Enterprise AI deployments are running into a hard wall. Token budgets are exploding. Chief information officers report burning through annual allowances before June. Google says it has a fix.

At its I/O developer conference Tuesday, the company unveiled Gemini 3.5 Flash, a model that challenges a long-held industry assumption: that the most capable AI models must also be the slowest and most expensive to run. According to Google, the model outperforms its own top-tier Gemini 3.1 Pro on nearly every major benchmark while generating output tokens at four times the speed.

The cost savings are dramatic. CEO Sundar Pichai told reporters that companies running roughly one trillion tokens per day on Google Cloud could save more than $1 billion annually by shifting 80 percent of their workloads to a mix of Flash and other frontier models. The claim marks one of the most significant shifts in enterprise AI economics since large language models entered corporate computing.

The Benchmark Gap That Disappeared

For years, organizations had to choose between intelligence and speed. The smartest models were slow and expensive. Faster models made more mistakes. IT teams built complex routing systems to send simple queries to cheap models and complex tasks to expensive ones.

Gemini 3.5 Flash collapses that distinction. On Terminal‑Bench 2.1 it scores 76.2 percent. On GDPval‑AA it reaches an Elo of 1656. And on multimodal reasoning benchmarks such as CharXiv Reasoning it leads with 84.2 percent. Those numbers equal or exceed Gemini 3.1 Pro, a model Google positioned as its flagship just five months ago.

Yet Flash does this while costing one third to one half as much. An optimized variant inside Google’s Antigravity agent platform runs 12 times faster while maintaining identical quality.

Why This Matters

Enterprise AI adoption has been throttled by cost. Every customer support interaction, every legal document summary, every line of code an agent writes consumes tokens. At frontier model pricing, those tokens add up fast. Agentic workflows that autonomously execute multistep tasks burn even more.

Google says its model APIs now process 19 billion tokens per minute. Across all its surfaces, the company handles over 3.2 quadrillion tokens per month, a sevenfold increase in one year. The trend is industrywide.

Flash changes the math. CIOs can now route a far wider range of workloads to a single model without sacrificing accuracy. Engineering overhead shrinks. User experience becomes consistent. And the billion‑dollar savings Pichai cited could accelerate AI projects that were previously shelved due to budget constraints.

The model currently occupies what independent analysis firm Artificial Analysis calls the top‑right quadrant of its intelligence versus speed index. No competitor holds that position. If the performance holds at scale, Google may have shifted the economic ground beneath the entire enterprise AI market.