Gemini 3.8 Flash in Antigravity: Complete Guide

What Shipped

On September 2, 2026, Google shipped Gemini 3.8 Flash into Antigravity — the agent-first IDE where models are selected per task rather than per project. The headline numbers: $0.75/$3.75 per 1M tokens (input/output), 54.9% on HLE-Verified, and the 1M-token context window that was previously a 3.1-Pro-tier feature inside the IDE. The announcement post positions it as the “most intelligent workhorse model” built for long-horizon coding and autonomous agents — workhorse meaning the default model you leave selected for most of a workday, not the reasoning-max tier you invoke for hard problems.

The calendar math is the quiet story. This is the third Flash release in six weeks: 3.6 Flash, then 3.7 Flash (positioned for compute-efficiency-first workflows, and still available), now 3.8 Flash. Cadence like this changes how you should treat any single launch post — the model you read about today is the baseline you’ll be compared against in the next post, so switching decisions should weight evidence over launch language.

3.8 Flash is available in Antigravity (Pro and Ultra tiers), the Gemini API, AI Studio, and Gemini Enterprise. If you are on the API rather than the IDE, the pricing page carries the same numbers — and 3.7 Flash remains selectable there for compute-efficiency-first work.

The Announcement

The launch came from the Antigravity account, and the wording is worth reading precisely because the claims are measurable: workhorse positioning, long-horizon coding, autonomous agents.

Google Developers followed the next day with the same anchor — complex developer-tool building — which matters because it scopes the claim to a domain rather than a generic capability. A model that makes complex developer-tool building easier is a claim you can test against your own repo; a model that is broadly smarter is a claim you can only take on faith.

The quota signal came from Mohan ( @_mohansolo ), Gemini on Antigravity: quota resets across Antigravity because TPUs are melting with 3.8 Flash usage. The phrasing is casual; the operational meaning is not. Launch-week usage is exceeding provisioned capacity, which is exactly the window where your own benchmark runs compete with everyone else’s for the same silicon.

The topjohnwu reply (the Magisk developer) raised the agy gyro topic in the same thread — and that gap is covered in its own section below because it is the one part of this launch that affects developers who live in the CLI rather than the IDE.

Benchmark Evidence

The launch post leads with three vendor-reported benchmark beats and one verified multi-step reasoning score. Read them with the caveat column in mind: these are launch-day claims from the vendor, and the benchmark names are agent-domain-specific rather than generic-code benchmarks.

The HLE-Verified number deserves its own read. Humanity’s Last Exam at 54.9% is a verified multi-step reasoning score across STEM, humanities, and professional fields — verified meaning the answers were checked rather than self-reported. For context, 3.7 Flash positioned lower on the same benchmark, and the workhorse tier historically trails the reasoning-max tier by double digits. If your workload is multi-step reasoning rather than raw code generation, run your own eval before reassigning the default model slot; if it is agentic coding, the launch benchmarks are closer to your domain and worth taking more seriously.

Pricing and Quotas

3.8 Flash launches at the same price as 3.7 Flash: $0.75 per 1M input tokens and $3.75 per 1M output tokens. No launch premium. That is the single most important switching fact in this post — the cost side of the equation is unchanged, so the decision reduces to capability and quota behavior rather than budget.

Inside Antigravity, quotas are tier-based (Pro and Ultra) and metered per model. The launch-week quota resets documented above are the practical caveat: during the reset window, your long benchmark runs share silicon with launch traffic. The Antigravity changelog is the source to watch for quota policy changes; the pricing page is the source for API-side numbers.

Switching Models in Antigravity

In the IDE, model selection is per agent run: open the model selector (or run /model), pick Gemini 3.8 Flash, and the current run uses it. The 2.2.1 build added dynamic resizing to the selector so long model names stay readable — a small change that matters when the model list grows this fast.

Selection is per run, not global. Scripted or CI runs do not always inherit your last manual selection, so if you rely on 3.8 Flash in automation, set it in the run configuration explicitly and verify it sticks before you trust it. This is the single most common failure mode people hit with fast model cadence: the model they benchmarked is not the model their automation actually used.

# In the Antigravity IDE
/model gemini-3.8-flash        # switch the current run

# In the agy CLI (see the registry gap below)
agy model set gemini-3.8-flash # only if your agy version knows the model

Thinking Levels

The launch post introduces thinking levels as a first-class control: developers can adjust reasoning depth and latency per task, and 3.7 Flash remains available for workflows that prioritize compute efficiency. Practically, the levels trade output tokens for reasoning depth — higher thinking burns more output quota per turn, which is the resource your Pro/Ultra plan actually meters.

