AI Models
Gemini 3.8 Flash Arrives With More Reasoning and More Benchmark Caveats
Google launched Gemini 3.8 Flash and 3.8 Flash Cyber on September 2. General Flash keeps the introductory 3.7 price and adds effort controls across consumer, developer, and enterprise surfaces; Cyber is limited to trusted defenders. The release's benchmark tables and partner results need exact settings and independent reproduction before they support broad superiority claims.
Citation-ready: Google launched Gemini 3.8 Flash on September 2, 2026, at $0.75 per million input tokens and $3.75 per million output tokens, while limiting the Flash Cyber variant to trusted defenders.
What happened and why it matters
No. They justify testing the exact model and effort setting, while real selection still depends on task completion, tool use, latency, token use, failures, safeguards, and accepted-output cost on the buyer's workload.
Official Google model launch and developer documentation
Primary reference: Google: Introducing Gemini 3.8 Flash and 3.8 Flash Cyber. Kaleido Field checked the event date and the article's attributed facts against this source.
| Source date | September 2, 2026 |
|---|---|
| Checked by Kaleido Field | September 3, 2026, 09:20 CST |
| Source function | current model-release analysis separating model variants, access, input modes, effort controls, price, first-party and external benchmark administration, partner tests, safeguard policy, and workload reproduction |
Effort is part of the model configuration
Google says 3.8 Flash may use more reasoning steps and tool calls on complex work, especially at higher effort. A faster model name therefore does not guarantee the shortest run or lowest accepted-task cost.
A comparison should freeze the model version, effort, tools, prompt, context, retries, stop conditions, cache, region, price date, and pass rule, then retain traces for failures as well as successes.
Cyber access changes the operating boundary
General Flash is available across the Gemini API, Gemini Enterprise, and specified consumer surfaces. Flash Cyber uses more permissive defensive-security safeguards and is routed through the Fairwind trusted-access program.
Security comparisons need the target authorization, sandbox, vulnerability class, reproduction, patch test, disclosure path, false positives, unsafe outputs, and reviewer decision in addition to a benchmark score.
Evidence boundary
Official product facts: model IDs and variants, launch date, input and output price, supported inputs, token limits, capabilities, effort settings, consumer and enterprise availability, and limited Cyber access. Google-reported or launch-curated evidence: DeepSWE, HLE-Verified, CyberGym, internal security tests, and named partner results. Externally administered but launch-reported evidence includes Vals and CWE-Bench results. Not established: independent reproduction across every table, universal price-performance lead, stable production latency, failure rate, or superiority on a specific customer workload.
FAQ
What is the general Flash price?
$0.75 per million input tokens and $3.75 per million output tokens at launch.
Who can access Flash Cyber?
Google says trusted defenders can receive prioritized access through the Fairwind Program.
Are all launch results independent?
No. The release combines Google-run, partner-reported, and externally administered benchmark evidence that must be read separately.