Google shipped three new Gemini models on July 21, 2026: Gemini 3.6 Flash, Gemini 3.5 Flash-Lite, and a narrow, security-focused variant called Gemini 3.5 Flash Cyber, with Gemini 3.6 Flash going generally available (GA) in the Gemini API that same day at $1.50 per million input tokens and $7.50 per million output tokens, according to Google. The announcement on Google’s official blog frames the release around one goal: getting more done with fewer tokens. Gemini 3.6 Flash replaces Gemini 3.5 Flash as Google’s mid-tier workhorse model, and it arrives with a lower output price, faster responses, and double-digit gains on several coding and reasoning benchmarks.
Gemini 3.5 Pro, the tier above Flash, did not ship in this release. Google says it remains in partner testing, with general availability promised once it clears internal checks. That gap matters for how this launch should be read: Google chose to ship its cheaper, faster tier first and let the flagship keep cooking.
The rollout covers a wide surface at once. Gemini 3.6 Flash and Gemini 3.5 Flash-Lite are live now in the Gemini API through Google AI Studio and Android Studio, inside Google Antigravity for developers, across the Gemini Enterprise Agent Platform and Gemini Enterprise app, and in the Gemini app, where Google says 3.6 Flash is now available to everyone rather than a limited rollout group. Gemini 3.5 Flash-Lite is also rolling into Google Search. That breadth points to a distribution push rather than a narrow developer preview, and it lands at a moment when OpenAI, Anthropic, and xAI have each pushed out their own mid-tier refreshes this year. It is Google’s answer for the price-sensitive, high-volume segment of that fight, the API traffic behind coding agents, support bots, and search features where cost per token decides which model wins the contract.
Don't miss new tech stories on Google
Add Tech Insider once in the Google app and our stories appear in your news suggestions.
The Benchmark Numbers: Gemini 3.6 Flash vs. Its Predecessor
Google’s benchmark disclosures show gains across several categories rather than one cherry-picked score. On DeepSWE, which measures real-world software engineering tasks, Gemini 3.6 Flash scores 49%, up from 37% for Gemini 3.5 Flash, a 12-point jump confirmed independently by Ars Technica’s launch coverage. On MLE-Bench, which tests machine learning engineering workflows, the gap is even larger: 63.9% against 49.7%, a gain of more than 14 points. OSWorld-Verified, a computer-use benchmark, climbed from 78.4% to 83%. GDPval-AA, an Elo-style knowledge-work benchmark, moved from 1349 to 1421.
- DeepSWE (coding agent tasks): 37% to 49%
- MLE-Bench (ML engineering): 49.7% to 63.9%
- OSWorld-Verified (computer use): 78.4% to 83%
- GDPval-AA (knowledge-work Elo): 1349 to 1421
None of those gains would matter much if the model got slower or pricier to run. Google says the opposite happened: Gemini 3.6 Flash uses roughly 17% fewer output tokens than Gemini 3.5 Flash to finish comparable tasks, a figure tracked by the Artificial Analysis Index and consistent with Google’s own July 2026 disclosures. The model’s knowledge cutoff also moved forward, from January 2025 to March 2026, a 14-month jump that 9to5Google’s July 2026 coverage flagged as one of the release’s more practical upgrades for developers working with recent libraries, APIs, and events.
Gemini 3.5 Flash-Lite’s Bigger Generational Jump
The smaller Flash-Lite tier actually posted larger percentage gains than its bigger sibling, just from a lower starting point. Compared with the prior Gemini 3.1 Flash-Lite, the new 3.5 Flash-Lite jumped from 31% to 54% on Terminal-Bench 2.1, from 60.1% to 72.2% on GDM-MRCR v2, and from an Elo of 642 to 1140 on GDPVal-AA v2. Google says the new Flash-Lite runs at roughly 350 output tokens per second and even beats the larger Gemini 3 Flash on SWE-Bench Pro (54.2% versus 49.6%) and OSWorld-Verified (74% versus 65.1%), despite sitting in the cheapest pricing tier Google sells.
