When did Gemini 3.6 Flash launch?

Google announced Gemini 3.6 Flash, 3.5 Flash-Lite, and 3.5 Flash Cyber on 21 July 2026. 3.6 Flash and 3.5 Flash-Lite are available the same day across Gemini API, AI Studio, Android Studio, Gemini Enterprise surfaces, and the Gemini app; Flash-Lite is also rolling into Search.

How much does Gemini 3.6 Flash cost on the API?

Google’s launch post lists $1.50 per million input tokens and $7.50 per million output tokens for 3.6 Flash. 3.5 Flash-Lite is listed at $0.30 input and $2.50 output per million tokens. Always re-check ai.google.dev pricing before you budget.

Is 3.6 Flash better than 3.5 Flash?

Google positions 3.6 Flash as better on coding, knowledge work, and multimodal tasks while using fewer output tokens (about 17% fewer on the Artificial Analysis Index claim) at a lower list output price. Treat vendor evals as a starting point and re-run your own agent harness before swapping production defaults.

What is Gemini 3.5 Flash-Lite for?

Flash-Lite is the speed and unit-cost tier in the 3.5 class. Google cites ~350 output tokens per second on Artificial Analysis and prices it for high-throughput agent steps, search-style subagents, and document processing. Use it when latency and QPS dominate, not when a single hard coding ticket needs maximum quality.

Can I use Gemini 3.5 Flash Cyber?

Not as a general public API model. Google states 3.5 Flash Cyber is limited to governments and trusted partners via the CodeMender pilot for finding and fixing vulnerabilities. Dual-use risk is the reason for restricted access.

Is Gemini 3.5 Pro available yet?

As of the 21 July 2026 post, 3.5 Pro is testing with partners and not broadly available. Google also says pre-training for Gemini 4 has started. Plan near-term agents around Flash / Flash-Lite, not Pro.

Should I switch my production agents to 3.6 Flash immediately?

If you already run Gemini Flash agents, yes for a controlled A/B: same tickets, same tools, score cost × review minutes × tool-call count. If you live in Claude Code or Cursor with another frontier model as default, treat this as competitive pricing pressure first, not an automatic migration.

Gemini 3.6 Flash: What Changed, Pricing, Agents

Gemini 3.6 Flash is Google’s 21 July 2026 workhorse model for production agents: stronger coding and knowledge work than 3.5 Flash, with a claimed cut in output-token waste and list API rates of $1.50 / $7.50 per 1M tokens. The same post ships 3.5 Flash-Lite for high-throughput steps and a restricted 3.5 Flash Cyber path inside CodeMender.

This is a PromptHive launch analysis, not a press rewrite. Facts come from Google’s announcement and related DeepMind model-card links. Vendor benchmarks are their measurements — we translate them into decisions for people already using Gemini, building on the Gemini API, or choosing agent stacks.

Verified 2026-08-05 against Google’s public 21 July 2026 launch materials. Model IDs, plan packaging, and API rates can move; re-check official pricing before you budget.

The short answer

3.6 Flash — new default workhorse: better coding / knowledge / multimodal claims, ~17% fewer output tokens vs 3.5 Flash on the Artificial Analysis Index claim, $1.50 in / $7.50 out per 1M tokens.
3.5 Flash-Lite — fastest 3.5-class tier Google highlights (~350 tok/s on AA), $0.30 / $2.50 per 1M, aimed at volume and low latency.
3.5 Flash Cyber — specialized for vulnerability find-and-fix only via CodeMender, governments and trusted partners.
Availability — Gemini API, AI Studio, Android Studio, Gemini Enterprise surfaces, Gemini app; Flash-Lite also rolling to Search. 3.5 Pro still partner-testing.
PromptHive take: if your agents already run on Gemini Flash, re-baseline on 3.6 Flash and push routine sub-steps to Flash-Lite. If you only use consumer Gemini in Gmail/Docs, this is a quieter upgrade — still start from our beginners guide.

What was officially announced

From Google’s 21 July 2026 Keyword post (Tulsee Doshi on behalf of the Gemini team):

Model	Official role	Headline economics (Google)
Gemini 3.6 Flash	Workhorse for agents: coding, knowledge work, multimodal	$1.50 / $7.50 per 1M input/output; fewer tokens per task vs 3.5 Flash
Gemini 3.5 Flash-Lite	Fastest / cheapest 3.5-class for high throughput	$0.30 / $2.50 per 1M; ~350 output tokens/s (Artificial Analysis claim)
Gemini 3.5 Flash Cyber	Cyber-focused model paired with CodeMender agents	Limited access; not a general public chat model

Google also states:

3.5 Pro is testing with partners; broad availability “as soon as it’s ready.”
Pre-training for Gemini 4 has started (future, not a ship date).
Computer use is a built-in client-side tool via Gemini API and Gemini Enterprise for the new Flash-class models.
Frontier Safety work on CBRN and cyber-offense misuse for 3.6 Flash, with a public model card.

