Google's July 21, 2026 launch added three Gemini models to its API in a single announcement, but its flagship tier is still missing.
Google launched Gemini 3.6 Flash, Gemini 3.5 Flash-Lite, and Gemini 3.5 Flash Cyber on July 21, 2026, delivering a major push across the Flash tier of its Gemini family. The three models target different workloads: 3.6 Flash improves coding and knowledge-task quality while cutting output-token usage by 17% compared to 3.5 Flash; 3.5 Flash-Lite delivers 350 tokens per second for high-throughput agentic pipelines at $0.30 per million input tokens; and 3.5 Flash Cyber, paired with Google's CodeMender agent framework, focuses on vulnerability detection and patching for government and trusted-partner access only. All three are part of Google's stated strategy to build AI agents "at scale," according to the official Google Blog announcement.
Key Takeaways
- Gemini 3.6 Flash costs $1.50/$7.50 per million tokens (input/output), down from 3.5 Flash, with 17% fewer output tokens on average and up to 65% fewer on DeepSWE benchmarks.
- Gemini 3.5 Flash-Lite runs at 350 output tokens/second for $0.30/$2.50 per million tokens, outperforming 3 Flash on several agentic benchmarks.
- Gemini 3.5 Flash Cyber is a security-tuned variant restricted to a limited pilot via CodeMender for governments and trusted partners. No public API pricing.
- Gemini 3.5 Pro remains delayed despite earlier promises; Gemini 4 training has already begun.
What changed in Gemini 3.6 Flash
Gemini 3.6 Flash is the next iteration of Google's cost-efficient Flash line, built on developer feedback from the 3.5 Flash release. The model costs $1.50 per million input tokens and $7.50 per million output tokens, making it noticeably cheaper than its predecessor on every workload. According to Artificial Analysis benchmarking, it reduces output token usage by 17% versus 3.5 Flash and by up to 65% on DeepSWE, a coding-focused evaluation that measures how efficiently an agent can fix software bugs. On the same benchmark, 3.6 Flash scores 49.0% compared to 3.5 Flash's 37.0%, meaning it not only uses fewer tokens but fixes more bugs correctly.
The model also improved across multiple evaluation categories. MLE Bench (machine learning research capability) jumped to 63.9% from 49.7%. OSWorld-Verified (computer-use tasks) reached 83.0% versus 78.4%. On GDPval-AA v2, a knowledge-work benchmark, it scored 1421 compared to 3.5 Flash's 1349. Customers including Hebbia and Harvey have noted particularly strong performance on multimodal tasks like document parsing and chart analysis, according to The New York Times coverage of the launch.
Security improvements accompany the quality upgrades. 3.6 Flash ships with enhanced Frontier Safety safeguards in the CBRN (Chemical, Biological, Radiological, Nuclear) and cyber-offense domains, making it more resistant to jailbreak attempts while reducing false refusals on legitimate use cases. The model card provides full technical details.
Flash-Lite: raw speed for agent workloads
Gemini 3.5 Flash-Lite targets a different niche entirely: throughput. At 350 output tokens per second as measured by Artificial Analysis, it is the fastest model in the 3.5 series. Its price point, $0.30 per million input tokens and $2.50 per million output tokens, positions it as a high-volume engine for agentic search, document processing, and other pipeline work where latency matters more than nuanced reasoning.
Now on agentic evaluations, 3.5 Flash-Lite beats 3 Flash on several specific tasks. On SWE-Bench Pro it scores 54.2% versus 49.6%, and on OSWorld-Verified it reaches 74.0% versus 65.1%. That means for certain coding and computer-use tasks, you can use a faster, cheaper model and get better results. Terminal-Bench 2.1 put it at 54% against 31% for the prior generation, and GDPval-AA v2 showed a jump from 642 to 1140, as documented in this APIDog comparison.
3.5 Flash-Lite is rolling out in Google Search alongside its API availability, expanding its reach beyond just developers. It includes computer use as a built-in client-side tool, allowing it to handle multi-step workflows without external tooling.
Flash Cyber: AI for cybersecurity, gated access
Gemini 3.5 Flash Cyber takes a completely different approach. Built on the Flash foundation and fine-tuned specifically for vulnerability detection, validation, and patching, it is not available through any public API or self-serve platform. Instead, Google is deploying it exclusively through CodeMender, its code-security agent framework, in a limited-access pilot for governments and trusted partners.
Within CodeMender, multiple 3.5 Flash Cyber agents work together to produce combined security reports. Google says the model achieves competitive performance at the frontier on the CyberGym benchmark. It has been tested internally on Chrome's production commit-scanning pipeline, reporting from the Google Blog announcement. The dual-use nature of vulnerability-finding tools drove the decision to restrict access, according to The New York Times.
