Gemini 3.6 Flash Guide: API, Pricing and Benchmarks

Google released Gemini 3.6 Flash, Gemini 3.5 Flash-Lite, and Gemini 3.5 Flash Cyber on July 21, 2026. This guide focuses on the developer decision: which model to call, what it costs, which launch claims survive a methodology check, and what to test before changing production traffic.

The research cutoff is July 21, 2026. Product status, pricing, and model capabilities below come from Google's live API documentation, release notes, model cards, evaluation methodology, and launch posts; GitHub's Copilot rollout is sourced separately from GitHub. No paid Gemini endpoint or gated Flash Cyber system was run for this article.

What Google launched

This is a three-model announcement, but only two models are generally available through the Gemini Developer API.

Both public models accept text, image, video, audio, and PDF input and produce text. Their dedicated model pages list a 1,048,576-token input limit, a 65,536-token output limit, and a March 2026 knowledge cutoff. They support thinking, caching, code execution, file search, function calling, Google Search and Maps grounding, structured output, URL context, Batch, Flex, and Priority inference. The main capability difference in the live model tables is Computer Use: 3.6 Flash lists it as Preview, while the Flash-Lite model page lists it as unsupported. A separate documentation conflict is covered below.

Google describes Gemini 3.6 Flash as building on 3.5 Flash; it does not disclose a new architecture. The 3.6 model card carries forward the earlier model's architecture and training-data description. What changed is behavior, efficiency, price, evaluation results, and the newer March 2026 knowledge cutoff.

What changed from Gemini 3.5 Flash

For migration work, compare the production assumptions that changed.

Google says 3.6 makes fewer unwanted code edits, needs fewer reasoning steps and tool calls, and is less verbose. These are launch claims, not guarantees for an arbitrary repository. The stable interface, tool surface, and context limits make an A/B rollout relatively straightforward; the changed behavior is exactly why a blind model-string replacement is risky.

There is a meaningful cost effect even before input caching or tool charges. Google's launch chart cites roughly 17% fewer output tokens on the Artificial Analysis Intelligence Index, while the standard output price falls from $9 to $7.50 per million tokens. If a workload reproduced that token reduction, its output-charge ratio would be:

0.83 × ($7.50 / $9.00) ≈ 0.69

That is an illustrative 31% reduction in output charges, not a total-task savings guarantee. Input tokens, cached context, reasoning behavior, Search queries, tool calls, retries, and response-quality failures still affect the bill.

For the full history of the predecessor's thinking levels and Interactions migration, read the Gemini 3.5 Flash developer guide. The sections below concentrate on the July 21 delta.

Pricing and availability

The following prices are the Gemini Developer API list prices as of July 21, 2026. Amounts are USD per million tokens; output prices include thinking tokens. Google lists separate charges for tools such as Search grounding, so these values do not describe the full cost of every agent run.

Google also lists storage charges for cached content and higher Priority-inference rates. Recheck the live pricing page before setting a budget or publishing a customer-facing cost calculator.

Product availability is not one global switch:

What the benchmark results do and do not show

Google's headline results consistently favor 3.6 over 3.5 Flash, but the configurations are not identical across every row. The table keeps only results whose values and method can be mapped to Google's launch chart and July 2026 evaluation methodology.

All Gemini results are pass@1 unless Google notes otherwise. The methodology says smaller benchmarks may be averaged over multiple trials. Non-Gemini numbers in the broader model card are often provider self-reports, which is why this guide does not build a cross-vendor ranking from the launch table.

Token efficiency deserves its own artifact because a shorter answer can be either efficient or incomplete.

The right local question is not “Does 3.6 use fewer tokens?” It is “Does 3.6 pass the same acceptance tests with fewer total billed tokens, retries, and tool calls?” The smoke-test plan below measures that outcome.

Where Gemini 3.5 Flash-Lite fits

Flash-Lite is not a cheaper 3.6 setting. It is a separate stable model based on Gemini 3.1 Flash-Lite, with a lower standard price and a default minimal thinking level. Google positions it for low-latency, high-throughput jobs such as document parsing, classification, search, translation, and subagent execution.

Google cites Artificial Analysis at 350 output tokens per second. That is a third-party speed measurement surfaced in the official launch post, not a latency guarantee for your region, prompt length, concurrency, or tool stack.

The launch chart reports the following two Flash-Lite comparisons. Google's methodology says Gemini 3.5 Flash-Lite results use gemini-3.5-flash-lite with high thinking and default sampling unless a row says otherwise.

