The pace at which AI news arrives has started to feel different. It used to be exciting — a new model, a new benchmark, a new product launch. Now it feels more like watching a market recalibrate in real time. Every few weeks, another signal lands that quietly shifts what I believe is coming next.

I wrote about this in my most recent article — a pricing experiment by Anthropic in April that briefly removed Claude Code from its Pro subscription. It was reverted in two days, but it told a story: the economics of AI at scale don’t yet support the subscription prices we’re paying. Since then, the signals have only gotten louder. And I’ve spent the past few months building enough with AI — at work, in my personal projects, in the fabric of how I organise my days — to feel the shape of what’s coming.

Here’s my updated prediction. I think the window we’re currently in — cheap, powerful, broadly accessible AI — is shorter than most people are assuming.

The Market Signals That Changed My Thinking

On June 2, 2026, President Trump signed an executive order on AI and cybersecurity, directing federal agencies to develop a formal process for benchmarking and assessing frontier AI models before wide release. Days later, OpenAI announced three new models — GPT-5.6 Sol, Terra, and Luna — and at the U.S. government’s explicit request, limited their initial rollout to a “small group of trusted partners” whose participation was shared with the government. Not the public. Not developers. Not small businesses. A vetted group, with broader access coming “in the coming weeks.”

OpenAI says they believe in broad access, and I believe them. But the pattern matters more than the timeline. We now live in a world where government policy has formal influence over who gets to use the most capable AI models — and when. That’s new.

At the same time, Bloomberg reported in June that the industry is shifting away from flat-rate subscriptions toward token-based usage pricing — significantly more expensive at scale. According to data tracked by Ramp, business token usage grew over 1,000% between January 2025 and April 2026. The infrastructure costs behind that growth are enormous, and flat subscriptions were never designed to fund them indefinitely.

Two signals — access restriction at the frontier, pricing pressure across the board — arriving in the same month. That’s not a coincidence.

My Prediction (With All the Usual Caveats)

No one can accurately predict how this unfolds. The AI market is moving too fast, the economics are still being figured out, and the regulatory picture is being written in real time. But pressed for a prediction — here’s where I think we’re heading.

Within the next six to twelve months, access to the largest and most capable AI models will become a serious cost consideration for most individuals and small businesses. Flat subscriptions will be replaced or supplemented by token-based pricing for anything beyond basic use. And if recent government action is any guide, the most powerful frontier models may face staged rollouts as a matter of policy — not just economics. The era of “almost unlimited and almost free” is a chapter, not the whole book.

That doesn’t mean AI is going away. It means the AI you can freely access today may look different in a year’s time. And it means the workflows built around unlimited frontier access are carrying more risk than they might appear to.

Why Local Models Are Having Their Moment

This is where the hardware story becomes genuinely interesting — and I think it’s being underreported.

On June 3, 2026, Google DeepMind released Gemma 4 12B — an open, multimodal model that handles text, image, audio, and video natively on a laptop with 16GB of RAM. By Google’s own account, it nearly matches the twice-as-large 26B model across standard benchmarks. It’s available today through Ollama, LM Studio, or llama.cpp, under the Apache 2.0 license.

This is not the local AI story of two years ago, where running a model locally meant accepting significantly degraded output on a powerful desktop rig. This is a capable, multimodal model that fits in the footprint of most developer laptops. The hardware has caught up faster than most people are tracking.

Local models won’t match frontier models on raw capability. The quality gap is still real, and there are tasks where a large cloud model is genuinely better. But for a growing range of everyday work — summarising, writing, coding assistance, classification, research — a well-configured local model running on your own hardware is a realistic alternative. And for the workflows that matter most, it’s becoming a resilience floor.

There’s another shift worth watching: local models aren’t just improving — they’re specialising. Over the past year or so, new models have surfaced at a rate of roughly one or two per week. The early wave was generalists trying to compete broadly. What I expect over the next twelve months is specialisation at a much finer grain.

We already have coding-focused models — but they’re still broad, covering any language and any stack. What comes next, I think, mirrors how specialised human development teams already work: a model purpose-built for Kotlin and Android, one optimised for Swift and Xcode, another tuned for web front-end. The same logic that leads you to hire a Swift specialist rather than a generic programmer, applied to models themselves.

The economics reinforce this. A model trained deeply on a narrow domain doesn’t need to carry the weight of everything else. It can be smaller, faster, and — for the right task — still outperform a much larger generalist. Smaller doesn’t mean weaker; for a well-scoped task, it means better. And that’s where small, specialised agents and small, specialised models start to compound: pair a tightly scoped agent with a model purpose-built for exactly that kind of work, and you have a combination that will outperform a sprawling general-purpose setup on both quality and cost.

What This Means for How You Build

Here’s the practical takeaway — and it connects directly to something I’ve argued in earlier articles: the choices you make about how you structure AI workflows matter far more than which model you’re using today.

If you’ve been building agents and automation pipelines that depend on frontier reasoning and vast context windows, you’re building something that may become significantly more expensive — or less freely accessible — in the near future. That’s not a reason to stop. It’s a reason to be intentional.

Two habits, in particular, are worth starting now:

Keep your agents small and specialised. A single large agent trying to handle everything is optimised for the most capable (and expensive) models. A set of smaller, purpose-built agents — each doing one well-defined task — can be served by much lighter models. This is better design anyway; it’s also more resilient.

Break work into smaller, concrete steps. The temptation when working with frontier models is to hand over large, complex tasks in one shot — and it often works. Local models handle smaller, well-defined steps far more reliably. If you’re used to thinking in large tasks, start thinking in epics, then tasks, then specific steps. The discipline required to work well with a smaller model is the same discipline that makes workflows robust and maintainable in general — and it dramatically reduces hallucinations along the way. Worth noting: this discipline isn’t limited to software. Breaking work into well-defined steps is a habit that pays off across most areas of life — whether you’re managing a house renovation, planning a trip, or organising anything with moving parts. The project mindset transfers.

The workflows designed to run on local models are the workflows that will still work when the frontier becomes more expensive to reach.

A Bet Worth Making Now

I’m not predicting doom. I’m still paying for AI subscriptions, still using frontier models daily, still finding them genuinely transformative. But I’ve started to treat local model readiness the same way I treat backup practices: something to invest in consistently, not just when things go wrong.

Building AI workflows that are future-proof — small, specialised, thoughtfully broken down — is good practice regardless of what the market does. If the access window stays wide open, you’ve lost nothing. If it tightens, you’re already prepared. And given what the market has been signalling this month, I know which way I’m betting.

The direction of travel is clear. The window is open. The question is how long — and whether you’re building for what comes after.

Are you already experimenting with local models alongside your cloud tools? Or are you still primarily building for the frontier and not yet thinking about a fallback? I’m curious what the recent market shifts have changed in how you’re approaching your own setups — let’s talk in the comments.

#AI #LocalLLM #AIWorkflows #FutureOfWork #AIStrategy #SoftwareDevelopment #Gemma4 #IndieHackers