Google Launches Three Gemini Models: How Flash Redefines Agent Cost Evaluation

Model role split

On 21 July 2026 Google released three Gemini 3.x models that serve distinct functions. Gemini 3.6 Flash is the work‑horse model for complex planning, multimodal analysis and code‑intensive tasks. Gemini 3.5 Flash‑Lite is a high‑throughput execution model for search, translation, data extraction and other low‑latency sub‑tasks. Gemini 3.5 Flash Cyber is a security‑focused sub‑model accessed only through the CodeMender pilot for government agencies and trusted partners. All three reached GA status on the Gemini API on the same day, but Flash Cyber is not publicly available.

Core improvement: reducing invalid agent loops

Official Gemini benchmarks show that 3.6 Flash improves scores on several agent‑oriented evaluations:

DeepSWE v1.1: 37 % → 49 %

MLE‑Bench: 49.7 % → 63.9 %

OSWorld‑Verified: 78.4 % → 83.0 %

GDPval‑AA v2: 1349 → 1421

In addition, 3.6 Flash reduces output token volume. On the Artificial Analysis Intelligence Index the average output drops by 17 % (≈28 k → 23 k tokens). On DeepSWE the reduction is about 65 % (≈276 k → 97 k tokens). Google attributes these gains to fewer unnecessary reasoning steps, tool calls and code‑modification loops.

Task‑level cost analysis

Standard Gemini API pricing (per million tokens) is:

3.6 Flash: input $1.50, output $7.50 (output price includes thinking tokens)

3.5 Flash: input $1.50, output $9.00

3.5 Flash‑Lite: input $0.30, output $2.50

Assuming a task that previously emitted 1 M output tokens at $9.00, a 17 % token reduction combined with the lower output price yields an estimated output cost of $6.23 (≈30.8 % saving). Real‑world costs also depend on input context size, caching, grounding, tool fees, call frequency, failure rates and retries.

Flash‑Lite as high‑throughput execution layer

Google reports that Flash‑Lite can generate 350 output tokens per second, the fastest speed among the 3.5 series. In the same official benchmarks the model improves:

Terminal‑Bench 2.1: 31 % → 54 %

GDPval‑AA v2: 642 → 1140

GDM‑MRCR v2: 60.1 % → 72.2 %

These gains indicate that tasks previously requiring higher‑tier models can now be handled by the Lite tier, offering a lower‑cost execution path.

Flash Cyber as a specialized security sub‑model

Flash Cyber is only reachable via CodeMender, which may invoke the model up to five times per report. In the CyberGym benchmark the combined system achieves an 83.2 % score. In a V8 JavaScript engine bug‑finding test Flash Cyber discovers 55 unique confirmed bugs, compared with 47 for 3.5 Flash and 36 for Claude Opus 4.6. The evaluation notes that competitor scores in CyberGym are self‑reported and that some security‑oriented models (e.g., Claude Opus 4.6) restrict execution of certain tasks, affecting comparability.

Emergence of “agent economics”

The three‑model lineup illustrates a shift from evaluating models solely by per‑token price or single‑turn quality to a broader “agent economics” view. Cost now incorporates output token count, context growth, reasoning rounds, tool calls, latency, failure‑rate and routing. Reducing redundant output and execution loops improves scalability and makes agent‑driven workflows more production‑ready.

References

Google Blog: Introducing Gemini 3.6 Flash, 3.5 Flash‑Lite, and 3.5 Flash Cyber (2026‑07‑21)

Gemini API Release Notes: GA announcement for 3.6 Flash and 3.5 Flash‑Lite (2026‑07‑21)

Gemini Developer API Pricing: standard‑layer prices for 3.6 Flash, 3.5 Flash, 3.5 Flash‑Lite

Google DeepMind: Introducing Gemini 3.5 Flash Cyber (2026‑07‑21)

Google Blog: Gemini 3.1 Flash‑Lite – Built for intelligence at scale (2026‑03‑03)

Code example

[1] Google Blog：Introducing Gemini 3.6 Flash, 3.5 Flash-Lite, and 3.5 Flash Cyber（2026-07-21）
[2] Gemini API Release Notes：Gemini 3.6 Flash与3.5 Flash-Lite GA公告（2026-07-21）
[3] Gemini Developer API Pricing：3.6 Flash、3.5 Flash、3.5 Flash-Lite标准层价格
[4] Google DeepMind：Introducing Gemini 3.5 Flash Cyber（2026-07-21）
[5] Google Blog：Gemini 3.1 Flash-Lite: Built for intelligence at scale（2026-03-03）