The landscape of available models is changing super fast right now: as I write this Anthropic’s Fable is the undisputed best for sheer intelligence, but is also painfully expensive. OpenAI’s Sol is very close behind and much more reasonably priced, but still prohibitively expensive if you’re not on one of their subscription plans. At this point raw intelligence is no longer the only thing I care about though. Even relatively weak models (compared to the frontier) such as Zai’s GLM 5.2 are quite intelligent and more than enough for many coding tasks, at a way lower price.

OpenAI seems to be leaning into this, it seems they are no longer just optimizing for sheer intelligence: as of July 31 their Terra and Luna models (the smaller variants in the 5.6 series) have had a massive price drop. Literally more than 80% off the previous (already low) price. China isn’t sleeping either: Deepseek dropped their latest checkpoint of their v4 flash model within 24 hours of the Luna price drop, and it’s somehow even cheaper than the newly discounted Luna models.

Again, none of these cheaper models are going to compete with Fable (or Sol) on sheer intelligence, but I think that’s kind of the point. To put it in more human terms, if you were CEO of a company and you had some employees on staff who’s hourly rates were 3x the hourly rate of your more rank and file employees, you would probably naturally reserve the expensive employees for the tasks where they were “worth it.”

The same dynamic is playing out with the models today, and the question that needs answering is where “good enough” is. One complication is that there are ever more knobs to twiddle, so which model is worth using is dependent on both the model and the thinking/effort budget you give it.

This graph (copied from this tweet) is interesting:

What’s interesting about it is that they overlap! Terra on xhigh can beat sol on medium, and Luna on max can beat Terra on medium! Crazier still, Luna max is roughly equivalent to GPT 5.4 on xhigh, but at bargain basement pricing! GPT 5.4 was state of the art just a short while ago. Terra is also impressive, it’s roughly equivalent to GPT 5.5 level performance, but the discount is not quite as steep enough in Terra’s case so it’s in a bit of an awkward middle situation price-wise.

All in all, this is the first time that it actually feels worth it to use non-frontier models. I’ve set Luna max as my default now and am honestly quite happy with it.

Craziest of all is that this is just on the American models side… When you put the Chinese models into the equation they are basically beating the American models on cost effectiveness at every step. I haven’t had a chance to try out Qwen 3.8 max yet, but if this (ChatGPT generated, so take with a grain of salt) chart can be trusted it seems like it may be the sweet spot:

So what does this all mean? Well, in my case it means my workflow is changing again (previous workflow here) for the first time in a few AI years (read: months). Will write about that more in a later post, but the TLDR is that these prices have finally made swarm mode viable. Swarm mode is great for large projects and free form review and adversarial review loops, but it’s a tradeoff: burn way more tokens for a higher chance of getting working software without much human involvement. That was previously price prohibitive, but with the intelligence of this new crop of “good enough” models the price difference becomes negligible while the number of human thought-hours saved goes way up. I don’t mind “wasting” a million tokens if the tokens are at Deepseek’s prices!

One final thing: so when do you actually need the big guns like Sol and Fable now? For me at least the answer is: less and less often. Fable is the best brainstorming and design partner I’m aware of, and Sol is a champ at reasoning about trickier things like concurrency, but at this point letting them implement directly is starting to feel like a waste of money/tokens. Brainstorm and plan with the big guys, then let the smaller models do the grunt work.