The price comparison has a tradeoff: 3.5 Flash-Lite's $0.30 input and $2.50 output rates are higher than 3.1 Flash-Lite's $0.25 and $1.50 standard rates. Test whether fewer failures and shorter workflows offset that increase. For simple deterministic extraction, the older model may still win on unit economics; for more capable subagent work, the newer model may need fewer escalations to 3.6.

Start Flash-Lite at minimal for classification and extraction, then raise it to low only when an acceptance test shows a quality gap. Do not copy the launch benchmark's high thinking setting into every high-volume request.

What Gemini 3.5 Flash Cyber is

Gemini 3.5 Flash Cyber is a security-specialized model built on 3.5 Flash and used inside CodeMender. Google says CodeMender can invoke multiple Flash Cyber agents and combine them into one report. On CyberGym, its launch configuration allows up to five Flash Cyber calls per final report, so that result represents a model-plus-agent system rather than one raw model response.

The dedicated Flash Cyber announcement provides one result that can be transcribed without reading values from an unlabeled chart:

Google says ten of the Flash Cyber issues were not found by the other two tested models. This remains a first-party, non-public evaluation; it does not establish an independently reproduced vulnerability-discovery rate or safe autonomous patching.

The deployment boundary matters more than the score. Flash Cyber was not released through the public Gemini API. Google plans to offer it soon through CodeMender only to governments and trusted partners in a limited pilot because vulnerability discovery is dual-use. There is no documented public model ID, price, self-service signup, or general release date.

Interactions API quickstart

Google now recommends the generally available Interactions API for new Gemini projects. It supports both launch model IDs and adds optional server-side conversation state, observable execution steps, and background tasks. The older generateContent API remains supported.

The current SDK documentation requires google-genai 2.3.0 or later for Interactions API support:

python -m pip install -U "google-genai>=2.3.0"
export GEMINI_API_KEY="replace-with-a-project-key"

This Python sample is adapted from Google's current Interactions and thinking examples. It was reviewed against the documentation but not executed against a billed endpoint for this article.

from google import genai

client = genai.Client()

task = """
Review this proposed API migration. Return:
1. the three highest-risk regressions,
2. one focused test for each regression,
3. a go/no-go recommendation.

Migration diff:
[paste a small, non-secret diff here]
"""

tests = [
    ("gemini-3.6-flash", "medium"),
    ("gemini-3.5-flash", "medium"),
    ("gemini-3.5-flash-lite", "minimal"),
]

for model, thinking_level in tests:
    interaction = client.interactions.create(
        model=model,
        input=task,
        store=False,
        generation_config={"thinking_level": thinking_level},
    )
    print(f"\n--- {model} / {thinking_level} ---")
    print(interaction.output_text)

store=False avoids retaining the interaction for server-side state, but it also prevents previous_interaction_id continuation and background execution. With the default stored mode, Google documents a 55-day paid-tier retention period and a one-day free-tier period, with shorter paid retention options configurable in AI Studio. Choose the state mode deliberately instead of copying a sample into a sensitive workflow.

For multi-turn agents, remember that previous_interaction_id preserves conversation history, not interaction-scoped parameters. Resend tools, system instructions, and generation configuration on every new interaction where they must apply.

What changes inside Google Antigravity

Google says Gemini 3.6 Flash is available in Antigravity on launch day. That makes it a practical candidate for iterative coding and agent teams, but the announcement does not say it is the universal default, that every account receives identical quotas, or that every existing 3.5 session is automatically upgraded.

A safe Antigravity adoption path is:

1. Select 3.6 explicitly in a disposable branch or project.
2. Re-run a task that previously caused excessive edits, repeated tool loops, or verbose output.
3. Compare accepted diff size, test pass rate, tool-call count, elapsed time, and credits consumed.
4. Repeat at medium and low thinking rather than changing both model and reasoning level at once.
5. Keep the prior model available as a rollback until the same repository-level tests pass.

The Antigravity 2.0 launch guide explains the broader desktop, CLI, SDK, and multi-agent surfaces. For budget behavior that is separate from raw Gemini API pricing, use the Antigravity credits and pricing guide. Those product credits should not be converted into API-token claims without current account-specific documentation.

GitHub Copilot rollout

GitHub announced a gradual Gemini 3.6 Flash rollout on July 21. It covers Copilot Pro, Pro+, Max, Business, and Enterprise across VS Code, Visual Studio, Copilot CLI, the GitHub Copilot cloud agent and app, JetBrains, Xcode, and Eclipse.

“Available” needs three qualifiers here:

• Rollout is gradual, so an eligible account may not see the model immediately.
• Business and Enterprise administrators must enable the Gemini 3.6 Flash Preview policy before organization members can select it.
• GitHub says the model is billed at provider list pricing under usage-based billing; the actual Copilot meter and entitlements belong to GitHub's billing documentation, not the Gemini API table above.

