How to Use Gemini 3.6 Flash: Speed, Price, and Limits

On July 21, 2026, Google shipped three Gemini models in one announcement: Gemini 3.6 Flash, Gemini 3.5 Flash-Lite, and a restricted security specialist called Gemini 3.5 Flash Cyber^[1]. No Pro model, no keynote, no claim of a new intelligence frontier. The entire release argues one thing: teams running agents in production need fewer tokens and lower latency per task, not a bigger model.

That framing tells you who it is for. Learning how to use Gemini 3.6 Flash well is mostly about the compounding math, because Google cut the output price and the model also emits fewer tokens to do the same work. Those two effects multiply. This guide covers the official benchmark data, where 3.6 Flash honestly loses to GPT-5.6 Luna and Claude Sonnet 5, the four API changes that break carried-over code, and which of the three models actually fits your workload.

TL;DR: the July 2026 Gemini Flash lineup

Two roadmap notes hide in the same post: Gemini 3.5 Pro is still in partner testing with no public date, and Google confirmed it has started its most ambitious pre-training run yet, for Gemini 4^[1].

On reAPI, Gemini 3.6 Flash runs $1.20 input and $6.00 output per million tokens, which is 80% of Google's published rate. Flash-Lite and Flash Cyber are not on the gateway.

The efficiency play

Gemini 3.6 Flash costs $1.50 per million input tokens and $7.50 per million output, with cached input at $0.15 per million^[1]. Its predecessor charged $9 per million output, so Google cut the output price by roughly 17% while improving benchmark scores across the board.

The sticker price is the smaller half of the story. Artificial Analysis measured 3.6 Flash consuming about 17% fewer output tokens than 3.5 Flash on identical workloads^[4], and Google reported token reductions of up to 65% on Datacurve's DeepSWE benchmark^[1]. The model also takes fewer reasoning steps and fewer tool calls to finish multi-step work.

For an agent that loops dozens of times per task, those effects stack: cheaper tokens, fewer tokens, fewer loops. A per-token price comparison understates the gap.

Official benchmarks against its own predecessors

Google's launch chart compares 3.6 Flash against 3.5 Flash and the older 3.1 Pro on four agentic benchmarks^[1].

Three more gains over the prior generation sit in Google's evaluation methodology: SWE-Bench Pro at 58.7% against 55.1%, Terminal-Bench 2.1 at 78.0% against 76.2%, and the GDM-MRCR v2 long-context retrieval test at full 1M-token depth, where 3.6 Flash roughly doubles both predecessors at 54.0% against under 27%^[5].

From the model card: the knowledge cutoff moved to March 2026 from January 2025, the context envelope stays at 1M input and 64K output tokens, and the model accepts text, images, audio, and video natively^[2].

Where 3.6 Flash wins and where it loses

Google's chart only compares Gemini against Gemini. Here is the cross-vendor picture using mid-2026 published results^[5].

No model sweeps the board, and 3.6 Flash is not trying to. It leads on OSWorld-Verified computer use, on both CharXiv chart-reasoning variants, and on both GDM-MRCR long-context tests^[5]. Those are the multimodal and long-context lanes Google has consistently owned.

It clearly trails GPT-5.6 Luna on agentic coding, on both DeepSWE and Terminal-Bench. It trails Claude Sonnet 5 on ML engineering and knowledge work, and Sonnet 5's 1607 Elo on GDPVal-AA is a different weight class from 3.6 Flash's 1421^[5].

The independent read agrees. Artificial Analysis scores 3.6 Flash at 50 on its Intelligence Index, rank 21 of 187 models tracked. That is not frontier intelligence, and Google is not claiming it is. The same evaluation ranks it first on output speed in its class at 303.6 tokens per second and describes its token usage as fairly concise, while noting that time to first token, at 11.54 seconds under high thinking effort, sits at the slow end of its tier^[4].

One methodology caveat before anyone quotes the cross-vendor table: vendor scores are generally reported at each model's maximum thinking configuration, and agent harnesses differ between vendors. Treat small gaps as noise and only large ones as signal.

The practical read. If the job is raw coding-agent horsepower, GPT-5.6 Luna and Claude Sonnet 5 still hold the top of the board, at several times the price. If the job is computer use, document-heavy multimodal work, or long-context retrieval at scale, 3.6 Flash offers the best score per dollar in its bracket.

Gemini 3.5 Flash-Lite

The second release targets the opposite end of the curve. Flash-Lite runs at 350 output tokens per second as measured by Artificial Analysis, priced at $0.30 per million input and $2.50 per million output, which is one-fifth and one-third of 3.6 Flash's rates^[1][4].

