Gemini 3.8 Flash: Specs, Benchmarks, and Agentic Upgrades
Google DeepMind releases Gemini 3.8 Flash and Flash Cyber, introducing recursive agentic training, 54.9% on HLE-Verified, and $0.75/$3.75 pricing.
What was announced
On September 2, 2026, Google DeepMind announced the release of Gemini 3.8 Flash alongside a specialized security counterpart, Gemini 3.8 Flash Cyber. Detailed in an official announcement by Google DeepMind co-authored by Senior Director of Product Management Tulsee Doshi and Gemini Security Lead Raluca Ada Popa, this release marks DeepMind’s third iteration of its Flash tier within a six-week window. The launch follows Gemini 3.7 Flash—which arrived on August 13, 2026—and the preceding Gemini 3.6 Flash release.
According to DeepMind, Gemini 3.8 Flash is positioned as their most intelligent workhorse model to date, targeting complex multi-step reasoning, autonomous agentic operations, and long-horizon software engineering. DeepMind confirmed that Gemini 3.8 Flash maintains the exact introductory pricing tier introduced with Gemini 3.7 Flash:
- Input Tokens: $0.75 per million tokens
- Output Tokens: $3.75 per million tokens
The release introduces two distinct variants built upon a shared architectural core:
- Gemini 3.8 Flash: A general-purpose reasoning and coding engine built for production software engineering and enterprise workflows requiring tool invocation and iterative problem solving.
- Gemini 3.8 Flash Cyber: A domain-tuned cybersecurity variant designed for automated vulnerability detection and patch generation. Access to the Cyber model is restricted to vetted defensive teams through DeepMind’s newly established Fairwind Program.
DeepMind claims that both models share the same foundational intelligence, which was refined using long-running recursive agentic loops. The team also noted that training the core foundation in cybersecurity contributed directly to the general model's reasoning and code-generation gains.
Alongside these architectural improvements, DeepMind highlighted the role of inference-time computation. Gemini 3.8 Flash introduces adjustable "effort levels." The model is engineered to exhibit greater diligence on complex prompts by issuing iterative tool calls and executing extended internal reasoning steps, which can consume more output tokens at higher effort settings. For token-sensitive deployments, developers are directed to configure lower effort levels or remain on Gemini 3.7 Flash, which DeepMind confirmed will remain fully supported for efficiency-first workloads.
Benchmark Evaluation: Coding, Reasoning, and Domain Agents
DeepMind’s announcement framed Gemini 3.8 Flash as a cost-effective alternative to larger frontier systems, asserting that the model approaches or exceeds the performance of higher-cost models across software engineering and professional reasoning benchmarks.
+------------------------------------+-------------------------+-------------------------+
| Benchmark | Gemini 3.7 Flash | Gemini 3.8 Flash |
+------------------------------------+-------------------------+-------------------------+
| DeepSWE v1.1 (Long-Horizon SWE) | 65.3% | "Outperforms most |
| | | larger frontier models" |
| HLE-Verified | Not disclosed | 54.9% |
| Vals Finance Agent V2 | Baseline | Outperforms 3.7 Flash |
| Harvey's Legal Agent Benchmark | Baseline | Outperforms 3.7 Flash |
| Input Pricing (per 1M tokens) | $0.75 | $0.75 |
| Output Pricing (per 1M tokens) | $3.75 | $3.75 |
+------------------------------------+-------------------------+-------------------------+
Software Engineering and Agentic Coding
On software engineering evaluations, DeepMind highlighted performance on DeepSWE v1.1 (Long-Horizon Software Engineering). While Gemini 3.7 Flash previously registered a 65.3% score on this benchmark (up from 49.0% on Gemini 3.6 Flash), DeepMind did not publish a specific percentage for Gemini 3.8 Flash. Instead, the announcement stated that 3.8 Flash outperforms most larger, higher-cost frontier models in autonomously resolving end-to-end engineering issues at a fraction of their operating cost.
DeepMind also showcased the model's iterative code-generation loop in a production setting, noting that Gemini 3.8 Flash generated a playable game from a simple text prompt using a looping instruction inside Google Antigravity. This mirrors prior internal demonstrations, such as Gemini 3.7 Flash orchestrating sub-agents alongside Gemini Omni or generating dynamic assets with Nano Banana.
Multi-Step Professional Reasoning
For broader reasoning evaluations, DeepMind reported that Gemini 3.8 Flash scored 54.9% on HLE-Verified. The benchmark measures multi-step reasoning capabilities spanning STEM disciplines, the humanities, and quantitative professional fields.
In specialized domain benchmarks, DeepMind stated that 3.8 Flash surpassed Gemini 3.7 Flash and competing frontier baselines on:
- Vals Finance Agent V2: Evaluating advanced financial modeling, quantitative reporting, and corporate data analysis.
