustified refusals up 1.1 points against 3.7 Flash. Small, but if your product sits in healthcare or finance, a new refusal on a legitimate prompt is a support ticket.
4. Probe the knowledge cutoff for your domain. Most domains reach March 2026, some stop at January 2025, and Google doesn’t say which. Ask the model about events from the first half of 2026 in your field and see what it knows.
5. Model the 2027 bill now. Standard rates go from $0.75 and $3.75 to $1.50 and $7.50 per million tokens on January 1, 2027. If the migration case only closes at introductory pricing, it doesn’t close.
6. Match workloads to Batch or Priority. Overnight document runs belong on Batch at $0.375 and $1.875. Interactive agents may need Priority. Mixed traffic on Standard is usually the most expensive way to run both.
7. Canary 5 to 10 percent of traffic behind the same eval harness. Compare success rate and cost per success against the 3.7 Flash baseline for at least a week before flipping the default.

The surprise in our own model swaps is almost never the input side. It’s output tokens. A model that reasons longer writes a longer transcript even when the final answer is word-for-word identical, and that line item only appears after the migration if nobody measured it before. Measure it before.

Frequently Asked Questions About the Gemini 3.8 Flash Launch

The most common questions about the Gemini 3.8 Flash launch come down to price, timing, limits, and how it differs from the model it replaced. Short answers below, each one standalone.

Is Gemini 3.8 Flash free to use?

Not on the API. Gemini 3.8 Flash costs $0.75 per million input tokens and $3.75 per million output tokens on the standard tier as of September 2, 2026, with Batch at half that rate. Consumers get it inside the Gemini app, AI Mode, and Google Sheets as part of a Google AI Pro or Ultra subscription. Google AI Studio is the cheapest place to try it before you commit to anything.

When does the Gemini 3.8 Flash introductory pricing end?

The introductory pricing ends on December 31, 2026. From January 1, 2027, standard rates double to $1.50 per million input tokens and $7.50 per million output tokens, and the Batch and Priority tiers double with them. The same end date applies across the whole Flash 3.x family, so switching between 3.5, 3.6, 3.7, and 3.8 Flash doesn’t buy extra time at the discount.

What is the context window of Gemini 3.8 Flash?

Gemini 3.8 Flash accepts up to 1,048,576 input tokens and returns up to 65,536 output tokens per request. That’s roughly 1M in and 64K out. The output ceiling is the one to plan around for long agent runs, since reasoning and tool-call transcripts count against it.

Is Gemini 3.8 Flash better than Gemini 3.7 Flash?

Yes, on Google’s published benchmarks. Gemini 3.8 Flash scored about 71.0% on DeepSWE in September 2026 against 65.3% for 3.7 Flash on the related v1.1 version in August 2026, and Google reports it leads on Vals Finance Agent V2 and Harvey’s Legal Agent Benchmark. The trade is that it may spend more tokens per task to get there, so 3.7 Flash remains the better pick when cost per call is the constraint.

What is Gemini 3.8 Flash Cyber and who can use it?

Gemini 3.8 Flash Cyber is a cybersecurity variant of Gemini 3.8 Flash, built for finding vulnerabilities and generating patches, and it’s available only through Google’s Fairwind Program. Access goes to trusted government authorities, critical infrastructure operators, and software maintainers by application. It shares the base model with 3.8 Flash but ships with more permissive cyber mitigations, which is why it’s gated.

Does Gemini 3.8 Flash support image or audio output?

No. Gemini 3.8 Flash outputs text only. It accepts text, images, video, audio, and PDF as input, but native image generation and speech output are outside its scope as of September 2, 2026.

What is the knowledge cutoff for Gemini 3.8 Flash?

The knowledge cutoff is March 2026 for most domains, with some domains limited to January 2025. Google doesn’t publish which domains fall under the earlier date. If your prompts depend on events from late 2025 or early 2026, test them directly before trusting the model’s recall.

Where to Start with Gemini 3.8 Flash This Week

Start with Gemini 3.8 Flash in Google AI Studio this week, and let a single representative task tell you whether the upgrade pays. Pick one agentic job you already run on 3.7 Flash, something with tool calls and a checkable result, and run it through both models at medium thinking. Log total tokens per completed task, including every retry. Then divide spend by successes.

That one number decides the next step. If 3.8 Flash wins on cost per success, canary a slice of traffic and hold it there for a week. If it loses, stay on 3.7 Flash and revisit at the next point release, which on Google’s current cadence may be weeks away.

Either way, get the comparison done before December 31, 2026, while both models still bill at the introductory rate.