Gemini 3.5 Flash: Benchmarks, Thinking & API Guide

Google DeepMind announced Gemini 3.5 Flash on May 19, 2026 — the latest model in the 3.x Flash line and the first member of the Gemini 3.5 family. The stable API model ID is gemini-3.5-flash, replacing the gemini-3-flash-preview identifier used during the preview window.

Two things stand out about the launch. First, only the Flash tier ships today — the official announcement confirms that Gemini 3.5 Pro is rolling out next month, which inverts the usual pattern of leading with the top-tier model. Second, Google's own published benchmark table shows a Flash-tier model leading Claude Opus 4.7 and GPT-5.5 on five separate evaluations — the cleanest signal yet that the Flash / Pro tier line inside Google is narrower than the marketing label suggests.

This guide covers what shipped today, the verified benchmark comparison versus Gemini 3.1 Pro, Claude Sonnet 4.6, Claude Opus 4.7, and GPT-5.5, the new thinking_level API surface, the migration path from gemini-3-flash-preview, and the practical question of when to actually reach for 3.5 Flash versus the alternatives. Every number below is anchored to the Google blog post or the ai.google.dev docs page published the same day.

The launch lands with a deliberately small surface area on day one. One model — Gemini 3.5 Flash — across many platforms, with Gemini 3.5 Pro held back for a release next month. The announcement frames 3.5 Flash as a small, fast model that nevertheless outperforms Gemini 3.1 Pro on challenging coding and agentic benchmarks, and runs roughly 4x faster than other frontier models on output tokens per second.

Read against the benchmark table in Section 02, that framing understates the result. 3.5 Flash leads the entire reported field — including Claude Opus 4.7 and GPT-5.5 — on MCP Atlas, Toolathlon, Finance Agent v2, CharXiv Reasoning, and MMMU-Pro. That is unusual for a Flash-tier model.

The chart below is the five-benchmark headline view: every evaluation where Gemini 3.5 Flash posts the highest score in Google's published table. The full multi-vendor table follows beneath. Methodology source: deepmind.google/models/evals-methodology/gemini-3-5-flash.

The full table compares 3.5 Flash against Gemini 3 Flash (the predecessor in the 3.x Flash line, released December 17, 2025), Gemini 3.1 Pro, Claude Sonnet 4.6, Claude Opus 4.7, and GPT-5.5. Bold cells mark the field leader per benchmark.

Where it doesn't lead

Pro-tier strengths persist where you'd expect them. Opus 4.7 wins on SWE-Bench Pro (64.3%) and Humanity's Last Exam (46.9%). GPT-5.5 takes Terminal-Bench 2.1 (78.2%), GDPval-AA (1769 Elo), OSWorld-Verified (78.7%), Blueprint-Bench 2 (36.2%), MRCR v2 at 128k (94.8%), and ARC-AGI-2 (84.6%). Long-context retrieval at the 128k slice is the most obvious soft spot for 3.5 Flash versus the Pro tier — if your pipeline depends on dense mid-context retrieval rather than agentic tool calls, GPT-5.5 still leads by a wide margin.

The biggest developer-facing change is the thinking control surface. The integer thinking_budget parameter that shipped with Gemini 3 Flash Preview is replaced by a string enum, thinking_level. The new values are minimal, low, medium (default), and high.

Two notes from the ai.google.dev migration page. The default level dropped from high to medium — meaning a naive port from gemini-3-flash-preview to gemini-3.5-flash will silently reason less than the old preview did unless you opt back in. Separately, Google retuned low-effort thinking specifically for code and agentic tasks, so "low" is no longer a hint to skip reasoning — it is a tuned setting that holds up on coding workflows.

