Google shipped three Flash models in roughly six weeks. Gemini 3.8 Flash, released September 2, is the one worth updating your code for. It scores 90.8% on Terminal-Bench 2.1 — up from 81.6% on its immediate predecessor — nearly doubled its Terminal-Bench 4.0 score, and is now the default model powering Google’s own Antigravity coding environment. The upgrade also kills temperature parameters, introduces a thinking-level API, and ships with a locked-down security twin that found a 13-year-old Chrome bug in under two hours.
What Actually Improved #
The benchmark gains are real, and they’re targeted squarely at agentic workloads. Terminal-Bench 2.1 jumped from 81.6% to 90.8%. Terminal-Bench 4.0 — which tests whether a model can operate a terminal, run commands, and complete technical setup tasks autonomously — nearly doubled, from 11.2% to 19.1%. OSWorld-2.0, measuring computer-use capability, went from 50.6% to 59.0%.
Google’s stated design goal: 3.8 Flash “works harder.” It runs more reasoning steps and calls tools more aggressively on complex tasks. That’s exactly what you want in an agent backbone — and exactly why general exam-style reasoning didn’t move much. That wasn’t the target.
For context: Claude Opus 5 scores 51.8% on Terminal-Bench 4.0 and 75.4% on OSWorld-2.0. Gemini 3.8 Flash isn’t dethroning frontier models. It’s closing the gap at Flash-tier pricing, which is the relevant comparison for most production workloads.
The API Change You Cannot Ignore #
Gemini 3.8 Flash deprecates temperature, top_p, and top_k. On 3.8 Flash, those parameters are silently ignored. Set frequency_penalty, presence_penalty, or candidate_count and you’ll get an active API error. The old integer-based thinking_budget is also gone.
The replacement is a three-level enum:
- low — minimal reasoning, fastest and cheapest; use for extraction, classification, tasks you verify downstream
- medium — the default; Google recommends this for code and agent work
- high — maximum reasoning budget; for complex multi-step logic and tool-heavy pipelines
One migration trap worth flagging: minimal thinking level is not supported on 3.8 Flash and throws an error. If your existing code uses it, strip it before flipping the model string. Otherwise the move from 3.7 Flash to 3.8 Flash is mostly a find-and-replace on the model name — the Gemini 3 family shares API conventions.
from google import genai
client = genai.Client()
interaction = client.interactions.create(
model="gemini-3.8-flash",
input="Analyze this payment pipeline for race conditions and rewrite the transaction locks safely.",
generation_config={
"thinking_level": "medium"
}
)
print(interaction.output_text)
The Cost Reality Check #
The pricing looks identical to 3.7 Flash: $0.75 per million input tokens, $3.75 per million output tokens through December 31, 2026 — after which rates double to $1.50 and $7.50. But “same rate” doesn’t mean “same cost.” Because 3.8 Flash generates roughly 30% more output tokens per task and takes more agentic turns, the same workflow costs approximately 40% more in practice.
The upgrade makes sense for workloads that benefit from better agent reliability — fewer retries, more accurate tool calls, cleaner code output. For high-volume, low-complexity tasks where 3.7 Flash already delivers, the extra token burn isn’t worth it. Run both models on a representative sample of your production traffic before committing.
Flash Cyber: What Google Is Keeping for Itself (For Now) #
Alongside 3.8 Flash, Google announced Gemini 3.8 Flash Cyber — a security-focused variant not available through the standard API. Access is gated behind the Fairwind Program, which restricts it to verified SOC teams, critical infrastructure operators, and accredited security researchers.
The results so far are hard to dismiss. Google’s Chrome Security team used Flash Cyber to generate correct vulnerability patches at 2.6 times the rate of the best comparable commercial models. The Cloud Vulnerability Research team found a critical vulnerability in under two hours — work that normally takes months. Most notably, Flash Cyber surfaced a bug that had been sitting in the Chromium codebase for 13 years, undetected by the hundreds of engineers who reviewed that code.
For most developers, Flash Cyber is aspirational today. But it signals where AI-assisted security tooling is heading: models fine-tuned for vulnerability discovery, with access gradually opening as Google stress-tests the safety model.
Integration and Availability #
Gemini 3.8 Flash is available now through Google AI Studio and the Gemini API. It’s the default model in the Antigravity SDK and the Antigravity Agent. Built-in capabilities include function calling, search as a tool, computer use, and code execution across a 1M-token context window with 64K output.
If you’re building agents and still on 3.7 Flash, the upgrade path is short. Strip the old sampling parameters, add thinking_level, swap the model string, and benchmark against your actual workloads. The benchmark improvements are real — just go in with clear eyes on the token cost tradeoff. See Google’s full announcement for the complete changelog.