- Harvey's Legal Agent Benchmark: Measuring legal analysis, contract comprehension, and statutory synthesis.
As teams track wider industry shifts toward autonomous agent architectures—such as those analyzed in our coverage of the GPT-6 Astra Architecture, Benchmarks, and Rollout Analysis—Gemini 3.8 Flash illustrates how providers are attempting to bring frontier-grade agentic benchmark scores down to mid-tier API pricing.
Background: Recursive Agentic Loops and Test-Time Compute
To understand the improvements claimed for Gemini 3.8 Flash, it is helpful to examine the engineering mechanics behind DeepMind's stated methodology: recursive agentic training loops and test-time compute allocation.
Pre-training & Domain Corpus
(Cybersecurity, Code, STEM)
│
▼
┌──────────────────────────────┐
│ Base Transformer Weights │
└──────────────┬───────────────┘
│
▼
┌────────────────────────────────────────────────┐
│ Recursive Agentic Reinforcement Loop │
│ │
│ ┌──────────────┐ ┌──────────────┐ │
│ │ Generate │ ───────> │ Execute Tool │ │
│ │ Hypothesis │ │ / Run Code │ │
│ └──────────────┘ └──────┬───────┘ │
│ ▲ │ │
│ │ ▼ │
│ ┌──────┴───────┐ ┌──────────────┐ │
│ │ Refine Model │ <─────── │ Verify State │ │
│ │ Parameters │ │ & Environment│ │
│ └──────────────┘ └──────────────┘ │
└───────────────────────┬────────────────────────┘
│
▼
┌──────────────────────────────┐
│ Gemini 3.8 Shared Core │
└──────────────┬───────────────┘
│
┌────────────────┴────────────────┐
▼ ▼
┌─────────────────────┐ ┌─────────────────────┐
│ Gemini 3.8 Flash │ │ Gemini 3.8 Flash │
│ (General Agentic) │ │ Cyber (Defenders) │
└─────────────────────┘ └─────────────────────┘
Recursive Evaluation and Refinement
DeepMind noted that Gemini 3.8 Flash was "accelerated by long-running agentic loops designed to recursively evaluate and refine the underlying models." In contemporary AI engineering, recursive agentic loops move beyond traditional static Reinforcement Learning from Human Feedback (RLHF).
In general multi-agent architectures, an evaluation harness executes the model within an isolated runtime environment (such as a containerized sandbox or a virtual terminal). The model generates hypotheses, executes code, inspects execution errors, and iterates on its output. When applied during post-training, the trajectory of self-correction—evaluating why an execution failed and revising the approach—is converted into synthetic fine-tuning datasets or used as an environment reward signal. By recursively using agentic execution traces to curate higher-fidelity training data, models learn to self-correct during inference before returning a final output.
Underlying inference efficiency remains critical for these iterative multi-step pipelines. Fast token generation and optimized attention mechanisms—such as the memory bandwidth improvements covered in How FlashAttention-3 Works: Hopper GPU Optimizations—are foundational to making high-frequency tool-calling loops economically viable in production.
Test-Time Compute and Effort Levels
DeepMind explicitly stated that 3.8 Flash "works harder" by executing extra reasoning steps and issuing iterative tool calls, noting that the model may emit more tokens to maximize benchmark accuracy.
In machine learning theory, this behavior uses test-time search (often referred to as test-time compute scaling). Rather than attempting to produce the final token sequence in a single greedy forward pass, the model allocates compute dynamically during inference:
- Drafting and Scratchpads: The system generates internal reasoning paths or sub-queries.
- Tool Verification: The model queries external tools, linters, compilers, or database environments.
- Iterative Revision: If intermediate verification checks fail, the model rewrites its planning logic before returning a finalized payload to the client.
Because this search process consumes additional output tokens, API costs increase linearly with reasoning length. By introducing configurable effort levels, DeepMind allows systems architects to balance latency, token spend, and reasoning depth depending on whether a task is an open-ended debugging assignment or a low-latency routing query.
Gemini 3.8 Flash Cyber and Defensive Systems
Alongside the general model, DeepMind introduced Gemini 3.8 Flash Cyber, a dedicated variant focused on defensive cybersecurity operations. DeepMind reported that rigorous training in cybersecurity directly drove the code comprehension and reasoning gains across the shared foundational model.
