Gemini 3.8 Flash is out: Google’s benchmarks, pricing, and Antigravity rollout

Google has released Gemini 3.8 Flash as a generally available model with the stable API ID gemini-3.8-flash. The launch is unusually consequential for agent builders: Google’s own comparison table shows 3.8 leading several coding, professional, multimodal, and reasoning evaluations, while the model keeps the same introductory token price as Gemini 3.7 Flash. Google DeepMind also says 3.8 is rolling out in Antigravity.

The decision for developers is not simply “switch to the newest model.” 3.8 is designed to work harder on difficult tasks, which can mean more thinking and tool-call tokens. The sensible first move is a canary: run the same agent tasks in 3.7 and 3.8, at the same effort level, and record success, retries, time, token use, and human cleanup.

Three takeaways

1. The release is official and GA. Google’s model page lists gemini-3.8-flash as stable, while its latest-model guide calls it generally available and production-ready. Read Google’s model documentation.
2. The benchmark lead is real but selective. Google’s published table puts 3.8 ahead on Vals Finance Agent v2, Harvey’s Legal Agent Benchmark, Terminal-bench 2.1, CharXiv Reasoning, HLE-Verified, and LABBench2, but not on every row.
3. Antigravity is part of the launch, not a rumor. Google DeepMind says 3.8 Flash is rolling out in Antigravity, and Antigravity’s own model page lists it in the selector across its documented plan tiers.

The release Google actually announced

Google’s launch article introduces two variants: Gemini 3.8 Flash for long-horizon software engineering, autonomous agents, and complex enterprise workflows; and Gemini 3.8 Flash Cyber for trusted defenders working on cybersecurity.

For general developers, the important line is the stable endpoint: gemini-3.8-flash. Google’s model page lists text, image, video, audio, and PDF inputs; text output; a 1,048,576-token input limit; a 65,536-token output limit; and support for caching, code execution, function calling, structured outputs, Search and Maps grounding, URL context, and computer use in preview. Thinking is configurable at low, medium, or high. minimal is not supported.

This is a Flash release, not the arrival of a new Pro model. Its pitch is a higher-quality workhorse that remains fast and comparatively cheap.

Google’s official benchmark table

The table below transcribes selected rows from Google’s official launch comparison. It is a useful snapshot of what Google believes 3.8 can do; it is not an EyesTech benchmark rerun.

Source: Google’s Gemini 3.8 launch article. Google links a separate evaluation-methodology document.

The interactive explorer embedded with this article lets readers switch among DeepSWE v1.1, Terminal-bench 2.1, CharXiv Reasoning, HLE-Verified, and OSWorld-2.0. It uses the same Google-published values shown above; it does not turn vendor-reported scores into an independent test.

The shape of the results matters more than a single headline score. 3.8 Flash is just behind Claude Opus 5 on DeepSWE, leads the financial and legal agent rows, edges Opus on Terminal-bench 2.1, and posts narrow wins on HLE-Verified and CharXiv. But Opus 5 remains far ahead on Terminal-bench 4.0 and OSWorld-2.0, while GPT-5.6 Sol leads the GDP.PDF row in Google’s full chart. That is a strong workhorse profile, not a universal frontier-model victory.

Google also says 3.8 Flash reaches 54.9% on HLE-Verified. That is the kind of number worth tracking, but it should be read as a vendor-published evaluation result with the model’s configuration and effort level attached—not as a promise that every production prompt will improve by the same margin.

The additional rows sharpen the picture. Gemini 3.8 trails GPT-5.6 Sol on GDP.PDF and Claude Opus 5 on the human-solvable BioMysteryBench set, but it leads the displayed competitors on the human-difficult BioMysteryBench cases. Those are useful evidence of task-shape variance, not a single intelligence ranking.

Why the benchmark can cost more than the price suggests

Google’s central product claim is that 3.8 “works harder.” On difficult tasks, the model takes smaller reasoning steps, calls tools iteratively, and verifies its work. That behavior is precisely what an agent builder wants when a task spans a repository, a browser, a terminal, and a test suite. It is also a billing variable.

Google’s latest-model guide gives three effort levels:

• Low: lower latency and lower token overhead for incident response, chat, drafting, and fast analysis.
• Medium: the default balance for complex code and agentic work.
• High: maximum reasoning and tool orchestration for difficult multi-step tasks.

The model does not expose a minimal level. Google’s migration notes also say to replace older thinking_budget settings with thinking_level, remove deprecated sampling parameters such as temperature, top_p, and top_k, and remove candidate_count for Gemini 3 and later.

That makes effort-level logging part of a fair comparison. If a 3.8 run succeeds where 3.7 loops, the extra tokens may be good economics. If both succeed but 3.8 spends twice as many tokens verifying a trivial change, the model is not automatically the cheaper route.

