By Nine days after Google I/O, most developers are still running Gemini 3.1 Pro. That’s a mistake. Gemini 3.5 Flash — released May 19 — outperforms 3.1 Pro on every benchmark that matters for agentic workloads, costs 40% less, and runs four times faster. The model that was supposed to be the budget option has quietly become the better option.
The Numbers That Settle the Argument
Google shipped 3.5 Flash with a claim that sounded like marketing: “strongest agentic and coding model yet.” The benchmark data backs it up.
| Benchmark | Gemini 3.5 Flash | Gemini 3.1 Pro |
|---|---|---|
| Terminal-Bench 2.1 | 76.2% | 70.3% |
| MCP Atlas (tool use) | 83.6% | 78.2% |
| GDPval-AA (Elo) | 1,656 | 1,314 |
| CharXiv Reasoning | 84.2% | — |
The GDPval-AA gap deserves attention: 342 Elo points is not a marginal improvement — that’s a different tier of agentic capability. The MCP Atlas score of 83.6% — measuring multi-step tool-chain reliability — beats not only Gemini 3.1 Pro (78.2%) but also Claude Opus 4.7 (79.1%). A Flash model outperforming current-generation Opus on tool use.
Speed compounds the case. At 280+ output tokens per second, Gemini 3.5 Flash runs roughly 70% faster than Gemini 3 Flash. For streaming-heavy applications and agent loops that need low latency, this is not a minor footnote.
Pricing: The Budget Math Flips
Better performance at lower cost is the developer dream. Here’s what Gemini 3.5 Flash actually delivers on Google’s official pricing:
| Model | Input ($/M tokens) | Output ($/M tokens) |
|---|---|---|
| Gemini 3.5 Flash | $1.50 | $9.00 |
| Gemini 3.1 Pro | $2.00 | $12.00 |
| Claude Sonnet 4.6 | $3.00 | $15.00 |
Cached input tokens drop to $0.15 per million — a 90% discount for repeated context. For pipelines that reuse system prompts or large documents across requests, that compounds fast. A pipeline processing 100 million tokens per day pays $150 with Flash versus $200 with 3.1 Pro. Over a year, that’s roughly $18,000 saved before factoring in output costs.
Budget arguments for staying on Gemini 3.1 Pro are gone. Flash is the financially rational default for most production workloads.
Thinking Levels: The New Dial
The old thinking_budget integer is out. Gemini 3.5 Flash introduces thinking_level, a string enum with four settings that control the quality-latency-cost tradeoff. Per the official API documentation:
- MINIMAL — Lowest latency, appropriate for simple completions
- LOW — Google’s recommended default for agentic coding and tool calling; generates 45% fewer tokens than MEDIUM with comparable quality on code tasks
- MEDIUM — The actual default; strong across most task types
- HIGH — Complex reasoning; time-to-first-token reaches ~17.75 seconds
For agent developers, start with LOW. Google specifically retuned it for code generation and tool-calling workflows — the two dominant patterns in production agent systems. Only escalate to MEDIUM or HIGH if you’re seeing quality problems with LOW.
Migrating from Gemini 3.1 Pro: Four Changes
The migration is straightforward. Four changes cover most codebases:
- Update the model ID:
gemini-3.1-pro→gemini-3.5-flash - Replace
thinking_budgetwiththinking_level(string enum) - Remove
temperature,top_p,top_kfrom your config (no longer recommended) - Add
idand matchingnameto all FunctionResponse parts
You’ll also need google-genai SDK v2.0.0 or later. Here’s what the updated call looks like:
from google import genai
from google.genai import types
client = genai.Client()
response = client.models.generate_content(
model="gemini-3.5-flash",
contents="Your prompt here",
config=types.GenerateContentConfig(
thinking_config=types.ThinkingConfig(thinking_level="low")
),
)
print(response.text)For Managed Agents via the Interactions API, the same model underpins Antigravity. One API call provisions an isolated Linux sandbox with reasoning, tool use, and code execution. The Managed Agents quickstart walks through the full setup.
When Flash Is the Wrong Choice
Not every workload fits.
If your primary use is code review that must be correct — SWE-Bench style tasks where wrong is worse than slow — Claude Sonnet 4.6 still leads. Flash’s Terminal-Bench advantage is in agentic execution, not surgical code correctness on hard problems.
For genuinely long-horizon, multi-step agent tasks, it may be worth waiting for Gemini 3.5 Pro, which Google confirmed is coming in June 2026 and is currently in testing. Per the DeepMind model card, Flash is positioned as the volume workhorse; Pro will be the ceiling model.
And if your infrastructure is already deep in the Claude Code ecosystem — hooks, permissions, memory configurations — the migration cost is real. Flash’s agentic advantages need to clearly outweigh switching friction before committing.
The Bottom Line
Gemini 3.5 Flash is the default choice for most production agentic workloads as of May 19, 2026. Not “worth evaluating” — default. It outperforms Gemini 3.1 Pro on the benchmarks that drive real decisions, costs 40% less, and runs 4x faster. The Flash label has stopped meaning what it used to mean.