Pricing Breakdown: Where Gemini 3.6 Flash Lands in the Market
Pricing is where Google made its most direct move against rivals. Gemini 3.6 Flash costs $1.50 per million input tokens and $7.50 per million output tokens. The input price matches Gemini 3.5 Flash exactly, but the output price drops from $9.00, a cut Ars Technica’s July 2026 launch coverage pegged at roughly 17%. Gemini 3.5 Flash-Lite costs $0.30 per million input tokens and $2.50 per million output tokens, up slightly from Gemini 3.1 Flash-Lite’s $0.25 and $1.50. Google is betting that developers will accept a small price increase at the Lite tier in exchange for benchmark scores that, in several cases, beat the previous full-size Flash model.
| Model | Input price (per 1M tokens) | Output price (per 1M tokens) | Context window | Status |
|---|---|---|---|---|
| Gemini 3.6 Flash | $1.50 | $7.50 | 1M in / 64K out | Generally available |
| Gemini 3.5 Flash (prior gen) | $1.50 | $9.00 | 1M tokens | Being phased out |
| Gemini 3.5 Flash-Lite | $0.30 | $2.50 | 1M tokens | Generally available |
| Gemini 3.1 Flash-Lite (prior gen) | $0.25 | $1.50 | 1M tokens | Prior generation |
| Gemini 3.5 Flash Cyber | Not published | Not published | 1M tokens | Invite-only pilot |
| Gemini 3.5 Pro | Not yet public | Not yet public | 1M tokens | Partner testing |
Inside Gemini 3.5 Flash Cyber, Google’s Security-Only Model
The most unusual model in this release is not the one with the highest benchmark score. Gemini 3.5 Flash Cyber is a fine-tuned variant built specifically for security vulnerability detection, and Google has plugged it directly into CodeMender, its internal system for finding and patching code flaws automatically. Access is restricted to governments and what Google calls trusted partners, and the company has not published pricing for it. That makes Flash Cyber closer to a pilot program than a public product launch, at least for now.
The move fits a broader pattern in AI model development this year: frontier labs increasingly ship narrow, task-specific variants alongside general-purpose flagships rather than betting everything on one model to do every job well. Google’s own developer documentation describes the Flash line as sustained frontier-level intelligence optimized for real-world tasks at a higher speed and lower cost, built for what the company calls the agentic era, and Flash Cyber applies that logic to one high-stakes task: finding exploitable bugs before attackers do. Google also says it strengthened Frontier Safety safeguards around CBRN and cyber-offense misuse across this release, an acknowledgment that a model built to find vulnerabilities could just as easily help create them in the wrong hands.
How Gemini 3.6 Flash Compares to GPT-5.6, Grok 4.5, and Claude Sonnet 5
Placed next to its closest rivals, Gemini 3.6 Flash lands in the middle of the pack on raw benchmark scores but at the sharper end on price. Figures compiled from the Artificial Analysis Index show OpenAI’s GPT-5.6 Luna and xAI’s Grok 4.5 both ahead of it on SWE-Bench Pro and DeepSWE, while Anthropic’s Claude Sonnet 5 leads the group on GDPval-AA and MLE-Bench. None of those three rivals matches the new model’s combination of price and benchmark movement in a single release cycle, though, and that combination is the actual selling point Google is making to developers who route high-volume traffic through an API.
Coding and Agentic Task Performance
On Terminal-Bench 2.1, Gemini 3.6 Flash scores 78%, behind GPT-5.6 Luna’s 84.7% and Grok 4.5’s 83.3%, but ahead of Gemini 3.1 Pro’s 73.8%. On DeepSWE, its 49% trails GPT-5.6 Luna’s 67% and Grok 4.5’s 54%, though it still more than doubles the 12% that Google’s own Gemini 3.1 Pro posts on the same test.
Reasoning and Long-Context Performance
The picture flips on long-context recall. Gemini 3.6 Flash posts 91.8% on GDM-MRCR v2, the highest score in the comparison set and well clear of Claude Sonnet 5’s 71.6% and GPT-5.6 Luna’s 74.8%. On CharXiv Reasoning, its 85.2% edges out every rival Google was measured against, including Claude Sonnet 5’s 77%.
| Benchmark | Gemini 3.6 Flash | GPT-5.6 Luna | Grok 4.5 | Claude Sonnet 5 |
|---|---|---|---|---|
| SWE-Bench Pro | 58.7% | 62.7% | 64.7% | 63.2% |
| DeepSWE v1.1 | 49% | 67% | 54% | 54% |
| Terminal-Bench 2.1 | 78.0% | 84.7% | 83.3% | 80.4% |
| MLE-Bench | 63.9% | 47.6% | 43.2% | 66.9% |
| OSWorld-Verified | 83.0% | 72.6% | Not disclosed | 81.2% |
| GDPVal-AA (Elo) | 1421 | 1584 | 1535 | 1607 |
| CharXiv Reasoning | 85.2% | 82.7% | 81.6% | 77.0% |
| GDM-MRCR v2 (128k) | 91.8% | 74.8% | 81.4% | 71.6% |
| Output price (per 1M tokens) | $7.50 | $6.00 | $6.00 | $15.00 |
That spread explains why Google is not claiming an outright win. Gemini 3.6 Flash is not the top scorer on most individual benchmarks against GPT-5.6 or xAI’s latest Grok releases. What it offers is a strong result at roughly half the output price of Claude Sonnet 5 and below Gemini 3.1 Pro’s $12.00 output rate, which is a different argument aimed at teams that care more about cost per million tokens than winning a single leaderboard.
