My focus at work is positioning my engineering org to best take advantage of AI. evaluating tools and strategies over the last year, here's the heuristic that I keep coming back to:
models > money > tools
A better model is worth several times it's increase in price. You can probably throw a number of sonnets at a task and eventually get an opus-like output, but maybe not. And it'll certainly be cheaper to use the better model.
and similarly, to get LLMs to write you better code, you could write the most elite development harness, recursive prompt templating framework, tiered memory hierarchy, bolt on a formal grammar system, the works. but you'd probably be better off just throwing all that shit into an LLM-as-judge or two, especially controlling for ongoing maintenance and upgrades of your pipeline. Quality is absorbed into intelligence, which is now measured in money. Just Use More AI!
Now for the end of the party:
Mythos is the first class of models to be gated to select companies before broad release. If you're competing against a company with early Mythos access, you may be bringing Opus to a Mythos fight. Not fun. Someone else has a better model than you, so you have to compensate by spending more money.
Next problem - the time of unlimited token spend without measurement is also coming to an end. many companies (Uber being in the news this week) are ditching their token spend leaderboards. Between ROI skecticism and lab price hikes, you may not be able to spend your way ahead anymore either.
Which leaves us to the end of the food chain: the tools. You still have complete control here. Expertise does still matter! We shrimp at the "application layer" are starting to dry out from a year long bender of token abundance. To borrow a favorite phrase of my CEO, what got us here won't get us there.
For a while I have often been a wet blanket on teammates designing highly complicated toolchains. "We should just use a single LLM call for this" is really lame to hear. Buying is less fun than building. But my tone is beginning to change at the edges - we can do it for Opus, but we won't be able to throw a Mythos call at every push on every PR. It's time to start shifting our strategy to prepare for the token market of 2027.