$ today_in_10_seconds
Google's newest Gemini models barely move the capability leaderboard — and that's the point. The headline number on 3.6 Flash is that it does the same agentic work using 17% fewer tokens at a lower price, because in 2026 the battle for agents is being fought on cost-per-task, not benchmarks. With three rival labs having just slashed prices weeks ago and inference costs down ~98% in two years, the era of the 'great token deflation' is now the defining force in what agents you can actually afford to run.
$ lead_story
Google's new Gemini Flash models compete on cost-per-task, not the leaderboard
Google released Gemini 3.6 Flash, 3.5 Flash-Lite, and a specialized 3.5 Flash Cyber — and the framing throughout is economic, not heroic. 3.6 Flash consumes 17% fewer output tokens than 3.5 Flash (up to 65% on some coding benchmarks), takes fewer reasoning steps and tool calls to finish multi-step workflows, and lands at a lower price of $1.50/$7.50 per million input/output tokens — explicitly to reduce the cost per agentic task. Flash-Lite pushes the other axis: 350 output tokens/second at $0.30/$2.50 per million, built for high-throughput agentic search and document processing. Both now ship with computer use as a built-in tool, and Google confirmed it has begun its 'most ambitious pre-training run yet' for Gemini 4. For builders, the message is blunt: the interesting optimization is no longer 'which model is smartest' but 'which model finishes the task in the fewest tokens at the lowest price' — because that's what determines whether your agent is economically viable at scale.
// why_it_matters: The frontier metric for agents has quietly shifted from capability to cost-per-completed-task — and that's the number Google chose to lead with.
$ also_worth_attention
The price war Google is responding to
Google's cost-focused launch didn't happen in a vacuum. Two weeks ago, three labs shipped flagship models within ~24 hours and all sold on price: SpaceXAI's Grok 4.5 at $2/$6 per million tokens, Meta's Muse Spark 1.1 at $1.25/$4.25, and OpenAI's Luna at $1/$6. For contrast, Anthropic's Opus 4.8 still runs $25 per million output tokens and GPT-5.6 Sol sits at $30 — and these mid-tier models deliver roughly 80% of the capability at about 5% of the cost. The insight: for most agentic workloads, the premium frontier models are becoming a luxury tier, while the real volume moves to a fiercely contested efficient middle.
link →The 'great token deflation' is the real story of 2026
Step back and the trend is staggering: achieving GPT-4-level performance cost roughly $60 per million tokens in early 2024; by early 2026 that fell by over 98%, to as little as $0.30–$0.75. Inference now accounts for an estimated 80–90% of total AI compute, which is why every lab is obsessed with squeezing it. But the caveat matters for agent builders: cheaper tokens don't automatically mean cheaper work — the metric that counts is cost per completed task, and small models (sub-27B) can now substitute for large ones on ~80% of agent subtasks. The insight: model routing, caching, and prompt compression are now core engineering skills, not afterthoughts.
link →The paradox underneath: prices crash while capex explodes
Here's the tension that makes this fragile: the same companies cutting token prices most aggressively are making the largest AI-infrastructure investments in history — Microsoft alone committed over $80 billion in data-center spend. That squeeze is reshaping strategy: Meta abandoned open weights to monetize the inference stage and fund its compute, while DeepSeek is designing its own inference chip to cut reliance on Nvidia. Google's disclosure that it has started pre-training Gemini 4 fits the same logic — win the efficiency war on today's models while spending enormously to secure tomorrow's. The insight: the token deflation is being subsidized by a capital arms race, and not everyone racing to the bottom on price will survive the bill.
link →
$ bigger_picture
The trend behind the headlines
Today reframes what 'progress' means in agentic AI. For two years the story was capability — whose model reasons best. In 2026 the decisive competition has moved to economics, because agents change the usage pattern entirely: an agent doesn't send one prompt, it runs loops of reasoning and tool calls that can burn tokens by the millions, turning AI spend into a line item finance teams now scrutinize like a second cloud budget. That's why Google led its Flash launch with token efficiency, why three rivals slashed prices this month, and why the 'great token deflation' — a ~98% cost collapse in two years — is the trend that actually governs which agents get built. But the deflation sits on top of a historic capex arms race, which means the cheap tokens you're building on are, for now, partly subsidized by labs betting they'll own the workflow later. The practical takeaway for builders: treat cost-per-completed-task as a first-class metric, get good at model routing and caching, and don't assume today's prices are the floor — or that they're permanent.
Artificial Analysis
Since the whole game is now cost-per-task, you need a neutral place to compare models on price, speed, and quality together rather than trusting each lab's launch chart. It tracks output tokens/second, price per million tokens, and quality indices across providers (Google cited it repeatedly in today's Flash announcement), making it the fastest way to sanity-check whether a 'cheaper' model is actually cheaper for your workload once token efficiency is factored in.
$ subscribe --free
Get the next edition in your inbox
// DAIly Agent covers agentic AI three times a week
// free, no spam, exit anytime