The next article initially appeared on Medium and is being republished right here with the creator’s permission.
Ask 10 builders which LLM they’d advocate and also you’ll get 10 totally different solutions—and virtually none of them are based mostly on goal comparability. What you’ll get as an alternative is a mirrored image of the fashions they occur to have entry to, those their employer accepted, and those that influencers they observe have been quietly paid to advertise.
We’re all dwelling inside recursively nested walled gardens, and most of us don’t notice it.
The entry downside
In company environments, the mannequin choice usually occurs accidentally. Somebody on the staff tries Claude Code one weekend, will get excited, tells the group on Slack, and instantly the entire group is utilizing it. No person evaluated options. No person ran a bakeoff. The choice was made by whoever had an organization card and a free Saturday.
That’s not a criticism—it’s simply how this stuff go. Nevertheless it implies that when that very same particular person tells you their favourite mannequin, they’re actually telling you which ones mannequin they’ve had essentially the most reps with. There’s a real studying operate at play: You get sooner, your prompts get higher, and the mannequin begins to really feel virtually intuitive. It’s not that the mannequin is objectively superior. It’s that you just’ve gotten good at utilizing it.
This issues greater than individuals admit, as a result of plenty of this area runs on emotions slightly than proof. Individuals really feel good about Opus proper now. It feels highly effective; it feels good; it feels such as you’re utilizing the perfect instrument accessible. And perhaps you might be. However ask somebody who’s paying for their very own tokens whether or not they really feel the identical means, and also you are inclined to get a extra calibrated reply. Pores and skin within the sport has a means of sharpening opinions.
The affect downside
There’s additionally some huge cash shifting by this area in ways in which don’t at all times get disclosed. Mannequin suppliers are spending actual price range to verify the proper individuals have the proper experiences—early entry, credit, invites to the proper occasions. Anthropic does it. OpenAI does it. This isn’t a scandal; it’s simply advertising, but it surely muddies the sign significantly. When somebody you observe is effusive a few mannequin, it’s value asking whether or not they arrived at that opinion by sustained use or by a curated demo atmosphere.
In the meantime, some builders—particularly these constructing within the open—will use no matter doesn’t value an arm and a leg. Their enthusiasm for a mannequin is perhaps extra about its pricing tier than its functionality ceiling. That’s additionally a sound sign, but it surely’s not the identical sign.
The alignment downside (the opposite one)
Then there are the geopolitical concerns. Some builders are intentionally avoiding Qwen and GLM as a result of considerations in regards to the international locations they originate from. Others are utilizing them as a result of they’re compelling, succesful fashions that occur to be dramatically cheaper. Each camps assume the opposite is being naive. It is a actual dialog that doesn’t have a clear reply, but it surely’s occurring largely below the floor.
What I’ve really been doing
I’ve been forcing myself to check outdoors my consolation zone. I’ve spent the final week utilizing Codex critically—not casually—and my expertise to date is that it’s practically indistinguishable from Claude Sonnet 4.6 for many coding duties, and it’s operating at roughly half the associated fee if you consider how effectively it makes use of tokens. That’s not a small distinction. I need to dwell with it longer earlier than I’ve a agency opinion, however “every week” is the minimal threshold I’d set for any mannequin analysis. Something much less and also you’re simply score your first impression.
I’ve additionally began utilizing Qwen and GLM-5 critically. Early outcomes are fascinating. I’ve had some compelling successes and some jarring errors. I’ll reserve judgment.
What I’ve observed with my very own Anthropic utilization is one thing value naming: I default to Haiku for well-scoped, mechanical duties. Sonnet handles virtually every part else with room to spare. Opus solely comes out once I want real breadth—structure questions, strategic framing, something with a genuinely extensive scope. However I’ve watched individuals in company environments depart the dial on Opus completely as a result of they’re not paying for tokens themselves. And right here’s the factor—that’s really not at all times to their benefit. Excessive-powered fashions overthink easy duties. They’ll add abstractions you didn’t ask for, restructure issues that didn’t want restructuring. When I’ve a clearly templated class to jot down, Haiku will get it proper at a tenth of the associated fee, and it doesn’t second-guess the design.
The factor we ought to be speaking about
Everybody final month was exercised about what Sam Altman stated about power consumption. High-quality. However I feel the extra urgent query is about advertising budgets and the way they’re distorting the collective understanding of those instruments. The benchmarks are beginning to really feel managed. The influencer protection is clearly formed. The entry packages create a constructive bias amongst individuals with the biggest audiences.
None of this implies the fashions are dangerous. A few of them are genuinely outstanding. However if you ask somebody which mannequin to make use of, you’re getting a solution that’s filtered by their employer’s procurement choices, the influencers they observe, what they’ll afford, and the way lengthy they’ve been utilizing that exact instrument. The reply you get tells you numerous about their scenario. It tells you virtually nothing in regards to the mannequin.
Take all of it with applicable skepticism—together with this publish.