The generational jump is larger here than anywhere else in the release^[1]:

• Terminal-Bench 2.1: 54.0% against 31.0% for 3.1 Flash-Lite, a 23-point move in agentic terminal coding.
• GDPVal-AA v2: 1140 Elo against 642, close to double on real-world knowledge work.
• GDM-MRCR v2 long context: 72.2% against 60.1%.

The comparison worth pausing on is against the older, bigger 3 Flash: Flash-Lite now wins SWE-Bench Pro at 54.2% against 49.6% and OSWorld-Verified at 74.0% against 65.1%^[1]. The cheapest current model beats the previous generation's mid-tier on coding and computer use.

Google positions it for high-throughput work: agentic search, bulk document processing, classification, subagent fan-out, with thinking levels configurable from minimal, the default, up to high^[1].

Gemini 3.5 Flash Cyber

The third model is the one Google will not sell you. Flash Cyber is a version of 3.5 Flash fine-tuned to find, validate, and patch security vulnerabilities, and it ships exclusively inside CodeMender, DeepMind's code-security agent, in a limited pilot for governments and trusted partners^[3].

The fine-tuning leans on assets few labs have: the OSV.dev database of more than 700,000 open-source vulnerabilities, over a decade of OSS-Fuzz results, and security-workflow data from Google-scale codebases including Chromium^[3]. Inside CodeMender, multiple Flash Cyber sub-agents analyze different code paths and merge findings into one report, which is the architectural bet: a cheap model called up to five times competing with single calls to much larger ones.

Read that honestly: Flash Cyber does not top the chart. OpenAI's dedicated GPT-5.5-Cyber leads by 2.4 points and Mythos 5 edges it by 0.6. Google's claim is competitive performance from a Flash-sized model, which is a cost claim rather than a crown^[3].

Where it does win is in Google's own deeper evaluations. On a V8 JavaScript engine test it confirmed 55 unique real issues against 47 for stock 3.5 Flash and 36 for Claude Opus 4.6, including 10 that no other model found^[3]. Google's Cloud Vulnerability Research team reports using it to find a memory-corruption vulnerability in a production service and build a working exploit chain, bypassing ASLR and W^X, within two hours^[3].

Which explains the leash. Google states the capability is dual-use, limits access to vetted defenders, deploys it only inside CodeMender's reporting workflow, and requires human approval before patches ship^[3].

How to use Gemini 3.6 Flash: four API changes that break old code

The model card describes 3.6 Flash as based directly on 3.5 Flash, the same natively multimodal sparse reasoning foundation and the same 1M/64K envelope, with the gains coming from post-training rather than a new base model^[2]. The request surface is where the work is.

1. Model IDs are gemini-3.6-flash and gemini-3.5-flash-lite^[1].
2. thinking_budget is replaced by a thinking_level string enum. 3.6 Flash defaults to medium, Flash-Lite to minimal^[1].
3. The classic sampling knobs are gone. temperature, top_p, and top_k were removed and should be deleted from request payloads^[1].
4. candidate_count is unsupported, and prefilled model turns, meaning conversations that end on a non-empty model role, are rejected^[1].

Both public models expose the full built-in tool suite, including computer use, search grounding, and code execution^[1]. Coming from 3.5 Flash, the migration is a model-string change plus deleting deprecated parameters. There is no new SDK surface to learn.

Which model should you pick

A routing heuristic straight off the published numbers:

• Default to 3.6 Flash for agent loops involving coding, computer use, document parsing, or anything multi-step. The token-efficiency gains compound exactly where agent costs explode.
• Route bulk, latency-sensitive work to Flash-Lite: extraction, classification, search summarization, subagent fan-out. Raise thinking_level only where sampled quality checks say you must.
• Keep frontier coding on frontier models. If DeepSWE-class autonomy is the job, the honest numbers say GPT-5.6 Luna and Claude Sonnet 5 still finish tasks 3.6 Flash cannot, and the premium may be worth it on low-volume, high-stakes runs.
• Flash Cyber is not a choice you get to make unless you are a government or an invited partner.