The honest guidance: there is no official per-level token table, and launch posts do not publish one. If you are quota-sensitive, benchmark one real task at each level against your own workload before committing a default. The “default to High, drop for mechanical work” heuristic is a starting point, not a source-backed rule.

Official Demos

The Antigravity blog post ships three official demo videos with the launch. The design-system demo is the clearest view of what workhorse-tier multimodal output looks like inside the IDE — minimal input to a full design system with implementations across 12 frameworks and an interactive theme editor:

The codebase-analysis demo is the one to watch if your work is closer to architecture than UI — the agent explores, documents, and produces an interactive artifact from an in-depth research pass over a real codebase:

Both demos are Google-produced and run on Google-chosen hardware and inputs. Treat them as capability ceilings demonstrated under ideal conditions, not as benchmarks of what you will see on your repository on launch day.

The agy CLI Registry Gap

Here is the part the launch posts do not cover. The agy CLI (Antigravity’s command-line surface) uses a static model registry when configured with a Gemini API key. New model releases do not appear there until a CLI update ships — meaning on launch day, CLI users configured via API key could not route to 3.8 Flash even though the IDE could.

topjohnwu (the Magisk maintainer) shipped a fix the same day: agy-gyro v0.1.2, a local retry proxy that adds model redirection so newly released models like gemini-3.8-flash route immediately instead of waiting for an agy update. It also stabilizes the CLI against transient Google Gemini API errors, which is its original purpose.

The trade-off is real and worth stating plainly: agy-gyro is a third-party proxy sitting between your CLI and Google’s API. It is open source, the author is well-known in the Android ecosystem, and the code is inspectable — but routing production traffic through it is a trust decision you should make with the repository open in front of you. The alternative is waiting for the official agy update, which for launch-day model access is exactly the gap this tool exists to close.

3.8 Flash Cyber and Fairwind

The same announcement shipped Gemini 3.8 Flash Cyber, a cybersecurity-focused variant with frontier-level vulnerability discovery and automated patching performance. It is not generally available: access runs through the Fairwind Program, Google’s gated track for vetted security researchers and defenders. For most Antigravity developers this is context, not an option — but it signals where Google is taking the Flash line’s specialized variants.

Caveats

All performance figures are vendor-reported. The DeepSWE, Vals Finance, Harvey, and HLE-Verified numbers come from Google’s launch post under Google-chosen configurations. No independent replication existed at publication time. If you are migrating a pipeline, benchmark on your own workloads — ideally the same task at 3.7 and 3.8 with identical thinking levels.

Launch-week quota behavior is volatile. The quota resets above happened during launch week and the reset cadence is not documented. Long benchmark runs in the first week compete with launch traffic for the same silicon.

The agy CLI registry lag is structural. Static registries update on CLI release cycles, not model release cycles. agy-gyro solves it for API-key users; IDE users are unaffected. If you live in the CLI with an API key, check your registry state before assuming a model is unavailable.

Use It If

Use Gemini 3.8 Flash if your work is agentic coding or long-horizon tool use in Antigravity and you want the newest workhorse at unchanged Flash pricing. The 1M context window reaching the Flash tier changes what fits in a single pass, and the launch benchmarks are at least in your domain.

Stay on 3.7 Flash if your priority is compute efficiency per token and your tasks do not need the longer context or the reasoning bump — Google explicitly kept it available for exactly that. And if you are a CLI-first developer on an API key, verify your model registry (or evaluate agy-gyro) before writing any of this into automation.

FAQ

Sources

Official sources first, then community and developer posts. All accessed September 2–9, 2026.

• Gemini 3.8 Flash in Google Antigravity — Antigravity blog (Sep 2, 2026)
• Introducing Gemini 3.8 Flash and 3.8 Flash Cyber — Google blog (Sep 2, 2026)
• Antigravity changelog (versions 2.11.0–2.12.2)
• Gemini API pricing
• @antigravity launch post (Sep 2, 2026)
• @googledevs launch context (Sep 3, 2026)
• @_mohansolo quota resets (Sep 3, 2026)
• @topjohnwu agy-gyro v0.1.2 (Sep 2, 2026)
• agy-gyro repository
• Gemini 3.8 Flash Cyber model page
• Fairwind Program

Related: Antigravity CLI deep dive, Antigravity 2.2.1 & the agy CLI, and Gemini 3.8 Flash: complete guide.