| This text initially appeared on Medium. Tim O’Brien has given us permission to repost right here on Radar. |
Whenever you’re working with AI instruments like Cursor or GitHub Copilot, the actual energy isn’t simply getting access to totally different fashions—it’s figuring out when to make use of them. Some jobs are OK with Auto. Others want a stronger mannequin. And typically it’s best to bail and change if you happen to proceed spending cash on a fancy downside with a lower-quality mannequin. If you happen to don’t, you’ll waste each money and time.
And that is the lacking dialogue in code era. There are just a few “camps” right here; the vast majority of individuals writing about this seem to view this as a fantastical and enjoyable “vibe coding” expertise, and some individuals on the market try to make use of this expertise to ship actual merchandise. If you’re in that final class, you’ve in all probability began to appreciate which you can spend a improbable amount of cash if you happen to don’t have a method for mannequin choice.
Let’s make it very particular—if you happen to join Cursor and drop $20/month on a subscription utilizing Auto and you’re proud of the output, there’s not a lot to fret about. However in case you are beginning to run brokers in parallel and are paying for token consumption atop a month-to-month subscription, this publish will make sense. In my very own expertise, a single developer working alone can simply spend $200–$300/day (or 4 instances that determine) if they’re attempting to deal with a mission and have opted for the most costly mannequin.
And, in case you are an organization and also you give your builders limitless entry to those instruments. Prepare for some surprises.
My Escalation Ladder for Fashions…
- Begin right here: Auto. Let Cursor path to a powerful mannequin with good capability. If output high quality degrades or the loop happens, escalate the problem. (Cursor explicitly says Auto selects amongst premium fashions and can change when output is degraded.)
- Medium-complexity duties: Sonnet 4/GPT‑5/Gemini. Use for targeted duties on a handful of information: strong unit exams, focused refactors, API remodels.
- Heavy raise: Sonnet 4 – 1 million. If I must do one thing that requires extra context, however I nonetheless don’t need to pay prime greenback, I’ve been beginning to transfer up fashions that don’t shortly max out on context.
- Ultraheavy raise: Opus 4/4.1. Use this out when the duty spans a number of tasks or requires lengthy context and cautious reasoning, then change again as soon as the large transfer is completed. (Anthropic positions Opus 4 as a deep‑reasoning, lengthy‑horizon mannequin for coding and agent workflows.)
Auto works superb, however there are occasions when you may sense that it’s chosen the mistaken mannequin, and if you happen to use these fashions sufficient, if you find yourself taking a look at Gemini Professional output by the verbosity or the ChatGPT fashions by the way in which they go about fixing an issue.
I’ll admit that my heavy and ultraheavy decisions listed below are biased in direction of the fashions I’ve had extra expertise with—your personal expertise may differ. Nonetheless, you also needs to have an identical escalation record. Begin with Auto and solely improve if you might want to; in any other case, you will be taught some classes about how a lot this prices.
Watch Out for “Considering” Mannequin Prices
Some fashions assist express “pondering” (longer reasoning). Helpful, however costlier. Cursor’s docs notice that enabling pondering on particular Sonnet variations can depend as two requests beneath crew request accounting, and within the particular person plans, the identical thought interprets to extra tokens burned. Briefly, pondering mode is superb—use it once you want it.
And when do you want it? My rule of thumb right here is that once I perceive what must be executed already, once I’m asking for a unit check to be polished or a way to be executed within the sample of one other… I normally don’t want a pondering mannequin. Then again, if I’m asking it to investigate an issue and suggest numerous choices for me to select from, or (one thing I do usually) once I’m asking it to problem my selections and play satan’s advocate, I’ll pay the premium for the very best mannequin.
Max Mode and When to Use It
If you happen to want large context home windows or prolonged reasoning (e.g., sweeping modifications throughout 20+ information), Max Mode may help—however it should eat extra utilization. Make Max Mode a short-term software, not your default. If you end up consistently requiring Max Mode to be turned on, there’s a very good probability you’re “overapplying” this expertise.
If it must eat one million tokens for hours on finish? That’s normally a touch that you simply want one other programmer. Extra on that later, however what I’ve seen too usually are managers who assume that is just like the “vibe coding” they’re witnessing. Spoiler alert: Vibe coding is that factor that folks do in shows as a result of it takes 5 minutes to make a foolish online game. It’s 100% not programming, and to make use of codegen, right here’s the key: It’s a must to perceive how one can program.
Max Mode and pondering fashions aren’t a shortcut, and neither are they a alternative for good programmers. If you happen to assume they’re, you will be paying prime greenback for code that may at some point must be rewritten by a very good programmer utilizing these identical instruments.
Most Vital Tip: Watch Your Invoice as It Occurs
A very powerful tip is to commonly monitor your utilization and utilization charges in Cursor, since they seem inside a minute or two of operating one thing. You may see utilization by the minute, the variety of tokens consumed, and in some instances, how a lot you’re being charged past your subscription. Make a behavior of checking a few instances a day, particularly throughout heavy periods, and ideally each half hour. This helps you catch runaway prices—like spending $100 an hour—earlier than they get out of hand, which is completely attainable if you happen to’re operating many parallel brokers or doing resource-intensive work. Paying consideration ensures you keep in charge of each your utilization and your invoice.
Preserve Monitor and Keep away from Loops
The opposite factor you might want to do is hold observe of what works and what doesn’t. Over time, you’ll discover it’s very simple to make errors, and the fashions themselves can typically fall into loops. You may give an instruction, and as a substitute of resolving it, the system retains operating the identical course of repeatedly. If you happen to’re not paying consideration, you may burn via plenty of tokens—and some huge cash—with out truly getting sound output. That’s why it’s important to look at your periods intently and be able to interrupt if one thing seems prefer it’s caught.
One other pitfall is pushing the fashions past their limits. There are duties they will’t deal with nicely, and when that occurs, it’s tempting to maintain rephrasing the request and asking once more, hoping for a greater outcome. In observe, that usually results in the identical cycle of failure, besides you’re footing the invoice for each try. Realizing the place the boundaries are and when to cease is vital.
A sensible method to keep on prime of that is to keep up a operating diary of what labored and what didn’t. Document prompts, outcomes, and notes about effectivity so you may be taught from expertise as a substitute of repeating costly errors. Mixed with maintaining a tally of your reside utilization metrics, this behavior will aid you refine your strategy and keep away from losing each money and time.