AI This Week: OpenAI's Three-Tier GPT-5.6, the End of Atlas, and the Shift to Smarter Systems

Big week. OpenAI shipped a new model family, quietly killed one of its own products, and — reading between the lines — signalled where the whole industry is heading next. Here’s what matters for anyone actually trying to get work done with these tools.

OpenAI ships GPT-5.6 as three tiers, not one model

The headline release is GPT-5.6, and the interesting part isn’t a single benchmark score — it’s how OpenAI packaged it. Instead of one model with a difficulty dial, GPT-5.6 arrives as three named tiers, each aimed at a different job:

• Sol — the flagship, built for the hardest work: complex coding, long-running agent tasks, research. Priced at $5 per million input tokens and $30 per million output.
• Terra — the everyday workhorse, pitched as matching the previous flagship’s quality at roughly half the cost. $2.50 input / $15 output.
• Luna — the cheap, fast tier for high-volume jobs like classification, summaries, and routine drafting. $1 input / $6 output.

Why this structure matters more than the raw capability: it gives you a real cost curve instead of one sticker price. The whole game with AI now is routing — sending the easy, high-volume work to the cheap model and saving the expensive one for tasks that genuinely need it. A three-tier lineup makes that routing decision explicit, and the gap is huge: Luna costs a fifth of Sol on both input and output. If you’re budgeting for AI features, that spread is the single biggest lever you have. (Our AI cost calculator now includes these kinds of tiers so you can model it.)

One caveat worth knowing: OpenAI previewed GPT-5.6 to the U.S. government before launch and started with a limited release, citing the model’s stronger capabilities in sensitive areas like cybersecurity. It reached general availability on July 9. This kind of pre-release government coordination is becoming a pattern for frontier models, and it’s worth watching how it shapes access going forward.

OpenAI sunsets its Atlas browser after nine months

In the same week, OpenAI announced it’s shutting down Atlas, the AI-powered web browser it launched only last October. It’ll stop working on August 9.

Atlas was OpenAI’s attempt to challenge Chrome by building ChatGPT directly into a browser. Less than a year later, the company has concluded the browser is a feature, not a destination — and it’s folding Atlas’s agent-browsing abilities into the ChatGPT desktop app and a Chrome extension instead.

There’s a lesson here that has nothing to do with browsers. Even the best-funded company in AI ships products that don’t land, and the smart move is to notice quickly and redirect rather than sink years into something that isn’t working. For any business experimenting with AI, that’s the right instinct too: try things in small, cheap bets, measure honestly, and don’t be precious about killing what doesn’t earn its place.

The real shift: from “bigger” to “smarter and cheaper”

Underneath both stories is the trend we think matters most this year. The race is no longer about who can train the biggest, most expensive model. It’s about who can build the smartest system around the models — one that picks the right tool for each task at the lowest cost.

Perplexity’s CEO put it well recently: the product going forward isn’t a single model, it’s the orchestration layer that dynamically routes each task to whichever model is most capable and cost-effective for that specific job. That reframing changes how you should think about adopting AI. You’re not marrying one model; you’re building a system that can swap models as prices drop and capabilities shift.

The other side of this: open-source models are getting good enough to handle a large share of everyday AI tasks, which puts real pricing pressure on premium proprietary tools. For businesses, that’s good news — it means the cost of useful AI keeps falling, and features that were too expensive a year ago may pencil out now. It also means you shouldn’t over-commit to any single vendor’s premium tier for work a cheaper model could do just as well.

One more thing: the scale of these companies

A financial snapshot doing the rounds this week put the size of today’s AI leaders in perspective. Companies like OpenAI, Anthropic, and SpaceX have reached valuations and growth rates that outpace essentially any tech startup of the last 25 years. Whatever you think about the hype, the money and momentum behind this shift are real — which is exactly why it’s worth staying oriented rather than tuning it all out.

The takeaway for your workflow

If there’s one thing to act on this week, it’s this: stop thinking about “which AI model is best” and start thinking about “which model for which task.” Route cheap work to cheap models, reserve the expensive ones for the jobs that need them, and keep your setup flexible enough to switch as prices fall. That’s how you get the benefit of all this without the bill that comes from defaulting everything to the most expensive option.

If you want to run those numbers for your own situation, our automation ROI calculator and AI cost calculator are built for exactly that. See you next week.