Gemini 3.6 Flash: Price, Benchmarks, and What's New

Google refreshed the workhorse tier of its Gemini lineup on July 21, 2026, releasing three models at once: Gemini 3.6 Flash, Gemini 3.5 Flash-Lite, and a security-specialized model called Gemini 3.5 Flash Cyber. The launch pushes down price and token consumption on the models most developers actually run in production, while the flagship the company promised for June—Gemini 3.5 Pro—remains in limited testing. Alongside the release, Google said it has begun what it called its “most ambitious pre-training run yet” for Gemini 4.

The framing is unusual. Rather than lead with a new frontier model, Google shipped incremental upgrades to its cheaper, faster tier and used the moment to advertise a generation that has not arrived. The result is a release that says as much about where the flagship is—delayed—as about what actually shipped.

What Gemini 3.6 Flash is

Gemini 3.6 Flash is the direct successor to 3.5 Flash, the mid-tier model Google positions for high-volume, latency-sensitive work: chat, summarization, classification, and increasingly agentic tasks that chain many model calls together. It ships with a 1-million-token context window, up to 64k tokens of output, native multimodal input, adjustable “thinking” controls, and built-in tools including Computer Use. Google advanced the model’s knowledge cutoff from January 2025 to March 2026.

The pricing is the headline. Gemini 3.6 Flash costs $1.50 per million input tokens and $7.50 per million output tokens, undercutting 3.5 Flash’s $9 output price while scoring higher across Google’s benchmark set. For teams running a model at scale, output price is the number that dominates the bill, and Google cut it while improving quality—the combination it needs to defend the tier against a wave of cheap, capable open-weight competitors.

The benchmark gains

On Google’s published numbers and third-party testing, 3.6 Flash beats not just its predecessor but, on several tests, the older Gemini 3.1 Pro:

• SWE-Bench Pro, a software-engineering benchmark: 58.7%, versus 55.1% for 3.5 Flash and 54.2% for 3.1 Pro.
• DeepSWE, Google’s internal coding-and-agent test: 49%, up from 37% for 3.5 Flash.
• OSWorld, a computer-use benchmark: 83%, up from 78.4%.
• On the independent Artificial Analysis Intelligence Index, 3.6 Flash scored 50, placing it above average among reasoning models in its price tier.

Just as notable as the accuracy gains is the efficiency story. Google reports 3.6 Flash uses about 17% fewer output tokens than 3.5 Flash to reach those scores on the Artificial Analysis index, and up to 65% fewer on the DeepSWE benchmark. Because output tokens are billed and also drive latency, cutting them compounds the sticker-price reduction—the effective cost per completed task falls faster than the per-token price alone suggests. That token economy is exactly what an AI coding assistant or a multi-step agent needs, where a single job may spend tens of thousands of tokens of internal reasoning before returning an answer.

Flash-Lite and Flash Cyber

The other two models target the ends of the range.

Gemini 3.5 Flash-Lite is the cheapest, lowest-latency option, aimed at simple high-throughput tasks—routing, extraction, short classifications—where a full Flash model is overkill. It rounds out a tier structure that lets developers match model cost to task difficulty rather than paying flagship prices for trivial calls.

Gemini 3.5 Flash Cyber is the more interesting outlier. Built on the 3.5 Flash foundation and specialized to detect, confirm, and remediate software vulnerabilities, it is designed to run inside Google’s CodeMender agent, which patches security flaws in code. Google is keeping it in a limited-access pilot for governments and trusted partners, citing dual-use risk: a model good at finding and fixing vulnerabilities is also a model good at finding them to exploit. That caution echoes a broader industry pattern of gating security-capable models behind vetting, and it puts Google’s own red-teaming posture on display—shipping the capability, but not to everyone. For defenders, an automated agent that closes real security holes in a codebase is a meaningful tool; for the same reason, Google is not handing it out freely.

The flagship that isn’t here

The subtext of the release is what Google did not ship. Gemini 3.5 Pro, the flagship model the company signaled for June, remains in limited testing with partners, having slipped past its originally announced date. Filling a launch slot with three refreshes to the Flash tier—and a teaser for Gemini 4—is the move of a company keeping momentum visible while its top-line model runs late.

Google said the Gemini 4 pre-training run is its most ambitious yet, but offered no date, benchmarks, or specifications. Announcing that a training run has started is a low-information signal; every large lab is always training something. What it does accomplish is narrative control: it reframes a delayed flagship as a stepping stone to a bigger one, and it lands in a month when rivals have shipped aggressively. Recent releases across the field—from Grok 4.5 to a run of open-weight models closing the gap with proprietary frontier systems—have kept pressure on Google to show it is still setting the pace.

Where you can use it

Google made 3.6 Flash and 3.5 Flash-Lite available through its developer surfaces—the Gemini API, AI Studio, and Vertex AI—and to consumers in the Gemini app. Flash Cyber, by contrast, is not generally available; access runs through the pilot program. The split matters for anyone planning around the release: the two general-purpose models are something you can build on today, while the security model is, for now, an announcement more than a product you can call.

What it means

The most concrete takeaway is economic. Google cut the price and token appetite of the model tier that carries the most real-world traffic, and it did so while raising benchmark scores. For teams building on Gemini, 3.6 Flash is a straightforward upgrade: lower cost per task, better coding and agent performance, and the same 1-million-token context they were already designing around. The efficiency gains matter most for agentic workloads, where token consumption—not the per-token rate—is usually what determines whether a use case pencils out. This is also a quiet reminder that context window size is now table stakes; the differentiation has moved to how few tokens a model burns to finish the job.

The competitive read. Shipping cheaper, more efficient Flash models is a direct response to the economics of the moment. A large language model tier is only defensible if it stays cheaper to run than the free alternatives are to self-host, and Google is fighting that battle on price and token efficiency rather than raw capability. The Flash tier, not the flagship, is where that war is fought—because it is where the volume, and the revenue, live. Google already leans on Gemini across products like AI-powered search, so the per-token cost of these models feeds straight into its own margins as well as developers’ bills.

The flagship gap is the risk. A delayed 3.5 Pro and a not-yet-real Gemini 4 leave Google narrating around an absence at the top of its lineup. Flash upgrades keep developers engaged and hold the price line, but they do not answer the question rivals are pressing: where is the model that beats the field outright? The teaser buys time. Whether Gemini 4 arrives before competitors define the next tier is the open question, and it is the one the Flash releases were partly designed to talk past.

What to watch next: whether Gemini 3.5 Pro ships on a firm date or slips further, whether Flash Cyber’s pilot expands beyond governments and trusted partners, and whether Google attaches real numbers to Gemini 4 before a rival forces its hand.