Surfaces listed for day-one / rolling access: Gemini API (AI Studio), Android Studio, Gemini Enterprise Agent Platform / Enterprise app, Gemini app, and Google Search (Flash-Lite rollout). 3.6 Flash is also called out in Google Antigravity.

What’s new compared with 3.5 Flash

Dimension	3.5 Flash era	3.6 Flash (announced)
Role	Prior Flash workhorse	Explicit agent-scale workhorse with efficiency pitch
Token efficiency	Baseline	Google cites ~17% fewer output tokens on Artificial Analysis Index vs 3.5 Flash; larger savings claimed on some DeepSWE-style runs
API list price	Higher prior Flash pricing (check current docs if you still run 3.5)	$1.50 / $7.50 — Google frames as lower cost per agentic task
Coding agents	Strong Flash tier	Claims higher precision, fewer unwanted edits / execution loops (e.g. DeepSWE 49% vs 37% in the post)
Knowledge / multimodal	Competitive	GDPval-AA and customer quotes (Hebbia, Harvey) on docs, charts, reports
Computer use	Evolving	Built-in tool path; OSWorld-Verified 83.0% vs 78.4% in Google’s table
Safety posture	Prior Flash safeguards	Enhanced CBRN + cyber-offense resistance claims; fewer beneficial-use refusals as a training goal

What did not change for readers: you still need tests, git, and human review for anything that edits production systems. A more efficient agent that loops less still needs a success signal — see our agents guide.

Features explained (without the brochure)

3.6 Flash as the “master” agent model

Google’s demos and customer framing push 3.6 Flash as the model that plans and orchestrates, not only the one that answers a single prompt. Efficiency matters here: fewer reasoning steps and tool calls reduce both bill and latency variance. That is the practical upgrade path for teams already on Managed Agents / multi-agent orchestration, not a new chat personality for casual use.

3.5 Flash-Lite as the volume tier

Flash-Lite is the model you put on subagents: extract features from a large catalog, translate receipts, fan out design variants, run high-QPS search-style steps. Google highlights configurable thinking levels — low for cheap high-volume work, higher for multi-step subagent load. On several coding/agentic evals in the post, Flash-Lite even beats older full Flash generations, which is a pricing trap if you leave every call on 3.6 Flash by habit.

Computer use as a first-class tool

“Computer use” moving into the API/Enterprise toolkit means agents can act in UI environments without every team reinventing browser automation glue. It does not mean unattended desktop control is safe. Keep human approval gates for anything that can spend money, send mail, or change infrastructure — same rule we use across MCP and coding agents.

3.5 Flash Cyber + CodeMender (limited)

Google’s cyber story is intentionally not “everyone gets a red-team model.” 3.5 Flash Cyber is fine-tuned for vulnerability discovery and patching, composed as multi-agent work inside CodeMender, with limited pilot access for governments and trusted partners. If you need general secure coding assistance, that is a different product path (code review agents, your existing IDE tools) — not Cyber Flash on the public Gemini app.

Safety you will feel

Enhanced jailbreak resistance on CBRN and cyber-offense paths can change refusal behaviour for dual-use prompts. Beneficial-use refusal reduction is the stated counterweight. Automation that assumed “Flash always answers” should log model ID, refusal rate, and fallbacks after the switch.

Pricing and availability

Channel	What Google states (public launch post)
API — 3.6 Flash	$1.50 input / $7.50 output per 1M tokens
API — 3.5 Flash-Lite	$0.30 input / $2.50 output per 1M tokens
Developers	Gemini API via AI Studio; Android Studio; Antigravity called out for 3.6 Flash
Enterprise	Gemini Enterprise Agent Platform; 3.6 Flash in Gemini Enterprise app
Consumer	Gemini app; Flash-Lite rolling into Search
3.5 Flash Cyber	CodeMender pilot — governments / trusted partners only
3.5 Pro	Partner testing — not general availability in this announcement

Consumer vs API trap: free and paid Gemini app packaging is a separate story from API list rates. For Workspace-native chat and Gmail/Docs behaviour, start with our Gemini beginners guide and tool review. For study/source-grounded work, NotebookLM remains the dedicated Google product — not a Flash SKU rename.

Agent unit economics: list price × tokens is only half the bill. Google’s efficiency claim (fewer output tokens, fewer tool loops) is the other half. Measure cost per closed ticket, not only rate cards.

