OpenAI and Anthropic each released a new flagship AI model in the first week of September, two days apart, at exactly the same price. If you run a business in New Zealand and you're wondering whether that matters to you, here's the short version. Same sticker price, different bill, and the right choice depends on the job, not the headlines.
What just happened?
Two things. On 1 September Anthropic released Claude Fable 5.1. On 3 September OpenAI released GPT-6 Astra. Both are the top model from each company, both are built for "agentic" work, and both list at the same price: US$10 per million words in, US$50 per million words out. (Technically it's per token, and a token is about three quarters of a word. That's all you need to know about tokens.)
"Agentic" is worth a sentence too. It means the AI doesn't just answer a question. It carries out a task with several steps, like reading a customer's email, checking the order in your system, drafting a reply and updating the record. That's the kind of AI most businesses actually want, and it's what both of these models were built for.
Which one is better?
Depends who you ask, and that's not a dodge.
OpenAI published a comparison table and Astra wins nearly every row of it. An independent testing outfit called Artificial Analysis ran its own index and scored Fable 5.1 higher on both of its headline measures. DataCamp's write-up, which is the most thorough we've read, lands on the sensible position: a vendor grading its own competitor deserves less weight than an independent test, but neither settles it.
What the numbers do agree on is where each model is strong. Astra is better at operating real software, clicking through apps the way a person does, and at maths and science-style problems. Fable 5.1 is better at long, hard reasoning, the kind where the model has to hold a lot in its head and think carefully. For a business, that translates roughly into: Astra for agents that drive your existing tools, Fable 5.1 for agents that have to work something out.
Same price, so why does one cost twice as much?
This is the part most people miss, and it's the part that matters for your budget.
The price is per token. But the two models don't use the same number of tokens to finish the same job. Fable 5.1 writes a lot more while it's thinking. So when Artificial Analysis measured what a task actually cost end to end, Fable 5.1 came out at about US$3.76 per task and Astra at about US$1.67. Same price list, more than double the bill.
Then it flips again. Anthropic made Fable 5.1 four times cheaper at reading saved context, the stuff an AI agent has to reread every time it takes a step: your instructions, your product list, the document it's working through. Systems built around that kind of rereading come out cheaper on Fable 5.1. And Astra charges extra once a single request gets very large, which Anthropic doesn't.
So the honest summary is: how your AI system is built decides the bill more than which model you pick. A sloppy build on the "cheaper" model will cost more than a careful build on the "dearer" one.
What should a NZ business actually do with this?
First, don't pick a model by reading leaderboards. Describe the job. If your agent has to click around your inventory system and produce a tidy spreadsheet at the end, Astra's the stronger fit today. If it has to read a 200-page contract and work out what's missing, Fable 5.1 is. Most real jobs sit somewhere in between, and that's what a scoping conversation is for.
Second, ask whether you need this tier at all. These are the biggest and most expensive models from each company. A lot of everyday business tasks run fine on the smaller, faster models one rung down, at a fraction of the cost. Paying flagship prices for a job a mid-tier model can do is the most common way we see AI budgets get wasted.
Third, and most important: don't marry either of them. These two models are two days apart and the next pair will be along within months. The AI systems Putti builds keep the model behind a clean boundary, so when the prices or the strengths change, you swap the engine and keep the car. The businesses that get burned are the ones whose whole system was wired to one vendor's quirks.
Where to start
If you've got a task in mind and you're not sure which model, or which tier, it needs, talk to us. We'll tell you plainly, including when the answer is a cheaper model than the one in the headlines. And if you're still working out what an agent is, start with what an AI agent actually is and when your business needs one, or see what AI implementation costs in New Zealand.
Figures in this post come from DataCamp's comparison (updated 5 September 2026), OpenAI's published benchmark tables and Artificial Analysis's independent index. They'll move. That's rather the point.