Gemini 3.8 Flash Cyber Architecture
┌─────────────────────────────────────────────────────────────┐
│ Vetted Defense Pipeline │
└──────────────────────────────┬──────────────────────────────┘
│
┌────────────────────┴────────────────────┐
▼ ▼
┌───────────────────────────┐ ┌───────────────────────────┐
│ Static/Dynamic Analysis │ │ Automated Patching │
│ │ │ │
│ • Control-flow analysis │ │ • Synthesizes repairs │
│ • Memory corruption paths │ │ • Validates regression │
│ • Logic vulnerability ID │ │ • Verifies mitigations │
└───────────────────────────┘ └───────────────────────────┘
│ │
└────────────────────┬────────────────────┘
│
▼
┌─────────────────────────────────────────────────────────────┐
│ Fairwind Program Distribution Model │
└─────────────────────────────────────────────────────────────┘
Frontier-Level Detection and Patching
Gemini 3.8 Flash Cyber is designed to assist software security engineers in two primary phases of the defensive lifecycle:
- Vulnerability Detection: Parsing large, complex codebases to detect subtle edge-case bugs, memory-safety failures, logic flaws, and unsanitized attack surfaces before adversaries exploit them.
- Automated Patching: Formulating, compiling, and testing targeted remediations that patch the vulnerability without introducing regressions to adjacent systems.
Auditing modern software targets often requires tracing complex exploitation strategies—such as those explored in How ARM64 Pointer Authentication Bypasses Actually Work—where subtle architectural interactions cause system compromise. By applying specialized training on attack graphs, binary structures, and vulnerability repositories, DeepMind claims 3.8 Flash Cyber achieves frontier-level performance across defensive workflows.
The Fairwind Program
To mitigate dual-use risks, DeepMind announced that Gemini 3.8 Flash Cyber will not be generally available on standard public API tiers. Instead, access is gated through the newly established Fairwind Program.
Under the Fairwind Program, Gemini 3.8 Flash Cyber is restricted to verified security practitioners and organizations acting as defensive teams. DeepMind framed this gated deployment as a safety mechanism designed to ensure that automated vulnerability discovery and exploit synthesis tools remain confined to defensive operations and infrastructure protection.
What we do not know yet
While DeepMind’s blog post provides an overview of performance benchmarks and deployment targets, several key technical specifications and architectural parameters were omitted from the announcement.
- Context Window Length: DeepMind did not publish context length specifications for Gemini 3.8 Flash or 3.8 Flash Cyber. It is currently unconfirmed whether the model maintains the 1-million or 2-million token windows associated with earlier Gemini Flash generations.
- Model Architecture and Parameter Count: The announcement does not disclose parameter metrics. It is unknown whether 3.8 Flash is a dense transformer or uses a Mixture-of-Experts (MoE) routing topology, nor has DeepMind provided information regarding total parameters or active parameters per forward pass.
- Numerical DeepSWE v1.1 Score: While DeepMind stated that 3.8 Flash outperforms most larger frontier models on DeepSWE v1.1, the post does not provide an exact numeric score for 3.8 Flash. For context, Gemini 3.7 Flash scored 65.3% on this benchmark.
- Granular Domain Benchmark Metrics: DeepMind highlighted that Gemini 3.8 Flash outperformed Gemini 3.7 Flash on Vals Finance Agent V2 and Harvey's Legal Agent Benchmark, but specific percentage increases, task breakdowns, and baseline numbers were not published.
- Fairwind Program Access Criteria: DeepMind did not publish the vetting criteria, geographic constraints, institutional requirements, or cost structure for security teams seeking access to Gemini 3.8 Flash Cyber via the Fairwind Program.
- Training Compute and Hardware: DeepMind shared no details regarding the training infrastructure, compute cluster topologies (such as TPU generations), FLOP budgets, or dataset compositions used during pre-training and recursive reinforcement loops.
- Inference Latency Profiles: The announcement noted that 3.8 Flash operates at the "same speed and low cost of 3.7," but raw latency metrics—such as Time to First Token (TTFT) and median tokens-per-second generation rates across differing effort levels—were not included.
Conclusion
Gemini 3.8 Flash represents an aggressive iteration cycle from Google DeepMind, marking three Flash-tier releases across six weeks. Rather than competing purely on raw parameter scale, DeepMind is focusing its Flash series on compute-efficient agentic problem solving, using test-time reasoning and recursive loop training to tackle complex coding and multi-step tasks.
By matching the existing $0.75 / $3.75 per million token pricing introduced by Gemini 3.7 Flash, DeepMind has delivered a 54.9% score on HLE-Verified and claimed improvements across DeepSWE v1.1 without raising the base API costs for developers. Additionally, the spin-off of Gemini 3.8 Flash Cyber highlights DeepMind's dual strategy: driving core reasoning improvements through demanding security tasks, while using gated distribution channels like the Fairwind Program to manage dual-use risks. As enterprise deployments scale up autonomous workflows, production performance will depend on how effectively developers tune effort levels to balance token expenditure against task accuracy.