Cursor, maker of the AI code editor most used by programmers worldwide, published a chart worth more than a hundred sales presentations. On the standardized coding test it created itself, CursorBench 3.1, its own model, Fable, delivers results well above all competitors: Composer 2.5, GPT-5.5 medium, Opus 4.8 high (Anthropic), Sonnet 4.6 high (Anthropic), Gemini 3.5 Flash (Google) and Kimi 2.6 (Moonshot). It costs more per task. And even so, it comes out cheaper in the end.
This text is not for programmers. It is for whoever signs software contracts, pays the AI bill at the end of the month, and needs to decide: is it worth paying more for the good model, or do you save on the entry-level one? The answer is in the chart. But only those who translate the chart into business language can see it.
What Fable is, and why it matters outside the world of programmers
Fable is an AI model specialized in writing code, created by Cursor (the company behind the code editor that many software teams use today to produce faster). Its direct competitors are well-known models: Claude (Opus and Sonnet) from Anthropic, GPT from OpenAI, Gemini from Google, Kimi from Moonshot, and Composer, from Cursor itself.
For anyone who hires software (internal or external), the AI model the vendor uses behind the scenes defines three concrete things: how long it takes to deliver, how much it costs per month of use, and how many bugs are left for the client to complain about later. This is not a technical detail. It is a budget decision.
CursorBench 3.1: the entrance exam for code models
Imagine an entrance exam just for AIs that write software. Same questions, same time, objective grading. Whoever gets more right scores higher. That is CursorBench 3.1, the latest version of Cursor's standardized test. The chart the company published compares how each model performs on this test, and how much each one charges per task solved.
The chart has two axes:
- Vertical axis (score): percentage of tasks answered correctly. The higher, the better.
- Horizontal axis (cost): dollars spent on average per task. The further to the right, the cheaper.
The Fable 5 high line stands alone up top, in the 70% correct range. The competitors all sit below 65%, several below 50%. For full reference: GPT-5.5 medium marks 58%, Opus 4.8 high (Anthropic's top model, known for being expensive) sits at 57%, Gemini 3.5 Flash at 49%, Kimi 2.6 at 47%, Sonnet 4.6 high at 47%. None come close to Fable.
"But Fable is the most expensive." Yes, and it is still the cheapest.
This is where most people do the math wrong. They look at the cost per call and decide by the cheapest. Whoever does this pays somewhere else.
Think of a renovation. You hire two painters. The first charges 100 reais a day and finishes the room correctly in three days. The second charges 50 reais a day, but paints crooked, misses corners, and you have to call them back to redo it. The second finishes in five days, cost 250 reais (50% more than the first), and delayed you a whole week on top of that. The "cheap" one turned out more expensive and slower.
With an AI model it is the same. A weak model makes more mistakes. When it makes a mistake, someone has to step in: review, send it back to be redone, fix it manually. Each error is one more round of calls (more AI cost), more programmer time stuck on rework (more salary burned), and more risk of a bug ending up in production (more cost at the end of the month, or worse, a client complaining in the leadership WhatsApp).
Fable's chart says exactly this: each task costs more, but you need fewer tasks to reach the result. And the result is cleaner. Total bill: lower. It is the first time in three years that "the most expensive per call" has become, by a wide margin, the cheapest per delivery.
What this changes when hiring (or being billed for) software
If you are hiring software developed with AI (and today practically all new software has AI involved at some point), the right questions change.
It is no longer "which AI is cheapest per call." It is:
- Which model does your team use, and why? Whoever saves on the entry-level model will charge you in time, rework, or bugs.
- How many attempts on average does your team need to close a coding task? The more attempts, the worse the model or the process is.
- Is the AI cost tied to delivery or to hours? If it is tied to hours, the vendor has no incentive to use a good model (they earn more the longer it takes).
These three questions separate the vendor who understands business from the vendor who is just passing the team's time on to you.
"If Fable is so good, will everyone use it for everything?" No, and that is fine.
A top model like Fable makes sense on a hard task: new logic, a large system reorganization, hunting a complex bug. For a simple task (generating a configuration file, writing an obvious test), a cheap model handles it without a problem. A mature team mixes models: it uses the expensive one where the expensive one is worth it, and the cheap one where the cheap one is enough.
It is the same logic as any operation. You do not send the partner to make photocopies, and you do not send the intern to negotiate a million-dollar contract. You distribute the task to whoever gives the best return per hour.
The subtler signal Fable's chart sends is another: the ceiling went up. On tasks where no AI model used to solve it properly (and the work went right back into the hands of the human programmer), Fable now solves it. This opens doors that were closed, and for you, the end client, it means more robust software, delivered faster, with fewer bugs in production. It is the kind of quiet advance that changes what your company can build over the next six months.
What to do with this information next week
Three practical moves for an owner, director, or operations leader:
- Ask your technology team (internal or contracted) which model is being used for each type of task. If the answer is "I use the cheapest for everything," you are saving the wrong way.
- Compare your AI bill against your real return. If the bill dropped 30% but the team needed 50% more time to deliver, you are losing money with the appearance of savings.
- Treat "which model" as a business decision, not a technical detail. It is like choosing a raw-materials supplier: the unit price is just the start of the calculation.
Whoever leads this conversation in companies over the coming months will not be the developer. It will be the manager who learned to read the right chart. Fable is not just a better model: it is a new yardstick for everyone who pays an AI bill. Whoever ignores the yardstick will keep thinking they are saving, while the competition delivers twice as much with half the rework.
(If you are still calibrating how your company should position itself around AI, it is worth reading why AI stopped being a promise and became real profit and why an AI agent soars in the demo and dies in production.)