Calling Gemini 3.6 Flash alongside your other models

Notice what that routing list actually says. Three of the four bullets send work somewhere other than Gemini. The honest conclusion from Google's own benchmark table is that 3.6 Flash should win some of your traffic and lose the rest, and the split moves every time a vendor ships.

reAPI exposes Gemini 3.6 Flash through an OpenAI-compatible /v1/chat/completions endpoint, so the client already calling GPT and Claude calls it too. Note the wire model id keeps Google's dot:

from openai import OpenAI

client = OpenAI(
    api_key="YOUR_REAPI_KEY",
    base_url="https://api.reapi.ai/v1",
)

resp = client.chat.completions.create(
    model="gemini-3.6-flash",
    messages=[{"role": "user", "content": "Summarize this report and pull out the numbers."}],
    max_tokens=16000,
    stream=True,
)

Rates on the gateway are $1.20 per million input tokens and $6.00 per million output, which is 80% of Google's published $1.50 / $7.50. The gateway also exposes a native Gemini endpoint type for this model, so SDKs written against Google's own format work without a rewrite. Current rates and the request surface live on reapi.ai/models/gemini-3-6-flash and reapi.ai/docs/gemini-3-6-flash.

To be clear about coverage: Flash-Lite and Flash Cyber are not on the gateway. Flash Cyber is not available to anyone outside Google's pilot regardless of how you call it.

FAQ

Is Gemini 3.6 Flash actually cheaper than 3.5 Flash?

Yes, on output: $7.50 against $9 per million tokens, roughly a 17% cut. Because the model also emits about 17% fewer tokens for the same work, effective cost per completed task drops more than the sticker price suggests. Cached input is $0.15 per million^[1][4].

What is the Gemini 3.6 Flash context window?

1M input tokens and 64K output tokens, with text, image, audio, and video accepted natively^[2].

Do I need to change code to upgrade from 3.5 Flash?

Minimally. Swap the model string to gemini-3.6-flash, replace thinking_budget with thinking_level, and remove temperature, top_p, top_k, and candidate_count from requests^[1].

Is Gemini 3.6 Flash better than GPT-5.6 Luna or Claude Sonnet 5?

Not at agentic coding. Luna leads DeepSWE 67% to 49% and Terminal-Bench 84.7% to 78.0%, and Sonnet 5 leads MLE-Bench and GDPVal-AA^[5]. Gemini 3.6 Flash wins computer use, chart reasoning, and long-context retrieval, at a fraction of the price.

Can I access Gemini 3.5 Flash Cyber?

Not unless you are a government agency or an invited partner in the CodeMender pilot^[3].

What happened to Gemini 3.5 Pro?

It remains in partner testing with no announced date, while Google has confirmed that Gemini 4 pre-training is already underway^[1].

How fast is Gemini 3.6 Flash?

Artificial Analysis ranks it first on output speed in its class at 303.6 tokens per second, but time to first token at high thinking effort is 11.54 seconds, which is slow for its tier^[4]. Stream anything a person watches.

How do I call Gemini 3.6 Flash with the OpenAI SDK?

Point the OpenAI client at https://api.reapi.ai/v1 and set model to gemini-3.6-flash, keeping the dot in the id. The endpoint is OpenAI-compatible, so no SDK rewrite is needed.

Reading a release that competes on unit economics

This launch is worth understanding as a pricing move rather than a capability move. Google shipped no Pro tier, made no frontier claim, and put its engineering into the two numbers that decide what an agent costs: price per token and tokens per task. Both went down about 17%, and they multiply. Against that, the cross-vendor table shows 3.6 Flash losing agentic coding to GPT-5.6 Luna and knowledge work to Claude Sonnet 5, which is the correct trade for a model in this bracket rather than a flaw.

The practical version: default agent loops to it, send bulk extraction lower, keep frontier coding on frontier models, and measure effective cost per completed task instead of per-token rates, because that is the only number in which the token-efficiency gain shows up. That is how to use Gemini 3.6 Flash without paying frontier prices for work it already does well, or expecting it to win the rows it plainly loses.

References

1. Google. Introducing Gemini 3.6 Flash, 3.5 Flash-Lite, and 3.5 Flash Cyber. Retrieved July 2026 from blog.google/technology/google-deepmind
2. Google DeepMind. Gemini 3.6 Flash model card. Retrieved July 2026 from deepmind.google/models/gemini
3. Google DeepMind. Introducing Gemini 3.5 Flash Cyber and CodeMender. Retrieved July 2026 from deepmind.google/discover/blog
4. Artificial Analysis. Gemini 3.6 Flash — intelligence, performance, and price analysis. Retrieved July 2026 from artificialanalysis.ai/models
5. OfficeChai. Gemini 3.6 Flash beats Gemini 3.5 Flash and Gemini 3.1 Pro on most benchmarks. Retrieved July 2026 from officechai.com

Further reading

• reAPI. How to use Claude Opus 5. reapi.ai/blog/how-to-use-claude-opus-5
• reAPI. Gemini 3.6 Flash endpoint reference. reapi.ai/docs/gemini-3-6-flash
• reAPI. Model catalog. reapi.ai/models