PromptHive hands-on analysis

We did not re-score every public leaderboard. Practical reading for PromptHive readers:

This is an agent packaging release, not a “new Gemini personality” moment. The useful mental model is a two-tier Flash ladder: 3.6 for hard orchestration, Flash-Lite for volume steps.
Token efficiency is the real product claim. A 17% output cut on AA is marketing-adjacent until you see it on your tool-using traces. Instrument output_tokens and tool-call counts before and after.
Computer use + agents only pays off with supervision. Pair with the same discipline we recommend for coding agents and MCP: smallest tools, explicit success checks, human merge.
Competition frame: Claude Opus-class and GPT-5.6 Sol still own “expensive smart” for many coding teams (Claude Opus 5, GPT-5.6 Sol/Terra/Luna). Gemini’s pitch here is cheaper capable agents at scale, especially if you already sit on Google Cloud / Workspace.
Cyber Flash is not your pen-test chatbot. Restricted access is a feature of the risk model, not a missing checkbox on the free tier.
Pro is still “soon.” Do not design a roadmap that requires 3.5 Pro next week.

Pros

Clear workhorse + volume tier split (3.6 Flash / Flash-Lite) with published API rates
Explicit efficiency story (fewer tokens / tool loops) that can lower cost per task
Broad surface list: API, Studio, Android Studio, Enterprise, Gemini app, Search (Lite)
Computer use called out as a built-in path for agent workflows
Safety materials and model cards linked from the launch story

Cons

Vendor evals (DeepSWE, OSWorld, AA Index) are not your production suite
3.5 Pro still not GA — middle tier of the broader family remains fuzzy for planners
Cyber model is gated; security teams cannot assume self-serve access
Consumer Gemini packaging vs API SKUs still confuses buyers who only see “Gemini” in the app
Multi-vendor shops still face glue cost (auth, tools, eval harness) even when tokens get cheaper

Best use cases

Use case	Fit
Multi-step coding / migration agents on Gemini API	Primary — 3.6 Flash
Document, chart, and report agents (multimodal knowledge work)	Primary — 3.6 Flash
High-QPS extract / classify / translate subagents	Primary — 3.5 Flash-Lite
Computer-use style UI automation with human gates	Strong on Google’s framing — start small
Everyday Gmail/Docs help	Consumer Gemini path — see beginners guide
Source-grounded study from your own files	Prefer NotebookLM, not Flash SKU shopping
Public unrestricted vulnerability exploitation	No — Cyber is limited and dual-use restricted
“I only want free autocomplete in any IDE”	Not this launch — see best free AI coding tools

How to trial the switch (about 45 minutes)

Pin model IDs in a branch of your agent config (3.6 Flash for the planner; Flash-Lite for one high-volume worker).
Replay ten real tickets or workflows you already log (not toy prompts).
Score: success rate, human interventions, wall-clock, output tokens, tool-call count, dollar estimate.
Keep tools and system prompts fixed for the first pass so the model is the only variable.
Only then retune thinking levels on Flash-Lite or shrink planner prompts on 3.6 Flash.

If cost-per-success drops without review minutes rising, make 3.6 Flash the default planner. If it burns more tools while looking “smarter,” your harness — not the rate card — is the bug.

When to stay on another stack

Situation	Prefer
Already deep in Claude Code / Anthropic Max	Claude Opus 5 analysis and Claude Code — switch only with a harness win
ChatGPT / Codex as the paid home	GPT-5.6 tiers, Codex vs Claude Code
Editor-first flat bill with multi-file agents	Cursor (model picker still matters)
Research grounded only in your PDFs	NotebookLM
MCP tool mesh across vendors	What is MCP? then pick models per tool risk

What this means for PromptHive clusters

Chatbots: consumer Gemini remains the Workspace-native pick; this launch mainly upgrades the API agent story behind it (Gemini review).
Agents: reinforces “cheap + reliable loops” over single-shot IQ. Update defaults in the complete AI agents guide mental model: planner vs worker tiers.
Coding: Gemini is more competitive for API-orchestrated coding agents; IDE-native defaults still live in Cursor / Copilot / Claude Code (coding guide).
Research / study: do not conflate Flash upgrades with NotebookLM’s source-grounded product (NotebookLM study guide).

Final PromptHive verdict

Gemini 3.6 Flash is a real workhorse update if you run agents on Google’s stack: better quality claims plus a token-efficiency and list-price story aimed at cost-per-task, not a new consumer chatbot brand. Pair it with 3.5 Flash-Lite for volume steps, treat Cyber as a closed security pilot, and do not block roadmaps on 3.5 Pro until it is actually GA. If you are outside Google Cloud and already happy on Claude or OpenAI coding products, run a harness comparison before you rewrite tooling — cheaper tokens do not erase migration cost.

CTA: Take one production agent, swap only the planner to 3.6 Flash, log tokens and review minutes for a week, then decide team defaults. Start from how to use Gemini if you are app-first, or the agents and coding guides if you are building loops.