Google Launches Gemini 3.8 Flash, Closes Performance Gap with Flagship Models

Google released Gemini 3.8 Flash and Gemini 3.8 Flash Cyber for trusted cyber defense organizations on September 2, 2026. The standard version supports a 1 million token context window and is optimized for programming, agents, and expert tasks; in Google’s internal tests, it achieved a 73.7% score in long-horizon software engineering, nearing Claude Opus 5’s 74.0%, while outperforming other tested models in financial agents, legal agents, and certain comprehensive reasoning tasks. The Cyber version can autonomously identify vulnerabilities and generate patches: in Google’s internal tests across 20 programming languages, its success rate exceeded 70%, and the Chrome team received 2.6 times more correct patches than those produced by large commercial models. The current API pricing is $0.75 per million input tokens and $3.75 per million output tokens; however, the model trades higher computational cost through additional reasoning steps and tool calls to enhance performance, with independent evaluations showing its per-task cost is approximately 40% higher than that of 3.7 Flash. The key takeaway from this release is that agent capabilities, previously confined to expensive flagship models, are now entering affordable, scalable “workhorse” models—but performance data remains primarily based on Google and partner tests, and the Cyber version is not available to general users.

Article author, source: DeepMind

Gemini 3.8 Flash was released just three weeks after Gemini 3.7 Flash and is Google’s third Flash version within six weeks.

In the past, "Flash" typically meant fast and low-cost, but with capabilities significantly weaker than flagship models. The positioning of Gemini 3.8 is changing: Google still calls it a "work model" suited for large-scale deployment, but now highlights advanced capabilities such as long-horizon programming, continuous tool usage, and professional analysis—rather than limiting it to lightweight tasks like chat and summarization.

The new model supports text, images, audio, video, and PDF inputs, with a 1 million token context window and a maximum output of 64,000 tokens. It can invoke search, functions, and computer operation tools. It is now officially available, not in test or preview mode, and can be accessed via the Gemini API, Google AI Studio, Gemini Enterprise, Google Antigravity, the Gemini app, Search AI Mode, and Google Sheets.

Google's hosted agent, Antigravity, has also adopted 3.8 Flash by default. This indicates that the model is truly designed for agents requiring repeated planning, tool invocation, result verification, and continuous refinement—not just single-turn Q&A.

In Google's DeepSWE v1.1 long-term software engineering benchmark, Gemini 3.8 Flash achieved a score of 73.7%, surpassing 3.7 Flash's 65.3% and GPT-5.6 Sol's 72.7%, and trailing Claude Opus 5 by just 0.3 percentage points.

This type of test requires the model to enter a real codebase, understand relationships between multiple files, locate issues, and make modifications that pass the tests. It more closely resembles the actual work of a programming agent than generating a standalone function.

In Vals Finance Agent v2, 3.8 Flash achieved 61.4%, while the tested 3.7 Flash, Claude Opus 5, and GPT-5.6 Sol scored 59.0%, 58.6%, and 53.8% respectively.

In the Harvey Legal Agent test, Flash 3.8 scored 10.0%, higher than Flash 3.7 at 8.8%, Claude Opus 5 at 6.7%, and GPT-5.6 Sol at 2.5%. However, all models achieved very low absolute scores in this evaluation; ranking first does not mean they can reliably handle complex legal tasks.

In the HLE-Verified multi-disciplinary reasoning test, 3.8 Flash achieved 54.9%, while GPT-5.6 Sol and Claude Opus 5 scored 54.5% and 54.4% respectively; the differences among the three are minimal and insufficient to demonstrate a stable, overall advantage for any single model.

These results were primarily published by Google according to its own configurations. Factors such as model architecture, agent framework, reasoning intensity, tool permissions, and test budget all influence the final scores; therefore, a more reasonable conclusion is that Flash-level models have approached the performance of some high-end flagship models, rather than Gemini decisively outperforming them across all tasks.

Google attributes the 3.8 Flash capability improvement to a straightforward design choice: making the model "try harder."

When faced with complex tasks, it employs more intermediate reasoning steps, repeatedly invokes tools, and actively checks已完成的工作. Developers can choose from three thought levels—low, medium, and high—with medium set as default; high level is suitable for difficult programming and multi-step reasoning, while low level is ideal for real-time chat, draft generation, and latency-sensitive tasks.

This reduces the likelihood of the agent stopping prematurely, missing steps, or completely deviating from its goal after a single tool call fails, but it will also consume more tokens.

In independent tests of Artificial Analysis, 3.8 Flash achieved an intelligence score of 59 at the highest thinking level, a 3-point increase over 3.7 Flash, and is comparable to GPT-5.6 Sol’s non-maximum reasoning configuration. Its average output speed is approximately 300 tokens per second, with an average completion time of 2.5 minutes per test item.

However, the average output tokens for 3.8 Flash increased by approximately 30%, and the number of execution rounds in Agent evaluations was also higher. Although the price per token did not rise, the average cost per task is about $0.58, roughly 40% higher than 3.7 Flash’s $0.40.

Therefore, "lower price per unit" does not mean "each task is necessarily cheaper." If the task itself does not require complex reasoning, running 3.8 Flash at high intensity may only increase time and cost. Google also recommends prioritizing efficient workflows by lowering the reasoning level or continuing to use the still-supported 3.7 Flash.