No pricing information has been released because there is no open pricing tier. This keeps Flash Cyber separate from the commercial conversation around the other two models.
The 3.5 Pro gap and what's next
While three Flash models arrived, Gemini 3.5 Pro, Google's promised flagship tier, remains delayed. According to TechCrunch's reporting, internal benchmark misses, especially on coding tasks, contributed to the delay. Decrypt confirmed that partner testing is still underway with no broad release date set.
Meanwhile, Google announced that training has begun for Gemini 4, describing it as "the most ambitious pre-training run yet." According to 9to5Google's coverage, the Flash-focused strategy suggests Google is prioritizing production-grade efficiency over a monolithic flagship in the near term.
The split strategy reflects a broader shift in how companies deploy AI. According to Unite.AI analysis, the industry is moving toward specialized models for specialized work rather than pursuing a single generalist tier. The cost savings are clear: 3.6 Flash delivers both higher accuracy and lower token consumption, which makes economic sense for teams running thousands of agent calls daily.
Where does this leave developers?
For teams building production AI agents, the choice between these models depends on workload priorities. If coding quality and token efficiency matter most, 3.6 Flash is the clear pick. If raw throughput and lowest cost per task are the priority, Flash-Lite's 350 tokens/second and $0.30/M input token rate are hard to beat. The benchmark data suggests Flash-Lite may actually outperform standard Flash on several coding evals, making the "Lite" label increasingly misleading as a quality indicator.
The 3.5 Flash Cyber tier doesn't factor into most developer decisions unless you qualify for the CodeMender pilot. For everyone else, 3.6 Flash and 3.5 Flash-Lite cover the spectrum from quality-focused to speed-focused workloads, and both are available today through Gemini API, AI Studio, Android Studio, and the consumer Gemini app.
Bottom line
Google's Flash lineup got serious today. Three models, one focused announcement, and a clear message: efficiency wins. But the absence of 3.5 Pro leaves a gap that Gemini 4 will need to fill eventually.
If you're evaluating Gemini for your stack, start with 3.6 Flash's pricing tiers. The token savings compound quickly at scale, and the benchmark improvements will show in your first week of migration.
Frequently Asked Questions
Q: When did Google release Gemini 3.6 Flash, 3.5 Flash-Lite, and 3.5 Flash Cyber?
A: Google announced all three models on July 21, 2026, via an official blog post from the Gemini team.
Q: What is the pricing for Gemini 3.6 Flash?
A: Gemini 3.6 Flash costs $1.50 per million input tokens and $7.50 per million output tokens, which is lower than 3.5 Flash pricing and reduces overall cost per agentic task.
Q: Is Gemini 3.5 Flash Cyber available to the public?
A: No. Gemini 3.5 Flash Cyber is restricted to a limited pilot through CodeMender for governments and trusted partners only, with no public API or self-serve access currently available.
Q: Why is Gemini 3.5 Pro not launching at the same time?
A: According to TechCrunch and Decrypt reporting, internal benchmark misses, particularly on coding tasks, caused the delay. Google says partner testing continues with no public release date.
Q: How does Gemini 3.5 Flash-Lite compare to standard 3.6 Flash on benchmarks?
A: On several agentic benchmarks including SWE-Bench Pro (54.2% vs. 49.6%) and OSWorld-Verified (74.0% vs. 65.1%), 3.5 Flash-Lite outperforms 3 Flash, despite being the "Lite" variant. This makes it a compelling option for high-throughput coding tasks.
References
- Introducing Gemini 3.6 Flash, 3.5 Flash-Lite, and 3.5 Flash Cyber -- Google Blog, July 21, 2026
- Gemini 3.6 Flash and Gemini 3.5 Flash-Lite: Halving Time per Task -- Artificial Analysis, July 21, 2026
- Google releases three new Gemini models, but no 3.5 Pro -- TechCrunch, July 21, 2026
- Google Releases Three New Gemini A.I. Models -- The New York Times, July 21, 2026
- Google Ships New Gemini Flash Models, But Pro Is Still Missing -- Decrypt, July 21, 2026
- Google launches Gemini 3.6 Flash and teases Gemini 4 -- 9to5Google, July 21, 2026
- Google Ships Three Gemini Flash Models as Its Flagship Slips -- Unite.AI, July 21, 2026
- Gemini 3.5 Flash-Lite vs 3.6 Flash: which one should you use? -- APIDog Blog, July 21, 2026
0 Comments