Why token count is the hidden tax on AI agents
Run an autonomous software agent in production and the bill arrives in tokens, not in hours. Every reasoning step, every tool call, every draft output adds to the meter. For a workflow that fires thousands of times an hour, a model that thinks too verbosely can quietly drain a budget.
That’s the problem Google is aiming at with its latest model releases. This week it unveiled Gemini 3.6 Flash and 3.5 Flash-Lite, two models built for the unglamorous work of background agents — the kind that process documents, parse filings, and patch code without a human watching every step.
The pitch is simple: fewer tokens per task, lower latency, and pricing that makes continuous reasoning loops viable. Not chat. Not creative writing. Just efficient, repeatable work.
Gemini 3.6 Flash: the math of fewer tokens
Google’s own documentation leads with a single number: 17 percent fewer output tokens than the previous 3.5 Flash, based on measurements from the Artificial Analysis Index. In specific synthetic tests like the Datacurve DeepSWE benchmark, the company claims token usage drops by up to 65 percent.
Pricing sits at $1.50 per million input tokens and $7.50 per million output tokens. That’s positioned for continuous reasoning loops, not on-demand queries.
The performance gains are measurable. On DeepSWE, 3.6 Flash scores a 49 percent success rate versus 37 percent for its predecessor. On MLE Bench, the jump is from 49.7 percent to 63.9 percent. And on Google’s GDPval-AA v2 test, which measures real-world knowledge work rather than coding puzzles, the new model scores 1421 against 1349.
Those numbers matter for teams that have hit the ceiling of what a cheaper model can do. The trade-off used to be stark: pay more for competence, or accept mediocrity to save money. Google is trying to close that gap.
Real deployments: Figma, Hebbia, Harvey
Figma has already integrated 3.6 Flash into its prototyping infrastructure. According to Matt Colyer, Figma’s Director of Product Engineering, the model lets developers iterate faster on design without sacrificing output quality.
Legal platform Harvey and research tool Hebbia route data through the model for multimodal document work — ingesting raw financial filings, parsing structure, reading embedded charts, and producing draft reports for human review.
Google also folded a client-side computer-use tool directly into the Gemini API and Gemini Enterprise platforms. That removes the custom middleware engineers previously had to build to let models operate on an OS. The OSWorld-Verified score climbs to 83.0 percent, up from 78.4 percent, with updated safeguards against chemical, biological, radiological, and nuclear misuse.
Gemini 3.5 Flash-Lite: speed for high-volume agents
Not every agent needs deep reasoning. Some just need to process documents and search at volume. That’s the niche for Gemini 3.5 Flash-Lite.
The Artificial Analysis Index clocked it at 350 output tokens per second — the fastest in the 3.5 series, per Google. Pricing runs at $0.30 per million input tokens and $2.50 per million output tokens. Cheap enough that engineering teams can route simple, high-volume subagent requests to a minimal thinking level, reserving higher reasoning for multi-step work.
On Google’s GDM-MRCR v2 long-context test, Flash-Lite hit a 72.2 percent success rate against 60.1 percent for its predecessor. Its GDPval-AA v2 score nearly doubled, from 642 to 1140. The model also carries the same native computer-use tool as 3.6 Flash.
Separately, Google says Gemini 3.5 Pro remains in partner testing ahead of a full release, and pre-training for the next Gemini 4 architecture is already underway.
Gemini 3.5 Flash Cyber: a restricted model for patching
Automated vulnerability scanners now surface flaws faster than most security teams can patch them. That gap is where Google positions Gemini 3.5 Flash Cyber.
The model is built to validate and remediate code vulnerabilities. Google reports performance on the CyberGym benchmark competitive with frontier models, though it hasn’t released those figures with the same detail as its consumer-facing models.
Distribution stays restricted to governments and vetted partners through a pilot programme. Google frames this as a safeguard against the model generating exploit code for offensive use.
Inside Google’s CodeMender security agent, multiple instances of 3.5 Flash Cyber run in parallel, cross-checking each other’s findings before producing a single remediation report that a human reviewer signs off on. That’s a useful pattern for any team deploying agents in sensitive environments: redundancy before trust.
How to access the new models
Engineering teams can integrate these models through the Gemini API via Google AI Studio, Android Studio, or the Gemini Enterprise Agent Platform. Consumers can also access the new models in the Gemini app, and 3.5 Flash-Lite is rolling out in Google Search.
The broader takeaway is that enterprise AI agent costs are becoming the battleground for model providers. As more companies move agents from pilots to production, the models that win will be the ones that deliver acceptable results at the lowest token price. Google’s latest releases are a clear bet on that future.
For more on related developments, see AI agent efficiency strategies and enterprise AI model pricing trends.
