Google's Gemini 3.6 Flash Cuts AI Agent Token Costs by Up to 65% on Long-Horizon Engineering Tasks — Gemini 3.5 Pro Testing Underway

Google Ships Three New Gemini Models

Google announced three new Gemini models on Tuesday, July 21, 2026: Gemini 3.6 Flash, Gemini 3.5 Flash-Lite, and Gemini 3.5 Flash Cyber, a model built specifically for security tasks. According to a report from ithome.com.tw, the three models target general-purpose AI agents, high-volume low-latency workloads, and software vulnerability patching respectively. VentureBeat's coverage of the launch corroborates the same rollout, reporting Google's pricing for Gemini 3.6 Flash and Gemini 3.5 Flash-Lite through its application programming interface (API) on the same day.

How Gemini 3.6 Flash Cuts Token Costs by Up to 65%

Google said token usage in some tests fell by as much as 65%, according to VentureBeat. VentureBeat specified that this 65% figure applies to long-horizon software engineering benchmarks such as DeepSWE, which measures how well agents complete multi-step engineering tasks from scratch. Separately, ithome.com.tw reported that according to the Artificial Analysis Index, Gemini 3.6 Flash's output token usage is 17% lower than Gemini 3.5 Flash, with fewer reasoning steps and tool calls needed to complete multi-step work.

Benchmark Gains for Gemini 3.6 Flash

ithome.com.tw reported specific benchmark movements between Gemini 3.5 Flash and Gemini 3.6 Flash. VentureBeat separately reported a knowledge-work benchmark comparison for the same model pair.

Gemini 3.5 Flash-Lite: Speed and Benchmark Gains Over Gemini 3 Flash

According to Inside.com.tw, Google claims Gemini 3.5 Flash-Lite — aimed at agentic search, document processing, and other low-latency, high-throughput work — outperforms the higher-specification Gemini 3 Flash on select programming and agent benchmarks:

VentureBeat further reported that Artificial Analysis measured Gemini 3.5 Flash-Lite as about twice as fast as the prior-generation Gemini 3.1 Flash-Lite.

Technical Specs: Context Window, Output Limits, and Multimodal Support

Inside.com.tw reported that the official Gemini model page lists Gemini 3.6 Flash as supporting up to 1 million input tokens along with multimodal content — text, images, video, audio, and PDF. VentureBeat's review of the official model cards adds further detail: both Gemini 3.6 Flash and Gemini 3.5 Flash-Lite share a 1-million-token input context window, a max output limit of 64,000 tokens, and the same knowledge cutoff date of March 2026.

Pricing Comparison: Gemini 3.6 Flash and 3.5 Flash-Lite vs. Prior Generations

ithome.com.tw reported Gemini 3.6 Flash pricing at $1.50 per one million input tokens and $7.50 per one million output tokens, both lower than Gemini 3.5 Flash — whose output price VentureBeat separately put at $9 per one million tokens. Inside.com.tw reported Gemini 3.5 Flash-Lite at $0.30 per one million input tokens and $2.50 per one million output tokens, with an output throughput of 350 tokens per second. VentureBeat confirmed the same $1.50/$7.50 and $0.30/$2.50 pricing pairs through Google's API listing, and added that the prior-generation Gemini 3.1 Flash-Lite remains priced at $0.25/$1.50 per one million tokens — still, per VentureBeat, Google's "most cost-efficient" model — but runs 2x slower than the new Gemini 3.5 Flash-Lite.

Gemini 3.5 Flash Cyber: A Specialized Security Model With Limits

ithome.com.tw reported that, citing potential misuse risk, Google will for now only make Gemini 3.5 Flash Cyber available through a limited pilot program to government and trusted partners. VentureBeat reported the model's performance on CyberGym, a real-world software vulnerability benchmark, where Google's CodeMender tool was capped at a maximum of 5 model calls per task, per VentureBeat:

VentureBeat also reported that Google did not disclose an exact numerical price for Gemini 3.5 Flash Cyber, stating only that it is fine-tuned "at a lower price per token than larger models."

Gemini 3.5 Pro and Gemini 4: What Comes Next

ithome.com.tw reported that Gemini 3.5 Pro is currently being tested by partners and will be released broadly once ready, and that the next-generation Gemini 4 has already begun pretraining at what Google describes as its largest scale to date. VentureBeat reported the same status through an on-the-record quote: Google technical staffer Logan Kilpatrick wrote on X, "Gemini 3.5 Pro is currently testing with partners and we plan to make it broadly available as soon as it's ready."

What This Means

The numbers reported across the three outlets point to a pricing and performance ladder rather than a single clean upgrade. Gemini 3.6 Flash costs more per token ($1.50/$7.50) than Gemini 3.5 Flash-Lite ($0.30/$2.50), yet Google's own claim — a 17% cut in output tokens and up to 65% savings on long-horizon benchmarks like DeepSWE — is about using fewer tokens per task, not a lower per-token price. Similarly, VentureBeat's reporting shows the newest, priciest Flash-Lite model ($0.30/$2.50) is twice as fast as the older, cheaper Gemini 3.1 Flash-Lite ($0.25/$1.50), which VentureBeat notes Google still calls its "most cost-efficient" option — meaning cost-efficiency and speed are being marketed as separate axes rather than a single upgrade path. On the security side, Gemini 3.5 Flash Cyber's 83.2% CyberGym score sits within roughly 2.4 percentage points of the four comparison models VentureBeat cited, but unlike the other two new Gemini models, it ships without a published price and with access restricted to a government/trusted-partner pilot — a narrower, less quantified release than Flash and Flash-Lite. Meanwhile, both outlets' reporting on Gemini 3.5 Pro and Gemini 4 places the next tier of models still in partner testing and pretraining respectively, indicating the cost and performance gains detailed here apply only to the Flash-tier lineup released on July 21, 2026.