Google has released Gemini 3.8 Flash, a new model designed to handle long software-engineering tasks, autonomous agents and complex professional workflows.
Arriving only three weeks after Gemini 3.7 Flash, it is Google’s third Flash release in six weeks. That unusually rapid schedule shows how quickly the competition between AI companies is shifting away from occasional flagship launches and toward continuously updated models built for everyday production workloads.
Google presents Gemini 3.8 Flash as its most capable Flash model yet. It performs more reasoning steps, uses tools more persistently and verifies its work during difficult tasks. Early independent evaluations support the idea that it is a meaningful improvement, particularly for coding and agentic workflows.
There is, however, an important trade-off. Gemini 3.8 Flash can generate more tokens and spend longer completing a task. Its price per million tokens may initially look unchanged, but its real cost per completed job can still rise.
That distinction may matter more than the headline benchmark scores.
What Google has actually released
The Gemini 3.8 launch includes two models built on the same underlying intelligence:
- Gemini 3.8 Flash, the generally available model for developers, businesses and paying Gemini users;
- Gemini 3.8 Flash Cyber, a specialized version for vulnerability discovery and automated software patching, available only to approved organizations.
Gemini 3.8 Flash is not a new Gemini Pro flagship. It is the latest version of Google’s faster and more economical Flash line, which the company increasingly positions as a practical workhorse for applications operating at scale.
The model accepts text, images, audio and video as inputs and produces text as output. It retains a context window of one million tokens and supports responses of up to 64,000 tokens.
Developers can select three reasoning levels: low, medium or high. Medium is the default. A lower setting favors speed and reduced token consumption, while the highest setting allows the model to spend more time reasoning and coordinating tools.
Gemini 3.8 Flash is generally available through the Gemini API and Google AI Studio. It also powers Google’s Antigravity agent by default and is offered through Gemini Enterprise. Consumers can access it through the Gemini app, AI Mode in Google Search and Google Sheets with a Google AI Pro or Ultra subscription.
This wider distribution complements Google’s effort to turn its assistant into something capable of performing actions, a transition explored in our article on how Gemini Live is becoming an AI agent.
Gemini 3.8 Flash “works harder”
The central change is not a larger context window or a new type of input. It is the way the model approaches complex work.
Google says Gemini 3.8 Flash “works harder.” On difficult requests, the model can break the problem into smaller reasoning steps, call tools repeatedly and verify its progress before producing a final result.
For a coding agent, this could mean inspecting several files, identifying dependencies, editing the code, running tests, reviewing the errors and making further corrections. A simpler chatbot might produce an answer after one pass, while an agent may loop through the problem several times.
This approach is particularly relevant for long-horizon tasks, where success depends on maintaining an objective across many operations rather than producing one convincing response.
Google reports that Gemini 3.8 Flash performs strongly on DeepSWE v1.1, a benchmark focused on solving complex software-engineering problems from beginning to end. The company also reports improvements on professional agent benchmarks for finance and legal work, along with a score of 54.9% on HLE-Verified, which evaluates advanced reasoning across multiple fields.
These figures are promising, but several of the comparisons come from Google’s own evaluations. They should therefore be treated as evidence of progress rather than proof that Gemini 3.8 Flash is universally better than every competing model.
Independent testing confirms progress—with qualifications
Independent measurements from Artificial Analysis provide a more nuanced picture.
At the highest reasoning level, Gemini 3.8 Flash scored 59 on the Artificial Analysis Intelligence Index, three points above Gemini 3.7 Flash. Within this particular evaluation, that placed it alongside GPT-5.6 Sol at a sub-maximum reasoning setting and Grok 4.6 at medium effort.
The largest improvements appeared in agentic evaluations involving tool use, terminal-based coding and realistic multi-step work. This supports Google’s claim that the update is primarily aimed at software engineering and autonomous workflows rather than ordinary conversational use.
| Artificial Analysis measurement | Gemini 3.8 Flash result |
|---|---|
| Intelligence Index, high reasoning | 59 |
| Gemini 3.7 Flash score | 56 |
| Estimated cost per task, high reasoning | $0.58 |
| Estimated cost per task, medium reasoning | $0.41 |
| Estimated cost per task, low reasoning | $0.24 |
| Output speed, high reasoning | About 300 tokens/second |
| Average task time, high reasoning | About 2.5 minutes |
The model is therefore fast and competitive for its price. It is not dominant in every category, however. Ars Technica notes that its improvement on OSWorld 2.0, which tests an agent’s ability to operate a computer, still leaves it well behind Claude Opus in that particular evaluation.
The appropriate conclusion is not that Gemini 3.8 Flash has defeated every larger model. It is that Google has moved a relatively fast and inexpensive model closer to frontier-level performance on several commercially important tasks.