Pricing: the same introductory rate as 3.7

Google’s current paid standard rate for 3.8 Flash is:

The full Google API pricing page also lists batch and flex options at lower rates, context caching, grounding charges, and separate free/paid data-use terms.

For a transparent illustration, 1M input tokens plus 200K output tokens costs:

(1 × $0.75) + (0.2 × $3.75) = $1.50

The same token shape costs $3.00 at the documented post-2026 standard rate. These are token-only examples. A real agent bill can add tool-generated context, grounding requests, retries, cache storage, taxes, and foreign-exchange costs.

Gemini 3.8 Flash is now in Antigravity

This is the part many coding-agent users will notice first. In a Google DeepMind post on X, Google says Gemini 3.8 Flash is “rolling out now” in Antigravity and via the API in Google AI Studio and Android Studio.

Google’s Antigravity model documentation now lists Gemini 3.8 Flash as a selectable reasoning model. The same page marks it available for Free/Google AI Plus, Google AI Pro, Google AI Ultra, and Enterprise plans. Google’s API guide goes further: the managed Antigravity agent and the Antigravity SDK use Gemini 3.8 Flash by default.

Google’s launch demos show 3.8 Flash building a 3D wizard-castle game from a prompt, creating a playable DOS version of Google Maps, and generating a USGS-backed topographic map in Antigravity. Those examples demonstrate the intended workflow—plan, write code, run it, inspect the result, and iterate—but they are demonstrations, not independent reliability tests.

The 3.js demo is real—but it is an AI Studio visualizer

Google DeepMind’s Gemini model page identifies Hardware Anatomy as a Three.js visualizer built in Google AI Studio. That makes it a useful demonstration of Gemini 3.8 Flash generating and iterating on a rich browser interface. It does not prove physical accuracy, production reliability, or that every 3D workflow runs in Antigravity. EyesTech embeds the official demo media below and preserves the Google attribution.

If 3.8 appears in your Antigravity model menu, start with a fixed task set:

1. Use the same repository and prompt with 3.7 and 3.8.
2. Keep the effort level explicit and identical.
3. Record tool calls, wall-clock time, output tokens, test-pass rate, retries, and human cleanup.
4. Keep a 3.7 fallback until the difference survives several fresh tasks.

Official launch posts worth bookmarking

Google’s social posts are useful because they confirm the release surface in the company’s own words:

• Google AI: “We’re introducing Gemini 3.8 Flash” — highlights reasoning effort, multi-step work, and an Antigravity game demo.
• Google AI: Gemini 3.8 availability — lists Gemini, Search, Sheets, AI Studio, Android Studio, Antigravity, Stitch, and Gemini Enterprise surfaces.
• Google DeepMind: two new Gemini models — separates 3.8 Flash from the cyber-specific variant.
• Google DeepMind: Antigravity and API rollout — confirms the Antigravity rollout and consumer/API access routes.

What this changes

For coding-agent teams, Gemini 3.8 Flash is worth testing immediately. The official numbers suggest a meaningful improvement over 3.7 on long-horizon coding, professional agent work, and several reasoning tasks, while the introductory price is unchanged.

For high-volume systems, the model’s effort controls are as important as its benchmark rank. Use low effort where latency and predictable spend dominate; reserve medium or high for tasks where better first-pass accuracy and fewer failed loops pay for the extra tokens.

For India-based teams, the published price is a USD list price—not a final rupee cost. Google’s public pages reviewed for this launch do not state India-specific tax handling, latency, data residency, or an enterprise SLA. Confirm those details against the billing account and contract before making a procurement decision.

Gemini 3.8 Flash Cyber is a separate, gated model. Google says it is available to trusted defenders through Fairwind, with claims including more than 70% success on an internal real-world vulnerability-discovery benchmark and a 47.2% pass@1 on CWE-Bench. Those figures should not be treated as general 3.8 Flash results.

Who should care

• Agent and coding-tool builders: run a 3.7-versus-3.8 canary now, especially on multi-file changes and tool-heavy tasks.
• Cost-sensitive API teams: model the full token bill, including thinking tokens, and recheck the January 2027 price step.
• Antigravity users: select 3.8 explicitly, then compare it with 3.7 on your own repository before changing defaults.
• Enterprise and India deployments: verify region, support, data-use, tax, and SLA terms; they are not established by the public launch announcement alone.
• Cybersecurity teams: treat Flash Cyber as a separate access-controlled program, not a public model alias.

Update note

Checked 2 September 2026. This article uses Google’s launch post, live developer documentation, Antigravity documentation, and official Google/DeepMind X posts. EyesTech did not independently rerun the benchmarks or publish a production test. Recheck the model, pricing, Antigravity, and evaluation pages before acting on this article after launch day.

Suggested internal links: Why China Is Betting on “Flash” AI Models and Claude Fable 5.1: benchmarks, pricing, and the cache-first case for agentic work.

If you adopt 3.8, keep the 3.7 canary and publish the first cost-per-success result to the EyesTech Model Reality Index.