TL;DR: Google on Sept. 2, 2026 introduced Gemini 3.8 Flash and Gemini 3.8 Flash Cyber, its third Flash release in six weeks, priced at $0.75 per million input tokens and $3.75 per million output tokens on introductory terms.

Third Flash release in six weeks

Google introduced Gemini 3.8 on Sept. 2, 2026, according to the company announcement titled Introducing Gemini 3.8 Flash and 3.8 Flash Cyber.

The announcement was credited to Tulsee Doshi, Senior Director of Product Management, and Raluca Ada Popa, Gemini Security Lead at Google DeepMind.

The company described the launch as building on the momentum of 3.7 Flash from three weeks ago and marking the third Flash release in only six weeks.

Google described Gemini 3.8 as its best reasoning and coding model yet, offered at the same speed and low cost as 3.7, according to that Google announcement.

Coverage by Ars Technica framed the release as the third Flash model release in just six weeks.

That Ars Technica report, published Sept. 2, 2026 by Ryan Whitwam, noted that Google had not released a frontier-level Gemini Pro AI model since early 2026.

The same report said a promised Gemini 3.5 Pro now looked less likely to appear.

Two variants on shared intelligence

Gemini 3.8 introduces two variants, according to Google's announcement.

Gemini 3.8 Flash is described as the company's most intelligent workhorse model, for software engineering, agentic tasks, and multi-step reasoning in specialized domains.

Gemini 3.8 Flash Cyber is described as the company's most capable cybersecurity model, for vulnerability detection and automated patching.

Ars Technica described the standard Flash as a workhorse model for agentic tasks and software development.

That publication described Gemini 3.8 Flash Cyber as running on the same foundations but tuned for vulnerability detection and mitigation.

Google said both releases are powered by the same foundational intelligence, according to the Google announcement.

The company said that shared core was further accelerated by long-running agentic loops designed to recursively evaluate and refine the underlying models.

Google said coding and reasoning gains across that shared core were driven by innovations including rigorous training in cybersecurity.

Introductory API pricing

Gemini 3.8 Flash is available at the same introductory price as 3.7 Flash, according to Google.

The introductory price is $0.75 per million input tokens and $3.75 per million output tokens.

Google's DeepMind model card titled Gemini 3.8 Flash - Model Card lists the same input price of $0.75 per 1M tokens without caching and the same output price of $3.75 per 1M tokens.

That model card lists regular pricing in parentheses as $1.50 for input and $7.50 for output.

Ars Technica reported that API access is available at an introductory rate through the end of the year at $0.75 per million input tokens and $3.75 per million output tokens.

The same Ars Technica report listed the regular price as $1.50 per million input tokens and $7.50 per million output tokens.

That report said new models would likely be available before the price changes.

For comparison, the DeepMind model card lists Claude Opus 5 at $5.00 input and $25.00 output, Claude Sonnet 5 at $2.00 input and $10.00 output, GPT-5.6 Sol at $4.00 input and $20.00 output, and GPT-5.6 Terra at $2.00 input and $12.00 output.

Coding and long-horizon engineering scores

Google said Gemini 3.8 Flash delivers substantial gains from 3.7 Flash, often approaching the performance of higher-cost frontier models.

On DeepSWE v1.1 for long-horizon software engineering, Google said 3.8 Flash outperforms most larger frontier models in autonomously solving complex engineering problems end to end at a fraction of the cost.

The DeepMind model card lists DeepSWE v1.1 at 73.7% for Gemini 3.8 Flash and 65.3% for Gemini 3.7 Flash.

The same table lists Claude Opus 5 at 74.0%, Claude Sonnet 5 at 53.8%, GPT-5.6 Sol at 72.7%, and GPT-5.6 Terra at 69.6% on DeepSWE v1.1.

Ars Technica reported that Gemini 3.8 Flash is now at the top of the DeepSWE leaderboard for solving complex software engineering problems at a lower cost at the discounted rate.

That report said benchmark numbers showed marginal improvement over Gemini 3.7 Flash in most tests, with larger gains in coding evaluations.

On Terminal-bench 2.1 for agentic terminal coding, the DeepMind model card lists 89.4% for Gemini 3.8 Flash and 85.8% for Gemini 3.7 Flash.