# Python — generate with thinking_level
from google import genai

client = genai.Client()

response = client.models.generate_content(
    model="gemini-3.5-flash",
    contents="Plan a 3-step MCP workflow to file a GitHub issue.",
    config={
        "thinking_config": {
            "thinking_level": "medium"  # minimal | low | medium | high
        }
    },
)

print(response.text)

Other API changes worth knowing

• Google removed its recommendation for temperature, top_p, and top_k. Strip hardcoded sampling values from production configs unless an eval explicitly justifies the override.
• Function responses now require matching id and name fields between the function call and the response. Update any tool-execution layer that previously relied on positional matching.
• Thought preservation across multi-turn is now automatic. Google flags that this can increase token usage on long sessions — budget accordingly.
• Thinking with encrypted reasoning context now combines with structured outputs, code execution on images, and the full tool set (Search grounding, URL context, function calling) without the prior compatibility caveats.

The ai.google.dev model reference for gemini-3.5-flash gives the load-bearing specs you need before adopting. Inputs span text, image, video, audio, and PDF. Output is text only — image generation, audio generation, and the Live API are explicitly not supported on this model. Anyone shipping voice agents or live conversational UX needs to stay on a different model in the Gemini family.

Pricing was not stated in the Google blog post or the ai.google.dev docs page on launch day. The official Gemini API pricing page now lists Gemini 3.5 Flash at $1.50 / Mtok input and $9.00 / Mtok output on the standard tier, with context caching at $0.15 / Mtok (plus $1.00 / Mtok-hour storage) and batch / flex tiers at $0.75 / $4.50 — a 50% discount across the board. The priority tier runs $2.70 / $16.20. For reference, GPT-5.5 is $5 / $30 and Claude Opus 4.7 is $5 / $25, so 3.5 Flash sits at roughly 3.3x cheaper on input and 2.8-3.3x cheaper on output than the current Pro tiers from OpenAI and Anthropic.

The migration is small in lines of code and slightly larger in behavior. Three changes touch most codebases.

1. Rename the model

# Before — preview identifier
model = "gemini-3-flash-preview"

# After — stable identifier
model = "gemini-3.5-flash"

2. Swap thinking_budget for thinking_level

# Before — integer budget
config = {
    "thinking_config": {"thinking_budget": 8192}
}

# After — string enum (default is "medium")
config = {
    "thinking_config": {"thinking_level": "medium"}
}

Picking the right level is the actual decision. If your previous runs used a high thinking budget, set high explicitly — the new default of medium will silently reduce reasoning effort otherwise. If your workload is mostly tool calling or code generation, try low first — Google retuned it for that exact case.

3. Drop sampling parameter overrides

# Remove these from your default config
# temperature = 0.7
# top_p = 0.95
# top_k = 40

Google removed the prior recommendation for these values. Defaults are considered tuned for the model. Keep them only if you have a specific eval that justifies the override.

For the broader picture of how Gemini's agentic stack composes — Antigravity, the IDE surface, and what 3.5 Flash means alongside them — our existing guide on Google Antigravity and the Gemini agentic IDE still applies; only the Flash-tier model behind the agent has changed.

Gemini 3.5 Flash ships across Google's full surface area on day one. The announcement lists six platforms:

• Gemini app — global rollout.
• AI Mode in Google Search — global rollout.
• Google Antigravity — developer platform.
• Gemini API — Google AI Studio and Android Studio.
• Gemini Enterprise Agent Platform.
• Gemini Enterprise.

Two related releases ship alongside the model. Gemini 3.5 Pro is confirmed in development and rolling out next month — the announcement does not give a specific date. Gemini Spark, a personal AI agent built on 3.5 Flash, is rolling out to trusted testers today, with a Beta for Google AI Ultra subscribers in the US planned for next week.

For operators tracking the broader frontier-model picture, our cross-vendor comparison on GPT vs Opus vs Gemini at the Pro tier covers how the Gemini lineup stacks up to OpenAI and Anthropic on coding, reasoning, and price-performance — the 3.5 Flash result sharpens that picture without redrawing it.

On the published benchmarks, Gemini 3.5 Flash makes the most sense when you want frontier-grade agentic behavior at a Flash-tier latency and cost profile. The matrix below maps the verified scores to concrete workload picks.