| This text initially appeared on Medium. Tim O’Brien has given us permission to repost right here on Radar. |
While you’re working with AI instruments like Cursor or GitHub Copilot, the actual energy isn’t simply gaining access to completely different fashions—it’s realizing when to make use of them. Some jobs are OK with Auto. Others want a stronger mannequin. And typically you must bail and swap for those who proceed spending cash on a fancy drawback with a lower-quality mannequin. When you don’t, you’ll waste each money and time.
And that is the lacking dialogue in code technology. There are just a few “camps” right here; nearly all of folks writing about this seem to view this as a fantastical and enjoyable “vibe coding” expertise, and some folks on the market try to make use of this expertise to ship actual merchandise. If you’re in that final class, you’ve most likely began to understand that you may spend a incredible amount of cash for those who don’t have a technique for mannequin choice.
Let’s make it very particular—for those who join Cursor and drop $20/month on a subscription utilizing Auto and you’re pleased with the output, there’s not a lot to fret about. However if you’re beginning to run brokers in parallel and are paying for token consumption atop a month-to-month subscription, this put up will make sense. In my very own expertise, a single developer working alone can simply spend $200–$300/day (or 4 occasions that determine) if they’re making an attempt to sort out a undertaking and have opted for the costliest mannequin.
And—if you’re 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 swap when output is degraded.)
- Medium-complexity duties: Sonnet 4/GPT‑5/Gemini. Use for centered duties on a handful of recordsdata: sturdy unit assessments, focused refactors, API remodels.
- Heavy elevate: Sonnet 4 – 1 million. If I have to do one thing that requires extra context, however I nonetheless don’t wish to pay high greenback, I’ve been beginning to transfer up fashions that don’t shortly max out on context.
- Ultraheavy elevate: Opus 4/4.1. Use this when the duty spans a number of initiatives or requires lengthy context and cautious reasoning, then swap again as soon as the massive transfer is completed. (Anthropic positions Opus 4 as a deep‑reasoning, lengthy‑horizon mannequin for coding and agent workflows.)
Auto works nice, however there are occasions when you’ll be able to sense that it’s chosen the improper mannequin, and for those who use these fashions sufficient, you already know when 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 selections listed here are biased in direction of the fashions I’ve had extra expertise with—your individual expertise would possibly range. Nonetheless, you also needs to have an identical escalation checklist. Begin with Auto and solely improve if it’s worthwhile to; in any other case, you will study 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 observe that enabling pondering on particular Sonnet variations can rely as two requests beneath group request accounting, and within the particular person plans, the identical concept interprets to extra tokens burned. In brief, pondering mode is great—use it while you want it.
And when do you want it? My rule of thumb right here is that once I perceive what must be performed 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. Alternatively, if I’m asking it to investigate an issue and suggest numerous choices for me to select from, or (one thing I do typically) 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
When you want large context home windows or prolonged reasoning (e.g., sweeping adjustments throughout 20+ recordsdata), Max Mode will help—however it is going to devour extra utilization. Make Max Mode a non permanent device, not your default. If you end up always requiring Max Mode to be turned on, there’s a superb probability you’re “overapplying” this expertise.
If it must devour one million tokens for hours on finish? That’s normally a touch that you just want one other programmer. Extra on that later, however what I’ve seen too typically are managers who suppose that is just like the “vibe coding” they’re witnessing. Spoiler alert: Vibe coding is that factor that individuals do in displays 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 important to perceive the best way to program.
Max Mode and pondering fashions usually are not a shortcut, and neither are they a substitute for good programmers. When you suppose they’re, you will be paying high greenback for code that can at some point must be rewritten by a superb programmer utilizing these identical instruments.
Most Vital Tip: Watch Your Invoice as It Occurs
A very powerful tip is to recurrently monitor your utilization and utilization charges in Cursor, since they seem inside a minute or two of working one thing. You’ll be able to see utilization by the minute, the variety of tokens consumed, and in some circumstances, how a lot you’re being charged past your subscription. Make a behavior of checking a few occasions 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 fully doable for those who’re working 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 it’s worthwhile to do is preserve monitor 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 would possibly give an instruction, and as a substitute of resolving it, the system retains working the identical course of many times. When you’re not paying consideration, you’ll be able to burn by means of a whole lot of tokens—and some huge cash—with out really getting sound output. That’s why it’s important to look at your periods intently and be able to interrupt if one thing appears to be like prefer it’s caught.
One other pitfall is pushing the fashions past their limits. There are duties they’ll’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 always results in the identical cycle of failure, besides you’re footing the invoice for each try. Understanding the place the boundaries are and when to cease is vital.
A sensible method to keep on high of that is to keep up a working diary of what labored and what didn’t. Document prompts, outcomes, and notes about effectivity so you’ll be able to study from expertise as a substitute of repeating costly errors. Mixed with keeping track of your dwell utilization metrics, this behavior will aid you refine your method and keep away from losing each money and time.
