Leaving the 'Pro' Seat Empty, Three Cheap-and-Fast Flash Models Instead — How Gemini's Refresh Targets the Token Bill of Keeping Agents Running

On July 21, Google refreshed three models in its Gemini "Flash" line all at once: the headliner Gemini 3.6 Flash, the cheaper Gemini 3.5 Flash-Lite, and the security-focused Gemini 3.5 Flash Cyber. None of these is a flashy top-tier model; instead, the update takes aim at the cost and speed of running agents in bulk and over long stretches in real-world settings. At the same time, the much-anticipated top-tier Gemini 3.5 Pro once again failed to appear — a clear signal of where Google is putting its energy right now.

The Pro seat sits empty as three lower-tier models line up

What stands out about this announcement is that the models being updated are the mid- and lower-tier "Flash (the workhorse)" and "Flash-Lite (the lightweight)" offerings. According to TechCrunch, the flagship 3.5 Pro — which hasn't been updated since February — was passed over yet again. Bloomberg reports that 3.5 Pro is delayed because it has failed to hit the company's internal performance targets. Google product lead Logan Kilpatrick said the company is "testing 3.5 Pro with partners and hopes to ship it soon," while also acknowledging that Google has "begun its largest-ever pretraining run, aimed at Gemini 4."

In the meantime, competitors have been shipping: OpenAI released GPT-5.5 and GPT-5.6, and Anthropic released Claude Opus 4.8 and Claude Sonnet 5. The picture is one of Google stalling at the flagship level and instead asserting its presence in the mid-tier with models that are "cheap, fast, and light on tokens."

The three models, seen through price tags and numbers

Here are the prices and key figures for each model (prices are per million tokens; all performance numbers are Google's own claims).

The main benchmarks for Gemini 3.6 Flash (versus the old 3.5 Flash) are as follows.

• DeepSWE (software engineering tasks): 49% (was 37%)
• MLE Bench (machine learning in practice): 63.9% (was 49.7%)
• OSWorld-Verified (automating PC operations): 83.0% (was 78.4%)

Where "17% fewer output tokens" pays off: the agent that never stops

What matters most in practical terms this time is not the performance gains but the fact that the model finishes the same job with fewer output tokens. Google says that on the independent Artificial Analysis Index, 3.6 Flash cuts output tokens by 17% compared with the old 3.5 Flash — and by up to 65% on DeepSWE. The output price also dropped from $9.00 to $7.50.

This "light on tokens" property is exactly what matters for use cases like coding agents. An agent calls the model dozens of times for a single request, stacking up reasoning, tool execution, and retries. When the output volume per task shrinks, the bill — which scales with "unit price × number of calls" — is affected, and the gap widens the more autonomous execution runs in bulk and over long periods. It's an update that tackles, from the model side, the very problem we've touched on repeatedly on this blog: token bills ballooning when you leave agents running.

Flash-Lite at 350 tokens per second — chewing through "search and read" at scale

The lower-tier Flash-Lite pairs one of the cheapest price tags around — $0.30 input, $2.50 output — with a speed of 350 tokens per second. Google positions it for "low-latency, high-throughput" work such as agent-driven search and processing large volumes of documents. On Terminal-Bench 2.1 it scores 54% (the old 3.1 Flash-Lite scored 31%) and 74.0% on OSWorld-Verified, raising the numbers on practical tasks even for a lightweight-tier model.

The rule of thumb for choosing between them is simple. It's now easy to set up a two-tier approach: use 3.6 Flash (or a competitor's higher-end model) for the core that handles judgment and design, and use Flash-Lite for the volume-driven steps like summarizing, classifying, and prepping search results.

Cyber "finds and fixes" vulnerabilities — but its distribution is kept tight

The third model, Flash Cyber, is a different animal. It's a security-focused model in which multiple Cyber agents coordinate within Google's autonomous agent CodeMender to discover and fix vulnerabilities and produce a consolidated report. In testing on V8 (the JavaScript engine used by Chrome and others), it reportedly detected 55 confirmed unique bugs, beating 3.5 Flash's 47 and Claude Opus 4.6's 36 (all figures reported by Google).

That said, this model is a limited release restricted to governments and a handful of trusted partners. An AI that finds vulnerabilities with high precision and even fixes them automatically is a powerful tool for defenders — but if misused, it could just as easily accelerate attacks. That double-edged nature is likely behind the decision to keep distribution tight. This is an area where people warn that "the very ability to find flaws could proliferate" alongside the convenience, and it's worth keeping in mind that it isn't something ordinary development teams can just pick up and use.

In production, what should you be checking right now?

From a development team's perspective, the following points are the near-term talking points.

• Recalculate costs: Shifting your agent's "prep" layer to 3.5 Flash-Lite or Flash-Lite and re-examining it by total output tokens can pay off more than the price list alone suggests.
• Breadth of access: 3.6 Flash and 3.5 Flash-Lite are available through Google AI Studio, Android Studio, GitHub Copilot, Antigravity, the Gemini Enterprise Agent Platform, and more. That makes them easy to drop into existing agent environments as a model.
• The numbers are self-reported: All of this round's benchmarks are figures published by Google. The safe approach is to run small tests on your own real tasks before deciding whether to adopt them.
• The flagship gap: For use cases that require the top tier to handle high-difficulty design and reasoning, it's worth treating the uncertain timing of 3.5 Pro's release as a given and keeping competitors' higher-end models on the table as well.

Because it isn't a flashy top-tier announcement, it may look understated — but it's an update that directly addresses the cost of continuously running agents. Including how Google fills the empty flagship seat, we'll be watching for its next move toward Gemini 4.

References: Google's official announcement (Gemini 3.6 Flash, 3.5 Flash-Lite, 3.5 Flash Cyber) / MarkTechPost / TechCrunch / GCN