As of December 31, 2026, the promotional price for Gemini 3.8 Flash is $0.75 per million input tokens and $3.75 per million output tokens, the same as the current price for 3.7 Flash.

Starting January 1, 2027, the standard pricing will change to $1.50 per million input tokens and $7.50 per million output tokens, exactly doubling. Developers should not consider the launch-period pricing as permanent when evaluating long-term costs.

Even when calculated at future standard pricing, it remains lower than some flagship models; however, for agents requiring the generation of tens of thousands of tokens and dozens of tool calls, it’s more appropriate to compare the total task cost rather than just the per-million-token figures listed on the price table.

Google also released Gemini 3.8 Flash Cyber. It shares foundational capabilities with the standard version but has undergone more focused cybersecurity training and employs looser cybersecurity restrictions, enabling it to perform vulnerability analyses that the standard model might refuse.

In the CyberGym vulnerability discovery benchmark, Google stated it has reached state-of-the-art levels. Since CyberGym primarily focuses on C and C++ projects, the company developed an internal test suite covering 20 programming languages and incorporating complex, real-world codebases; Flash Cyber achieved a success rate of over 70% in vulnerability detection.

Discovering vulnerabilities is only the first step. Google emphasizes that Cyber's training focuses on fixing issues rather than generating exploit code.

In the CWE-Bench patch test run by Collinear, Flash Cyber's first submission achieved a pass rate of 47.2%, compared to 47.8% for the leading flagship model under comparison. The results are very close, but Google states that Flash Cyber has lower operational costs.

This represents a shift in the cybersecurity agent’s goal from “informing engineers that there may be an issue here” to understanding vulnerabilities, writing patches, running tests, and verifying fixes. If the results are sufficiently reliable, it could reduce the time it takes for the security team to move from discovering a vulnerability to deploying a fix.

Google has used Flash Cyber for internal code security work.

Chrome’s security team stated that the number of correct vulnerability patches it generates is 2.6 times higher than those produced by larger commercial models. Cybersecurity company Wiz reported that Flash Cyber’s vulnerability recall rate exceeds other state-of-the-art models by 7.5 to 9.7 percentage points in its penetration testing benchmarks, at a cost 2.3 to 5.2 times lower.

Google Cloud's vulnerability research team also stated that the model identified a critical foundational vulnerability in less than two hours, whereas similar research typically takes months.

These results have practical reference value but cannot be considered independent, reproducible public evaluations. Google has not disclosed the specifics of the critical vulnerability, the list of comparative models, or the full execution traces; the Chrome data comes from within Google, and Wiz is a Fairwind partner. Therefore, they demonstrate that models are already capable of participating in real-world security tasks, but they do not prove that they consistently achieve equivalent performance across all codebases and vulnerability types.

Flash Cyber is offered through Google’s newly established Fairwind Program, which currently includes over 650 participants, primarily government cybersecurity agencies, critical infrastructure operators, software maintainers, and large security companies.

Participating organizations must limit access to internal personnel and implement security measures such as multi-factor authentication. When integrated with the CodeMender Agent, the model can locate, verify, and remediate vulnerabilities within the organization’s own cloud environment.

Regular Google Cloud customers can still use CodeMender with the public version of Gemini, but they cannot directly access the full capabilities of Flash Cyber.

This “one foundational model, two permission levels” release approach closely mirrors Anthropic’s previous strategy of dividing Claude into Fable and Mythos: general programming and analytical capabilities are made available to the market, while cybersecurity functions more easily exploitable for attacks are restricted through identity verification, access controls, and continuous monitoring.

The standard 3.8 Flash includes security restrictions related to chemical, biological, radiological, nuclear, and cyber threats; the Cyber version, which requires fewer cybersecurity restrictions to fulfill professional defense tasks, is only available to vetted organizations.

Google's model card states that, compared to 3.7 Flash, 3.8 Flash does not introduce any substantial new capabilities in high-risk areas targeted by the company's advanced security framework, and therefore does not trigger a higher risk rating. It also shows improved protection in Gray Swan indirect prompt injection tests.

However, the model card also discloses that 3.8 Flash performed 5.4 percentage points worse than 3.7 Flash in automated multilingual security testing. After manual review, Google stated that most of the differences were false positives or minor issues, but the full sample set and language-by-language data were not released.

The model still exhibits common issues of large language models, including hallucinations, occasional slow responses or timeouts, and excessive token usage to enhance performance. The knowledge cutoff date is listed as March 2026, but Google notes that reliable knowledge in some areas may still only be updated to around January 2025; verification via an internet connection is still required for the most current information.

What’s most noteworthy about Gemini 3.8 Flash isn’t that it leads by a fraction of a point on any single benchmark, but that Google is embedding long-horizon programming, professional analysis, and autonomous tool invocation capabilities into the Flash price tier.

The flagship model is suited for high-value, low-frequency, complex tasks; however, the agents actually deployed in enterprise customer service, programming pipelines, financial analysis, and security scanning may handle millions of requests per day. In these scenarios, speed, cost per unit, and task success rate often matter more than ranking first.

3.8 Flash also reveals a contradiction in this approach: the more adept a model is at repeated checking and sustained work, the more likely it is to generate additional tokens, increasing the per-task cost for “cheap models.” Therefore, the core of future Agent competition will not only be who answers best, but also who can best determine when to continue thinking, when to invoke tools, and when to stop.