The same benchmark lists Claude Opus 5 at 89.1%, Claude Sonnet 5 at 80.4%, GPT-5.6 Sol at 88.8%, and GPT-5.6 Terra at 87.4%.

On Terminal-bench 4.0 for general agent capabilities, the model card lists 19.1% for Gemini 3.8 Flash and 11.2% for Gemini 3.7 Flash.

That table lists Claude Opus 5 at 51.8%, Claude Sonnet 5 at 12.4%, GPT-5.6 Sol at 37.3%, and GPT-5.6 Terra at 23.6% on Terminal-bench 4.0.

Finance, legal and expert-reasoning results

Google said 3.8 Flash exhibits dependability for enterprise autonomy across specialized knowledge domains.

In quantitative and professional fields requiring advanced analysis and reporting, Google said 3.8 Flash outperforms 3.7 Flash and other frontier models in benchmarks including Vals Finance Agent V2 and Harvey's Legal Agent Benchmark.

The DeepMind model card lists Vals Finance Agent v2 for financial analyst tasks at 61.4% for Gemini 3.8 Flash and 59.0% for Gemini 3.7 Flash.

The same table lists Claude Opus 5 at 58.6%, Claude Sonnet 5 at 53.9%, GPT-5.6 Sol at 53.8%, and GPT-5.6 Terra at 54.4% on that finance benchmark.

On Harvey's Legal Agent Benchmark for complex legal workflows, the model card lists an all-pass rate of 10.0% for Gemini 3.8 Flash and 8.8% for Gemini 3.7 Flash.

That benchmark lists Claude Opus 5 at 6.7%, Claude Sonnet 5 at 5.0%, GPT-5.6 Sol at 2.5%, and GPT-5.6 Terra at 0.8%.

Google reported 54.9% on HLE-Verified for multi-step reasoning across STEM, humanities, and professional fields.

The model card lists HLE-Verified for multidisciplinary expert reasoning at 54.9% for Gemini 3.8 Flash and 53.6% for Gemini 3.7 Flash.

That table lists Claude Opus 5 at 54.4%, Claude Sonnet 5 at 31.0%, GPT-5.6 Sol at 54.5%, and GPT-5.6 Terra at 51.1% on HLE-Verified.

On GDPVal-AA v2 for knowledge work, the model card lists an Elo score of 1545 for Gemini 3.8 Flash and 1482 for Gemini 3.7 Flash.

That benchmark lists Claude Opus 5 at 1824, Claude Sonnet 5 at 1584, GPT-5.6 Sol at 1710, and GPT-5.6 Terra at 1528.

Computer use, documents and science benchmarks

On OSWorld-2.0 for agentic computer use, Google's model card lists a partial score with batch tool enabled of 59.0% for Gemini 3.8 Flash and 50.6% for Gemini 3.7 Flash.

The same OSWorld-2.0 table lists Claude Opus 5 at 75.4%, Claude Sonnet 5 at 42.6%, GPT-5.6 Sol at 62.6%, and GPT-5.6 Terra at 50.2%.

Ars Technica reported that Gemini 3.8 Flash improved over 3.7 Flash in the OSWorld-2.0 test of agentic computer use but remained behind market leader Claude Opus.

On GDP.PDF for expert PDF document comprehension, the model card lists an all-pass rate of 35.0% for Gemini 3.8 Flash and 34.0% for Gemini 3.7 Flash.

That benchmark lists Claude Opus 5 at 37.0%, Claude Sonnet 5 at 28.0%, GPT-5.6 Sol at 40.0%, and GPT-5.6 Terra at 29.0%.

On CharXiv Reasoning for information synthesis from complex charts without tools, the model card lists 86.2% for Gemini 3.8 Flash and 84.5% for Gemini 3.7 Flash.

That table lists Claude Opus 5 at 83.7%, Claude Sonnet 5 at 70.1%, GPT-5.6 Sol at 85.8%, and GPT-5.6 Terra at 85.9% on CharXiv Reasoning.

On LVBench for long video understanding, the model card lists 87.8% agentic and 87.1% static for Gemini 3.8 Flash, compared with 85.4% for Gemini 3.7 Flash.

