There is a story we tell ourselves about the artificial intelligence race, and it goes in one direction only. Every few months a lab unveils a model that is bigger, smarter, more powerful than the last, and the headlines dutifully record another step towards some hazily imagined superintelligence. It is a story about the frontier, always pushing outward. So it is worth pausing on a release that runs the other way. This week Anthropic, the company behind the Claude chatbot and one of the two or three outfits genuinely at that frontier, launched a new model whose main selling point is not that it is cleverer. It is that it is cheaper.
The model is called Claude Opus 5, and Anthropic's pitch for it is almost refreshingly unglamorous. It is not the smartest thing the company has ever built. It is, in the company's own framing, the one you should reach for every day.
The thing to understand about tokens
To see why this matters, you have to understand how these systems are actually paid for, because it is not like buying software in a box. AI models are rented by the token, and a token is roughly a fragment of a word, a few characters of text going in or coming out. Every question you ask and every answer the model gives is measured in these fragments, and you are billed accordingly.
Opus 5 costs $5 for every million tokens of text you feed it and $25 for every million it produces. That sounds abstract until you set it beside the alternative. Anthropic's most capable model, called Fable 5, charges $10 and $50 for the same amounts, exactly double. For a business running these models millions of times a day, that gap is not a rounding error. It is the difference between a project that makes financial sense and one that does not.
Almost as good, for half the money
The genuinely interesting claim is what you give up for that saving, which is: not very much. Anthropic says Opus 5 comes close to matching Fable 5's abilities while costing half as much to run, and on the specific tasks most businesses actually care about, ordinary knowledge work and software coding, it says the cheaper model is not merely close but sometimes better.
Outside voices lend this some weight. The head of one legal-AI firm reported that Opus 5 produced work as good as the previous top model while using around a quarter fewer tokens to get there. The chief executive of a coding-tools company said it approached the flagship's performance at half the cost, and was especially good at the unglamorous work of debugging, finding out why a program has broken. When independent testers agree with a company's own marketing, it is usually worth believing.
There is a subtlety buried here that the pricing alone hides. A model that charges the same per token but rambles more, using more tokens to reach the same answer, can end up more expensive in practice. Efficiency is not just the sticker price, it is how economically the model thinks. Opus 5's real advance is that it appears to do more with less.
Deliberate hole in the middle
The most revealing part of the release is what Anthropic chose to leave out. Opus 5 is deliberately weaker than the company's top models at one particular skill: finding and exploiting security vulnerabilities in software, the raw material of cyberattacks.
This is not an oversight, and Anthropic says so plainly. It intentionally did not train the model on offensive cyber tasks. The company's genuinely powerful capabilities in this area live in a restricted model called Mythos 5, which is not sold to the public at all, but handed only to governments and critical-infrastructure firms through a vetted programme. Opus 5's safety filters, the automated classifiers that block dangerous requests, are tuned to intervene far less often than the flagship's, because there is simply less dangerous capability to guard against.
The design is unusually candid about the trade-off. Opus 5 is allowed to help hunt for weaknesses in openly readable source code, work that helps defenders, but is walled off from the darker arts of generating actual exploits. Strong at defence, deliberately hobbled at offence. In an industry often accused of shipping first and worrying later, it is a choice worth noticing.
Why now
None of this happens in a vacuum, and the timing tells you something about where the AI business has arrived. For a few heady years, companies threw money at these models to see what they could do, with only a vague sense of what they were getting back. That experimental indulgence is ending. Inference costs, the price of running the models at scale, have become a line item that reaches the boardroom, and buyers now want to know what they are paying for.
Anthropic is responding to that mood, and to competitive pressure from OpenAI's rival systems and cheaper models coming out of China. A near-flagship model at half the price is a direct answer to a customer base that has grown cost-conscious. It also reflects a recent bruising episode: Fable 5's safety filters were criticised for tripping too often on legitimate work, and the model was briefly pulled from sale entirely to comply with a government export-control order over its cyber capabilities, before access was restored a couple of weeks later. Opus 5, less powerful and less restricted, sidesteps much of that friction.
Real significance
Step back, and the release marks a small but real inflection in the story the industry tells about itself. The pure race for raw capability, bigger and smarter at any cost, is being joined by a less thrilling but more consequential race: to make this technology cheap and reliable enough that ordinary businesses can actually afford to run it all day.
Anthropic still has its frontier model for the hardest, most autonomous jobs, the days-long tasks where you want the very best regardless of price. But its guidance to users is telling. Choose the expensive one only when you truly need it; the rest of the time, reach for the everyday model. That is not the language of a moonshot. It is the language of a technology settling down into something people use for work, which may in the end be the more important milestone.
