June's AI Models: Smarter at the Top, Cheaper Everywhere Else

The AI model race accelerated again in June:

• June 3 — Google released Gemma 4 12B, an open model that handles text, images and audio and runs locally on a machine with about 16GB of memory. Google says Gemma 4 has passed 150 million downloads.
• June 9 — Anthropic released Claude Fable 5, the public version of its most capable model, built for complex coding, research and long-running work. A less restricted sibling, Mythos 5, is limited to vetted partners.
• June 26 — OpenAI began previewing GPT-5.6 in three versions: flagship Sol, balanced Terra, and fast, low-cost Luna. Access starts with about 20 partner organizations, with broader availability promised in the coming weeks.

AI now comes in tiers

The important change isn't that the newest models are smarter. It's that AI companies now sell different models for different jobs: a powerful one for difficult work, a cheaper one for everyday tasks, and a fast one for high-volume automation.

The price spread makes the point. Per million input tokens, GPT-5.6 Luna costs $1, Terra $2.50 and Sol $5. Claude Fable 5 costs $10. The cheapest tier runs at a tenth of the price of the most expensive one.

Meanwhile, the gap between those frontier models and models you can run yourself keeps narrowing. Open models from Google, Alibaba and others can now reason, write code, use tools and power AI agents. They still don't consistently match the strongest proprietary models on the hardest tasks, but most business work isn't the hardest task.

Why this matters for small businesses

Until recently, using advanced AI meant sending your data to OpenAI, Anthropic or Google and paying for every request. Now there's another option:

Use frontier AI when you need maximum capability, and cheaper or local models for routine work.

That could mean a Claude or GPT model plans a project, while smaller models classify documents, handle customer requests or analyze internal data behind the scenes. Local models add one more benefit: sensitive data never leaves your building. Gemma 4 is Apache 2.0 licensed, so it's free for commercial use.

What to do

1. Sort your AI tasks by difficulty. Most will be routine: summarizing, classifying, drafting, extracting.
2. Route routine work to the cheapest model that handles it well. Test a cheaper tier before defaulting to the flagship.
3. Keep sensitive data on models you control. Customer records, financials and health information are prime candidates for a local model.

The AI market is becoming more capable and more economical at the same time. For small businesses, that matters more than which company has the smartest model this month.

---

Not sure which of your workflows belong on which tier? Let's talk.