Anthropic released Claude Opus 5 on July 24 and framed it as the everyday default for office work, priced at roughly half of its flagship Fable 5. On several benchmarks it lands within reach of Fable 5 while costing about 50 percent less per token, and it ships with a control that lets you pick how hard the model thinks on each request.
Key Takeaways
- Claude Opus 5 reaches near-Fable 5 quality on many tasks at about half the token price.
- A low, medium and high effort control lets users trade speed for depth on each prompt.
- Opus 5 posts new top scores on coding and knowledge-work evaluations, and Anthropic positions it as the model for daily office work.
Have an AI Sum Up This Article
ChatGPTOpus 5 Lands at Half of Fable 5’s Token Price
The model went live on July 24, and the pitch is unusually blunt for a frontier lab. Anthropic expects Opus 5 to serve as the default option for many day-to-day office needs, not as a rare heavy-lifting model you reach for once a week. It succeeds the version the company shipped when Opus 4.8 rewrote the rules on dynamic workflows, and it delivers clearly better output for the same running cost as that predecessor.
The price line is the headline. Opus 5 costs about 50 percent less per token than Fable 5, the company’s most advanced widely available model, while approaching it in many categories. Anthropic detailed the numbers in its official Opus 5 announcement published Friday, and the split it draws between capability and cost is the whole point of the release.
On the scoreboard, Opus 5 topped Fable 5 on eight of thirteen benchmark tests despite the lower price. It set new state-of-the-art marks on coding and knowledge-work suites like Frontier-Bench and GDPval-AA, the kind of evaluations that track real professional output rather than trivia. It still trails the specialized Mythos 5 on cybersecurity tasks, a gap Anthropic did not hide.
One number stands out for anyone tracking raw reasoning. On ARC-AGI-3, a benchmark built around solving problems the model has never seen, Opus 5 scored 30.2 percent, close to four times the mark posted by OpenAI’s GPT-5.6 Sol. That is the sort of gap that reframes a pricing story into a capability story, and it lands right as the token economics of the whole market are shifting, a move Anthropic already signaled when Fable 5 switched from flat subscriptions to metered API tokens.
A Low, Medium, High Dial on Every Answer
The feature that will change daily habits is the effort control. Claude Opus 5 lets the user set how much effort the model spends on a task, choosing between low, medium and high before it answers. Low returns a fast, cheap reply for routine work, high spends more compute and time on a hard problem, and the choice sits in the hands of the person sending the prompt.
For a product team, that dial is a budget lever. A support draft, a meeting summary or a quick rewrite can run on low and barely register on the bill, while a thorny migration plan or a legal review can run on high without switching models. Teams stop paying frontier rates for clerical work, which is exactly the layer where Anthropic says most usage now sits, a pattern it laid out when it admitted Claude runs more office work than code.
For developers, the migration is light. Opus 5 slots in as a drop-in successor to Opus 4.8 at the same cost, so the practical question is not whether to move but how aggressively to route everyday calls to low effort. The efficiency and safety work Anthropic folded into the model means fewer retries and fewer surprises, which matters more than a benchmark point when the model runs inside an agent loop all day.
The nuance is that the dial only pays off if teams actually use it. A shop that leaves every call on high inherits the capability but not the savings. The winners here are the teams that treat effort as a per-task decision, and the tooling around Opus 5 is clearly built to nudge them that way.
More articles on Horizon
- Gemini Closes In on One Billion Monthly Users
- Every AI Model Tested Cheated UK Safety Tests
- Anthropic Pays $1.5 Billion for Pirated Books
Half-Price Frontier Puts OpenAI on the Clock
A near-frontier model at half price does not stay a Claude story for long. It hands developers a cheaper default and forces every rival to answer on cost, not just on capability. OpenAI feels it most directly, since the ARC-AGI-3 gap lands on GPT-5.6 Sol, and the pressure was already building when Grok 4.5 undercut Claude Opus at a third of the price.
The response is predictable. Rivals either match the price and eat the margin, or they lean harder on a differentiator Opus 5 does not touch, which for now is the cybersecurity edge Mythos 5 still holds. Either way, the implicit market standard for a daily office model just got cheaper, and everyone else now has to explain why theirs costs more.
The timing is not an accident. Anthropic is preparing an IPO later this year, and a model that shows margin discipline while topping benchmarks is a clean story to take to the market, one it started building when it filed its IPO paperwork with the SEC. Opus 5 is a product launch, but it is also a proof point aimed at investors who want to see a frontier lab make money.
For the rest of 2026, the number that matters is not a benchmark but a habit. If teams really route their routine work to low effort, Opus 5 quietly resets what a professional AI model is supposed to cost.
Follow the story on Horizon.