Why Google Is Betting on Flash Instead of a Pro-Tier Release
Ars Technica’s framing of the launch is direct: Google is replacing Gemini 3.5 Flash with Gemini 3.6 Flash while Gemini 3.5 Pro, the more capable and more expensive tier, stays in testing. That sequencing looks deliberate. Flash-tier models handle the overwhelming majority of production API traffic, the coding agents, chatbots, and automated workflows that call a model thousands of times a day, where cost per call compounds fast. A Pro-tier model matters for the hardest reasoning tasks and for benchmark bragging rights, but it rarely moves Google Cloud’s API revenue the way a cheaper, faster Flash refresh does.
There is a simpler explanation too. Shipping is easier when the bar is lower. A model marketed for real-world tasks at higher speed and lower cost invites less scrutiny than a Pro-tier flagship expected to beat GPT-5.6 and Claude Sonnet 5 on the hardest reasoning benchmarks. Google banks a real, shippable win now, while Gemini 3.5 Pro keeps cooking behind closed doors.
Market Impact: What This Means for Enterprise AI Buyers
For engineering teams choosing where to route production traffic, a 17% output-price cut is not trivial at scale. A team spending $50,000 a month on Gemini 3.5 Flash output tokens would save roughly $8,500 a month by switching to Gemini 3.6 Flash at equal usage, before counting the added token-efficiency gain, which compounds the savings further on top of that. Google made the switch easy on purpose: the new model sits in the same API slot, available through the same Gemini Enterprise Agent Platform and Vertex AI tooling teams already have wired into their pipelines.
The bigger effect may land in procurement conversations. Enterprise AI budgets in 2026 increasingly get split across multiple model providers instead of locked to one vendor, and a cheaper, faster Flash tier hands Google’s sales teams a stronger argument in head-to-head evaluations against Amazon Bedrock and Microsoft’s Azure AI Foundry. Whether that argument holds depends on how quickly OpenAI and xAI answer with their own flash-tier price moves, something both companies have done within weeks of a Google pricing change in past cycles.
Historical Context: Google’s Flash Strategy Since Gemini 1.5
The Flash line has been Google’s answer to cost-conscious developers since Gemini 1.5 Flash arrived in 2024, and the company has iterated on it aggressively since, cycling through 2.0 Flash, 3.1 Flash-Lite, 3.5 Flash, and now 3.6 Flash within roughly two years. Each generation has followed a similar playbook: cut price or hold it flat, while pushing benchmark scores up through better training data and more efficient architectures rather than simply scaling parameter count. Gemini 3.6 Flash’s 17% output-price cut and 14-month-newer knowledge cutoff fit that pattern closely.
What differs this time is the explicit agentic framing. Google’s developer documentation for 3.6 Flash emphasizes multi-step orchestration and full-stack code refactoring, tasks that assume the model is calling tools and taking actions across a session rather than answering one prompt. That reflects where the wider industry moved during 2026: DeepSWE, Terminal-Bench, and OSWorld-Verified all score an agent completing a multi-step task, not just producing one good response.
The Road to Gemini 4: What Google Has Confirmed So Far
Google used the announcement to confirm two things about what comes next. First, Gemini 3.5 Pro remains in partner testing, with general availability promised only once it clears Google’s internal bar and no firm date attached. Second, Google confirmed it has begun pretraining Gemini 4, describing the run internally as its most ambitious pretraining effort to date. Beyond that acknowledgment, Google has not detailed what Gemini 4 targets, what it will cost, or when partner testing might begin.
Industry Data and Analyst Signals
Independent coverage moved fast, and it largely agreed. Ars Technica and 9to5Google each published detailed breakdowns within a day of Google’s blog post, converging on the same core figures for pricing and benchmark deltas, a level of agreement that is not automatic for same-day AI launch coverage. International tech press followed just as quickly: Japan’s Gigazine and the China-focused outlet 36Kr both ran writeups within a day of the English-language announcement, a sign Google timed this as a global release rather than a staged regional rollout.
The pricing data also shows where Gemini 3.6 Flash sits inside a broader price war across the model market. GPT-5.6 Luna undercuts it on price at $1.00 input and $6.00 output per million tokens. Grok 4.5 charges more upfront at $2.00 input but matches GPT-5.6 Luna’s $6.00 output price. Claude Sonnet 5, at the top of that set, costs $3.00 input and $15.00 output, roughly double the output rate Google is now charging. None of the three rivals has matched Google’s combination of a competitive input price and genuinely large benchmark gains inside one release cycle, which is the specific comparison Google’s product and pricing pages are set up to invite.