That benchmark lists Claude Opus 5 at 75.4%, Claude Sonnet 5 at 68.5%, GPT-5.6 Sol at 82.1%, and GPT-5.6 Terra at 78.9%.

On BioMysteryBench for bioinformatics research workflows, the model card lists 88.8% human-solvable and 56.5% human-difficult for Gemini 3.8 Flash.

The same table lists 87.1% human-solvable and 43.5% human-difficult for Gemini 3.7 Flash on BioMysteryBench.

On LAB-Bench2 for biology real-world research tasks, the model card lists 86.2% for Gemini 3.8 Flash and 82.1% for Gemini 3.7 Flash.

Designed to work harder with effort controls

Google said performance gains stem from a design choice in which 3.8 Flash works harder on complex tasks.

The company said the model executes extra reasoning steps and calls tools iteratively on complex tasks.

Google said the model might use more tokens to maximize performance, especially at higher effort levels.

The DeepMind model card says the model might use more tokens to maximize performance, especially at higher effort levels.

For applications where compute efficiency is the primary constraint, Google said developers can use lower effort levels to minimize token overhead.

Google said developers can continue to rely on Gemini 3.7 Flash, which remains fully supported for efficiency-first workloads.

The model card describes Gemini 3.8 Flash as based on Gemini 3.7 Flash and as supporting customizable effort levels to control quality, cost, and latency.

Inputs, outputs and distribution

The DeepMind model card lists inputs as text strings, images, audio, and video files, with a token context window of up to 1M.

The same card lists outputs as text, with a 64K token output.

The model card lists the knowledge cutoff date for Gemini 3.8 Flash as March 2026, while noting users may experience knowledge limited to January 2025 in some domains in line with the Gemini 3 model family.

The card lists distribution in the Gemini app, Gemini Enterprise Agent Platform, Google AI Studio, Gemini API, Google AI Mode, and Google Antigravity.

Google's announcement highlighted builds made with Gemini 3.8 Flash in Google Antigravity and Google AI Studio, including a wizard castle game, a DOS version of Google Maps, a topographic map using U.S. Geological Survey datasets, and a Hardware Anatomy Three.js visualizer.

Flash Cyber for defenders

Gemini 3.8 Flash Cyber is available to a set of trusted defenders via the Fairwind Program, according to Google.

Google described Flash speed and cost as enabling quick iteration in cybersecurity work.

On CyberGym for autonomous vulnerability discovery, Google said Gemini 3.8 Flash Cyber demonstrates frontier-level performance.

Google said it surpasses both 3.5 Flash Cyber and significantly larger frontier models on that benchmark.

Google also evaluated the Cyber model on an internal benchmark covering vulnerabilities across codebases spanning 20 programming languages.

On that internal benchmark, Google said the model reaches a success rate exceeding 70%, representing a leap over previous models.

Google said it invested in vulnerability fixing from the start and prioritized fixing over offensive capabilities including exploitation.

On CWE-Bench, run by Collinear for patching capabilities, Google reported a pass@1 of 47.2% for Gemini 3.8 Flash Cyber, compared with 47.8% for a leading frontier model at significantly lower cost.

Chrome, cloud and partner testing

Google said it is already using Gemini 3.8 Flash Cyber to secure code across Google.

The Chrome Security team found that 3.8 Flash Cyber produced 2.6 times more correct patches to vulnerabilities in Chrome than the best commercial models that are much larger, according to Google.

Ars Technica reported the same Chrome finding as a 2.6x increase in patch accuracy with the new model.

Ars Technica also reported that Google's Cloud team said 3.8 Flash Cyber found a critical vulnerability in two hours.

That report cited statements from partners including Wiz and Palo Alto Networks about Gemini 3.8 Flash Cyber.

Availability on Sept. 2

Gemini 3.8 Flash was available across the Google ecosystem starting Sept. 2, 2026, according to Ars Technica.

Gemini 3.8 Flash Cyber is limited to trusted testers and governments, according to that report.

Google described Cyber availability as access for a set of trusted defenders through the Fairwind Program.

Ars Technica reported that access to Gemini 3.8 Flash in the Gemini app requires a Pro or Ultra subscription.

The same report said AI Studio remains available for users who want to tinker with the model without charge.