Eight days ago I published a post arguing that the age of the single flagship model was over, and that the skill worth building was routing: matching each task to the cheapest model that clears its bar. On Thursday, Anthropic answered the argument directly. It shipped Claude Opus 5, agreed with the premise, and moved the dial inside the model.
That is the whole story in one line, but it is worth slowing down, because the move is cleverer than a benchmark bump and it changes what routing means in practice.
What actually shipped
Opus 5 is Anthropic's fourth model in under two months. It costs five dollars per million input tokens and twenty-five per million output, exactly what Opus 4.8 cost. The pitch is that it comes close to the intelligence of Fable 5, the company's frontier public model, at half of Fable 5's price.
Notice the thing Anthropic is careful not to claim. It does not say Opus 5 is its smartest model. Fable 5 still holds that spot, and a restricted model called Mythos 5 sits above both. Opus 5 is pitched as something more useful than smartest: the one you reach for by default, every day, without thinking about the bill. It is the new default on Claude Max and the strongest model on Claude Pro, and it is live on the API, Claude.ai, Claude Code, and Cowork from day one.
Here is the Anthropic ladder as it stands after Thursday.
| Model | Role | Cost /1M (in / out) |
|---|---|---|
| Claude Mythos 5 | Restricted, top-end research and hardest tasks | not publicly listed |
| Claude Fable 5 | Smartest public model, longest-horizon autonomy | $10 / $50 |
| Claude Opus 5 | The everyday default, near-frontier at half the price | $5 / $25 |
| Claude Opus 4.8 | Prior default, now superseded at the same price | $5 / $25 |
On top of the ladder sits a dial. Opus 5 lets you set how much effort it spends on a task, three settings in the apps, from low through high, with finer control on the API. There is also a Fast mode that runs about two and a half times quicker for double the token price. The benchmarks are strong where you would want them to be for daily work: state of the art on Frontier-Bench, more than double Opus 4.8 at a lower cost per task; within half a percent of Fable 5 on CursorBench at half the cost; roughly three times the next-best model on ARC-AGI 3. It is also, by Anthropic's own audit, the most aligned Opus they have shipped, which matters more than it sounds when the model is running unattended in an agent loop.
The dial is the story
Last week, routing meant choosing between models. The expensive reasoning model earned its price on the tangled refactor and the plan that had to be right the first time. The cheap, fast model handled the mechanical majority. You matched the task to the tier and paid accordingly.
Opus 5 takes that same idea and folds it into one model. Hold the model constant and turn the effort up for the hard part, down for the rote part. The clearest proof is buried in the CursorBench number: at maximum effort, Opus 5 lands within half a percent of Fable 5 at half the cost per task. That is the exact trade routing between models used to buy you, now available as a setting on a single model you never have to swap out.
This does not retire routing. It relocates the dial. The premise of the earlier piece was that best stopped being a property of a model and became a property of a model, a task, and a budget, together. Opus 5 is that idea made literal. The budget knob is now printed on the model itself, and you turn it per task instead of switching APIs.
I wrote about this in The Flagship Is Dead. Start Routing., and Opus 5 is the strongest confirmation of the thesis I could have asked for. The lab that could most credibly sell you one anointed model instead sold you a dial.
Same price, more model
Do not skip past the pricing line. Opus 5 costs what Opus 4.8 cost, and more than doubles it on the coding benchmark that matters most, with double-digit gains on organic chemistry and protein prediction thrown in. Same dollars, materially more model.
That is the floor dropping in real time. The story of the last year was capability at the top. The story of this year is that the price of a given level of capability keeps falling faster than most teams update their assumptions. Work that was too expensive to hand a model in the winter is comfortably economical now. I made the business version of this case in Custom Software Is Cheaper Than You Think Now, and every model in this wave has pushed the same direction, harder.
What it means if you actually ship
I do not read these launches as a spectator. I run agentic work every day, the command center that watches my products, eRateIQ, a family app, Anchor. Agentic work is exactly where token bills balloon, because the model does not answer once and stop. It loops, retries, verifies, and reconsiders, and every pass is billed.
That is the setting where a dial earns its keep. A model that lands near the frontier at half the frontier price, with a knob to spend less thinking on the boilerplate and more on the ambiguous middle, is a cost-discipline tool, not just a leaderboard trophy. The tell is a stat Anthropic slipped into the announcement: on a trading benchmark, Opus 5 is the strongest Opus yet while using roughly a seventh of the reasoning tokens Opus 4.8 burned. It is being tuned to think as much as the task needs and no more. For anyone paying the bill on an agent that runs unattended, that is the number to circle.
What the dial still cannot do
Here is the part the dial does not touch. It changes how hard the model thinks. It does not verify that the output is correct, it does not know your proprietary data, and it does not hold the judgment about which task deserves which setting. That judgment is still yours.
Last week I argued that the durable advantage had moved off the model, which everyone rents through the same API, and onto the things you build, own, and understand: your data, your workflow, the verification that tells you the answer is actually right. Opus 5 does not weaken that argument. It sharpens it. If the model is now a dial you turn, the value lives even more plainly in the hand on the dial. That was the point of Work Back from the Future State, and it survives every new model unchanged.
The next one is already coming
There will be another model soon, probably inside the month, and it will top some benchmark and light up the group chat. Opus 5 is the strongest evidence yet that the right response is not to crown it. It is to keep building the harness that lets you turn the dial, between models when you must and within one when you can, without rewriting anything.
The router used to be something you built yourself. Now part of it ships in the box. Build the part that still does not.