The hidden cost behind the unchanged token price
Gemini 3.8 Flash initially uses the same promotional API rates as Gemini 3.7 Flash:
| API usage | Through December 31, 2026 | From January 1, 2027 |
|---|---|---|
| Input tokens | $0.75 per million | $1.50 per million |
| Output tokens | $3.75 per million | $7.50 per million |
The introductory prices are scheduled to double at the beginning of 2027. Even before that increase, the cost of running Gemini 3.8 Flash cannot be understood from the per-token rate alone.
Artificial Analysis found that the model’s average output rose to approximately 48,000 tokens across its test workload, around 30% more than Gemini 3.7 Flash. Additional turns during agentic evaluations also contributed to an estimated cost of $0.58 per task at high reasoning, compared with $0.40 for its predecessor—an increase of roughly 40% despite identical token prices.
This does not necessarily make Gemini 3.8 Flash poor value. A model that uses more computation but completes a difficult job successfully can be cheaper than a less capable model that fails, requires repeated prompts or needs human correction.
It does mean that businesses should compare models using broader measures:
- total cost per successful task;
- reliability across repeated runs;
- latency and time to completion;
- number of tool calls;
- human review requirements;
- and the cost of failed attempts.
The same issue affects competing systems. As explained in our analysis of OpenAI’s GPT-5.6 Sol API price reduction, agentic workloads may involve dozens of model calls, making token efficiency as important as the advertised API rate.
Google itself acknowledges this trade-off. Developers working on cost-sensitive or latency-critical applications can lower the reasoning level or continue using Gemini 3.7 Flash, which remains supported.
Gemini 3.8 Flash Cyber targets software vulnerabilities
The second model announced by Google is more specialized—and more restricted.
Gemini 3.8 Flash Cyber is designed to discover vulnerabilities and generate software patches. It uses the same foundational intelligence as the standard model, but operates with cybersecurity-specific training and a more permissive set of safeguards for defensive work.
Google reports a success rate above 70% on an internal vulnerability-discovery evaluation covering complex codebases in 20 programming languages. On the external CWE-Bench patching benchmark, the model achieved a pass@1 score of 47.2%, close to the 47.8% recorded by a leading frontier model but at a lower claimed operating cost.
The company also says its Chrome security team obtained 2.6 times more correct vulnerability patches than with the best larger commercial models it tested. Wiz reported higher recall at a lower cost on its own penetration-testing benchmark, while Google’s Cloud Vulnerability Research team used the model to identify a critical vulnerability in less than two hours.
These results are significant, but several come from internal tests conducted by Google or its partners. Independent real-world validation will be important before drawing broader conclusions.
Gemini 3.8 Flash Cyber is not available as a normal public model. Access is provided through Google’s Fairwind Program to selected government authorities, critical-infrastructure operators and software maintainers. This restriction reflects the dual-use nature of advanced cybersecurity models: the same capabilities that help defenders locate a flaw could also assist attackers.
What users and developers should expect
For most consumers, the difference may not be immediately visible in everyday questions. Gemini 3.8 Flash is primarily optimized for tasks where the model must continue working, use tools and recover from intermediate problems.
Developers and businesses are more likely to notice the change when building:
- coding assistants that work across complete repositories;
- agents that coordinate several software tools;
- financial, legal or operational analysis workflows;
- systems processing large collections of documents;
- automated quality-control and verification pipelines.
The adjustable reasoning level is therefore an important part of the product. Using maximum effort for every request would waste time and tokens. A practical application could route simple work to the low setting, retain medium as its standard mode and reserve high reasoning for the most difficult tasks.
Gemini 3.8 Flash also retains the familiar limitations of large language models. Google’s model card warns that it can still hallucinate and may occasionally suffer from slow responses or timeouts. Its knowledge cutoff is listed as March 2026 for some domains, while information in others may be limited to January 2025. Access to Search or other external tools remains important for current information.
Google is turning Flash into its main competitive weapon
The speed of this release cycle may be as important as the model itself.
By launching three Flash updates in six weeks, Google is treating its production model as a rapidly evolving service rather than a product generation that remains unchanged for many months. The strategy fits the needs of AI agents, which can consume large numbers of tokens and therefore make performance per dollar a decisive factor.
Gemini 3.8 Flash does not represent a clean breakthrough in every capability. Its specifications remain close to those of its predecessor, several benchmark gains are incremental, and stronger reasoning comes with higher token consumption.
But the model demonstrates something potentially more consequential: advanced coding and agentic performance is moving into faster, less expensive tiers. Capabilities that recently required a premium frontier model are becoming available in systems designed for high-volume use.
For users, the update should make complex Gemini workflows more reliable. For developers, the decision is less straightforward. Gemini 3.8 Flash may deliver better results, but the optimal reasoning level—and the true cost of achieving those results—will need to be measured application by application.
Google has made Flash smarter. It has not made the economics simple.