What This Means for AI Coding Assistants and Agent Platforms
Coding assistants and agent platforms are the most price-sensitive consumers of Flash-tier models, since they call the underlying model dozens or hundreds of times per session to plan, edit, and verify code. A 12-point DeepSWE gain paired with a 17% price cut changes the math for any tool that lets developers pick a backend model, a group that already covers most major AI coding assistants on the market today.
The switch itself is a small change for developers already on the Gemini API. Google’s model documentation shows the update is limited to the model identifier in an existing request:
curl https://generativelanguage.googleapis.com/v1beta/models/gemini-3.6-flash:generateContent \
-H "x-goog-api-key: $GEMINI_API_KEY" \
-H "Content-Type: application/json" \
-d '{"contents":[{"parts":[{"text":"Summarize this pull request diff"}]}]}'No new authentication flow, SDK version, or account tier is required for the Flash and Flash-Lite models, according to Google’s documentation. That low switching cost is part of why a Flash-tier refresh, rather than a Pro-tier flagship, is the release most likely to shift real production traffic in the next few months.
Five Predictions for the Next Six Months of the Model Race
Gemini 3.5 Pro reaches general availability before the end of Q3 2026. Google’s own language, still in testing with availability promised, reads like a model close to done rather than early in development.
Coding agent platforms add Gemini 3.6 Flash as a routing option within weeks. The benchmark and price combination is difficult for cost-sensitive tooling to ignore once developers start comparing their monthly API bills.
OpenAI or xAI answers with a flash-tier price adjustment within the next quarter. Both companies have moved quickly on pricing in past cycles, and Google’s 17% output-price cut sets a new floor rivals will feel pressure to match.
Gemini 3.5 Flash Cyber’s access expands past governments and trusted partners if CodeMender’s results hold up. A specialized vulnerability-detection model with strong internal results is a hard thing for Google to keep fully private once enterprise security teams learn it exists.
Gemini 4 stays in pretraining through the rest of 2026. Google has not attached any release window to it, and Gemini 3.5 Pro has not even reached general availability yet, so a next-generation flagship launch this year looks unlikely based on the current pattern.
Frequently Asked Questions
What is Gemini 3.6 Flash?
Gemini 3.6 Flash is Google’s updated mid-tier AI model, which went generally available (GA) in the Gemini API on July 21, 2026, to replace Gemini 3.5 Flash. It costs $1.50 per million input tokens and $7.50 per million output tokens, per Google, and it posts double-digit gains over its predecessor on coding, computer-use, and knowledge-work benchmarks — including a jump to 1421 on the GDPval-AA Elo scale from 1349, as 9to5Google reported in July 2026.
When did Gemini 3.6 Flash launch?
Google announced Gemini 3.6 Flash, alongside Gemini 3.5 Flash-Lite and Gemini 3.5 Flash Cyber, on July 21, 2026, with availability starting the same day across the Gemini API, Gemini app, and Gemini Enterprise Agent Platform.
How much does Gemini 3.6 Flash cost?
Gemini 3.6 Flash costs $1.50 per million input tokens and $7.50 per million output tokens, a 17% cut on output pricing compared with Gemini 3.5 Flash’s $9.00 output rate.
Is Gemini 3.5 Pro available yet?
No. Google says Gemini 3.5 Pro remains in partner testing and will reach general availability once it clears internal checks. The company has not given a firm release date.
What is Gemini 3.5 Flash Cyber used for?
Gemini 3.5 Flash Cyber is a fine-tuned model built for security vulnerability detection. Google uses it inside CodeMender, its automated code-patching system, and access is currently limited to governments and trusted partners, with no public pricing.
How does Gemini 3.6 Flash compare to GPT-5.6 and Claude Sonnet 5 on price?
Gemini 3.6 Flash is more expensive than GPT-5.6 Luna’s $6.00 output rate but roughly half the cost of Claude Sonnet 5’s $15.00 output rate. On input price, it costs less than Claude Sonnet 5 and Grok 4.5 but slightly more than GPT-5.6 Luna.
Can developers switch to Gemini 3.6 Flash without code changes?
Mostly yes. Google’s documentation shows the switch only requires updating the model identifier in an existing Gemini API call. No new authentication flow, SDK version, or account tier is needed for the Flash or Flash-Lite models.
What is Gemini 4 and when will it launch?
Gemini 4 is Google’s next-generation flagship model, currently in what the company calls its most ambitious pretraining run to date. Google has not confirmed a release window, and with Gemini 3.5 Pro still in testing, a 2026 launch looks unlikely.
