Gemini 3.5 Flash costs 5.5 times more to run than the version it replaced. On agentic tasks, it outspends the pricier Gemini 3.1 Pro by 75 percent. OpenAI is moving prices up. Anthropic is moving prices up. Every headline this week treats this like a crisis. It isn't.
Rising AI costs are a filter. They punish the bolt-on crowd and reward the operators who actually rebuilt their workflows. If your AI strategy is "we have a ChatGPT license and a Jasper account," yes, this is bad news. If AI is the operating system underneath your marketing, the math gets better as the models get smarter.
Here's the take nobody is writing: expensive AI is cheaper than dumb AI if the workflow is right. This week handed us the receipts.
What the numbers actually say
The Decoder's benchmark data is worth sitting with. Gemini 3.5 Flash isn't just incrementally pricier. It's architecturally different. It needs more interaction steps than any rival tested on agent tasks. That's why the cost balloons. The model is doing more work per task, not just charging more per token.
That distinction matters enormously. A single-shot prompt getting more expensive is a nuisance. A model that completes a 12-step research and briefing workflow autonomously, even at 5x the per-token cost, can still be the better economic choice if it replaces 4 hours of analyst time.
The businesses panicking right now are the ones running single-shot prompts at scale. "Write me a product description." "Summarize this report." That's self-checkout at CVS. It saved a little labor but it didn't change the store. When the machine gets pricier, you feel every cent because the output was always disposable.
The cloud analogy nobody wants to hear
We've been here before. When AWS started maturing, the cloud bill shock hit every company that had just lifted-and-shifted their old server architecture into EC2. Same bloated process, now on someone else's hardware, now with a monthly invoice. Companies that actually rebuilt around cloud primitives, APIs, serverless, object storage, got faster and cheaper at the same time everyone else was complaining about costs.
AI pricing is following the same curve. Bolt-on AI is the lift-and-shift. You took your 2018 workflow, added a ChatGPT call in the middle, and called it transformation. When the API gets expensive, you pay more for the same mediocre throughput.
Operator AI is the rebuild. The workflow itself is different. Each step produces structured output the next step consumes. The model runs where it adds the most leverage, not everywhere. Cost scales with value, not with volume.
“Expensive AI is cheaper than dumb AI if the workflow is right.”
Karpathy chose Anthropic. Read that.
Andrej Karpathy, one of the architects of Tesla Autopilot and a founding member of OpenAI's core team, announced this week he's joining Anthropic. Not returning to OpenAI. Anthropic.
He said he wants to get back into frontier LLM research and called the next few years "especially formative." The man who helped build the most famous AI lab in the world looked at the landscape and picked a different destination. That's a signal, not a footnote.
Claude is increasingly built for long, multi-step agent loops. That's not marketing copy. That's architectural intent. Claude's extended thinking, its MCP integrations, its behavior in agentic workflows, these are deliberate design choices. Karpathy walking in the door suggests the next few model generations are being designed with exactly that in mind. For marketers who already built workflows around Claude, this is compounding advantage, not a footnote.
Where bolt-on breaks down in practice
Here's a specific scenario. A marketing director at a 9-location med spa chain is paying for a mid-tier AI writing tool, a separate AI social tool, and a ChatGPT Team license. Each tool runs independent queries. None of them share context. The campaign brief lives in a Google Doc. The keyword research lives in Semrush. The ad copy lives in the writing tool. Nobody knows what the other tool knows.
When AI pricing goes up 5x, that stack costs 5x more and produces the same siloed output. There's no compounding. There's no reuse. Every query starts from zero.
Contrast that with an operator workflow: the brief is a structured document the model reads. The keyword data feeds the content outline. The outline feeds the ad copy. The ad copy feeds the performance report that updates the brief. One workflow. Structured handoffs. The model runs where it creates the most leverage. When pricing goes up, you're paying more for a system that produces more. That's a different math problem entirely.
This is exactly what we laid out in Operator AI: The OS Underneath Your Marketing Stack. The post still reads like current news because the thesis hasn't changed. What changed is that pricing pressure just made the gap more visible.
DeepSeek enters the code agent race
One more signal worth clocking. DeepSeek is building a dedicated team in Beijing to develop its own AI coding agent, working title DeepSeek Code, a direct competitor to Claude Code, Codex, and Cursor. The job listings specifically call for experience with agent loops, MCP, and context engineering.
That language is telling. MCP is the protocol that lets AI models call external tools inside multi-step workflows. Context engineering is the discipline of structuring what the model knows at each step. These aren't features. They're the infrastructure of agentic AI.
When three of the most-watched AI organizations in the world are all racing to build better agent infrastructure, the direction of the industry is not subtle. The prompt box was the beginning. The workflow is the product.
The bet we're making
The businesses that absorb rising AI costs without flinching are the ones who already rebuilt their marketing workflow around AI as the operating system. Not bolted it on. Rebuilt around it. For them, a smarter model at 5x the price is still a bargain compared to the alternative.
The businesses that feel every price increase in their gut are the ones who bought an add-on tool and called it a strategy. Bolt-on AI is the 2019 Honda with a CarPlay screen. Operator AI is the Tesla. Same category. Structurally different everything underneath.
Pricing pressure just became the forcing function that makes the difference impossible to ignore. If you want to know where your current setup lands, the quiz takes three minutes and tells you exactly which path fits your team.