The Gemini API model can be GA while a specific Copilot integration uses a Preview policy. Those labels refer to different layers. Copilot also supplies its own system prompt, tools, context assembly, and agent harness, so a Gemini API benchmark should not be treated as a Copilot task-completion promise.

Migration checklist

Use this checklist for a controlled move from 3.5 Flash to 3.6 Flash:

• Pin gemini-3.6-flash in a feature flag or environment-specific configuration.
• Keep gemini-3.5-flash as the rollback during evaluation.
• Remove deprecated temperature, top_p, and top_k controls; Google's July 21 release notes mark all three deprecated.
• Make thinking_level explicit. Start at medium for 3.6 and minimal for Flash-Lite, then tune from measured failures.
• Do not send both the legacy thinking_budget and thinking_level in one request.
• Recheck tool schemas, structured-output validation, and multi-turn state. Resend interaction-scoped tools and generation settings after previous_interaction_id.
• Decide whether stored interactions are acceptable; use stateless requests where retention or data classification requires it.
• For Computer Use, run the client-side action handler in a sandbox, preserve confirmation gates, and test prompt-injection stops.
• Recalculate budget from input, cached input, output plus thinking, tool charges, retries, and failed tasks—not token list prices alone.
• Inspect the active model ID in logs. The July 21 release notes document the exact new stable ID but do not document a gemini-flash-latest alias change, so do not assume the alias has moved.

A smoke test that produces a decision

Run the same frozen corpus against 3.5 Flash, 3.6 Flash, and the relevant Flash-Lite thinking level. A useful first pass needs enough structure to reject an attractive but wrong answer.

Pre-register a pass threshold before looking at model names. For example: no regression in hidden-test pass rate, fewer unrelated diff lines, and at least 15% lower cost per accepted task. Keep raw failures, not just averages; one destructive computer-use action outweighs a small mean-score gain.

Then roll out in stages:

1. offline corpus;
2. shadow traffic with outputs hidden from users;
3. small internal cohort;
4. capped production percentage with automatic rollback;
5. full traffic only after cost, safety, and quality remain within thresholds.

Google's Gemini 3.6 Flash model card lists hallucinations and occasional slowness or timeouts as known limitations. Its reported safety results are Google-run automated evaluations, not independent production assurance. Keep output validation, timeout and retry budgets, source checks, and human confirmation for consequential actions even when local evaluations pass.

Use it or skip it

Use Gemini 3.6 Flash as the first evaluation candidate when 3.5 Flash already fits your application and output verbosity, coding loops, computer-use reliability, or per-task cost are the main constraints. Its stable ID, unchanged context limits, same input price, lower output price, and current tool support make the migration testable without an architectural rewrite.

Use Gemini 3.5 Flash-Lite for large volumes of bounded work where a smaller quality loss is acceptable and escalation is cheap. It is especially plausible for classification, extraction, routing, translation, and subagent fan-out. Verify the Computer Use documentation conflict before assigning it action-taking work.

Keep 3.5 Flash temporarily when you have a validated prompt-and-tool stack and no time to rerun acceptance tests. Stable does not mean behaviorally identical. A rollback model is cheaper than debugging a fleet-wide change without comparison data.

Skip Flash Cyber unless you are admitted to the pilot. There is no public API contract to integrate. Teams outside the limited program should evaluate generally available defensive tools rather than build against an undocumented future endpoint.

For adjacent model-selection context, compare the Gemini 3.1 Pro and Claude Opus workflow guide. For a different Gemini API shape involving generated media rather than text agents, see the Gemini Omni Flash developer guide.

Official sources and further reading

Launch, API, and pricing

• Google's Gemini 3.6 Flash, 3.5 Flash-Lite, and Flash Cyber launch
• Gemini 3.6 Flash model specification
• Gemini 3.5 Flash-Lite model specification
• Gemini Developer API pricing
• Gemini API release notes
• Interactions API overview
• Thinking levels and thought handling
• Computer Use implementation and safety guidance

Evaluation and safety

• Gemini 3.6 Flash model card
• Gemini 3.6 Flash evaluation methodology
• Gemini 3.5 Flash-Lite model card
• Gemini 3.5 Flash-Lite evaluation methodology
• Gemini 3.5 Flash Cyber technical launch

Product rollout and official social

• GitHub Copilot rollout announcement
• Google DeepMind's official 3.6 Flash comparison post

Related AgentPedia guides

• Gemini 3.5 Flash developer guide
• Google Antigravity 2.0 launch guide
• Gemini Omni Flash developer guide