Ninety Minutes Apart and Half the Price

Ninety minutes. That’s roughly the gap between Anthropic putting Claude Opus 5.5 into the world on September 22, 2026, and OpenAI following with GPT-6 Sol and GPT-6 Luna at half its previous API prices. Two of the most-watched labs in AI shipped new models on the same afternoon, within the length of a movie, days after both had publicly argued that frontier AI development should slow down.

I want to walk through what that actually means for you, especially if you’re someone who uses AI tools without building them.

What happened, in plain terms

Both companies had recently added their voices to calls for caution around advanced AI, citing existential risk. Then both released new models. Anthropic’s was Claude Opus 5.5. OpenAI’s were GPT-6 Sol and GPT-6 Luna, launched with token prices cut in half compared to what came before. Anthropic’s release also lands ahead of a planned IPO, according to the Financial Times.

The framing from both companies leaned commercial. These are cost-effective models. Cheaper to run, cheaper to call, pitched at businesses and developers who care about the bill at the end of the month.

A quick note on “tokens”

If you’ve never had to think about this: AI companies charge by the token, which is roughly a chunk of a word. Every question you ask and every answer you get costs tokens. When OpenAI halves token prices, it means the same amount of AI work now costs half as much for whoever is paying the API bill. That’s usually the company behind the app you’re using, not you directly, though it eventually shows up in what products cost and what they’re willing to give away free.

The apparent contradiction

Calling for a slowdown and then shipping models looks inconsistent, and plenty of people online have said so. But there’s a distinction worth pulling apart.

A slowdown argument is usually about capability at the frontier: making systems more powerful, more general, more capable of things nobody has tested for. Cheaper models are a different axis. They’re about making existing capability accessible to more people at lower cost. The labs can argue, with some internal logic, that they didn’t push the ceiling higher this week. They lowered the floor.

Whether that distinction holds up is a fair question. Making capable AI dramatically cheaper means far more of it gets used, in far more places, by far more people who haven’t thought much about what could go wrong. Cheap is its own kind of scale. If your concern is the total volume of AI acting in the world, a price cut may matter more than a capability bump.

Why this is actually an agents story

This site is about AI agents, and price is the single biggest thing standing between agents being a demo and agents being ordinary.

Here’s why. When you chat with an AI, you send one message and get one answer. That’s a handful of tokens. When an agent does a task for you, it loops: it reads, it plans, it tries something, it checks the result, it adjusts, it tries again. A single agent task might involve dozens of model calls. The token cost of one agent finishing one errand can dwarf the cost of a long conversation.

So when the price of model calls drops by half, the math on agents changes in ways that are easy to underestimate:

• Tasks that were too expensive to automate become worth automating.
• Agents can afford to double-check their own work, which tends to make them more reliable, not just cheaper.
• Products can let agents run longer before someone has to step in and approve the spend.
• Smaller companies and solo builders can ship agent features that previously only well-funded teams could afford to run.

That last one is the part I find most interesting. A price cut is a quiet form of access. The people who benefit aren’t the labs’ biggest customers, who were already fine. They’re the ones who looked at the numbers last year and decided the idea didn’t pencil out.

What to watch, without overreading it

I’d be careful about treating this week as proof of anything about anyone’s sincerity. Companies contain multiple positions at once. A safety team can genuinely want slower frontier progress while a commercial team ships an efficiency release that was in the pipeline for months. Both things are real, and neither cancels the other.

What I’d actually track is simpler and more useful: does the cost of doing real work with AI keep falling, and do the tools built on top of it get more capable of acting on your behalf rather than just talking to you? Two labs cutting prices within ninety minutes of each other suggests they’re competing hard on exactly that. Competition on price tends to be good for the person at the end of the chain.

The slowdown conversation isn’t over. But this particular release was about making AI cheaper, not smarter, and those are worth reacting to differently.

🕒 Published: September 23, 2026