In the last two weeks, the model I would have told you to use changed at least three times.
Anthropic shipped Claude Sonnet 5, and Fable 5 came back online after a short pause. OpenAI began rolling out the GPT-5.6 family. xAI released Grok 4.5 with a coding agent to match. Google put out Gemini Omni Flash. Moonshot dropped Kimi K3. Zhipu released GLM 5.2, which is now, by most accounts, the strongest open-weight coding model available. That is not a year of releases. That is a fortnight.
If your plan for keeping up with AI is to identify the best model and standardize on it, that plan is already broken. Not because you picked wrong, but because the question is wrong. The wave that just rolled through was not an argument about which model is best. It was the moment the category stopped having one.
Nobody shipped a model. They shipped lineups.
Look at what actually arrived, and notice the shape of it.
OpenAI did not release a single thing called GPT-5.6. It released a family: Sol for high-end reasoning, coding, and science; Terra for roughly the previous generation's quality at about half the price; Luna beside them. Anthropic's newest generation spans Opus, Sonnet, and Haiku, each aimed at a different point on the cost and capability curve. Google ships Gemini in Omni, Flash, and Nano sizes. The unit of release has changed. A lab used to ship a flagship. Now it ships a portfolio, tiered by price and pointed at different jobs.
Once you see that, the argument about which one is best falls apart, because best is no longer a property of a model. It is a property of a model, a task, and a budget, together.
Three shifts hiding inside the noise
Strip away the launch-day benchmark charts and three things are genuinely moving.
The tiers are the point
Sol and Terra are not two tries at the same target. Sol is priced to be worth thirty dollars per million output tokens on work where reasoning earns its keep. Terra is built to deliver most of the quality for a fraction of the cost on work where it does not. The lab is telling you, right there in the price sheet, that it no longer expects you to run everything through the smartest model. It expects you to choose.
The floor is dropping fast
The story of the last year was capability at the top. The story of this wave is price at the bottom. Capable inference keeps getting cheaper, quarter over quarter, faster than most teams have updated their spreadsheets. Work that was too expensive to hand to a model in January is comfortably economical now. I wrote about how that rewrites the build-versus-buy math in Custom Software Is Cheaper Than You Think Now, and the trend has only picked up speed.
Open weights caught back up
GLM 5.2 and Kimi K3 are not curiosities. On coding especially, the best open-weight models are now close enough to the frontier that the assumption closed source is simply better no longer holds automatically. For anyone with a real reason to run models on their own hardware, a regulated industry, sensitive data, a tight unit cost, that option is live in a way it was not a year ago.
Why best model is the wrong question
Here is the trap. You run an evaluation, you pick a winner, you standardize the stack on it, you train the team, you write the integration. It takes a quarter. By the time you finish, two of the labs you passed on have shipped something better and cheaper, and the model you committed to is a tier behind.
Leaderboards now turn over on a timescale shorter than most procurement cycles. Chasing the top of the list is a treadmill, and the treadmill speeds up every month. The useful question is not which model is best. It is which model clears the bar for this specific task at the lowest cost, and the honest answer is that it will be a different model in ninety days.
I made a version of this case about the business risk in The AI Labs Are Going Public. Don't Bet Your Business on the Winner. The July wave is the technical proof of the same point. Betting your architecture on one model is the same mistake as betting your company on one lab.
The skill that replaces picking
If you cannot pick a permanent winner, you route.
Routing means matching each task to the cheapest model that clears its bar. The expensive reasoning model earns its price on the work where judgment and correctness matter most, the tangled refactor, the ambiguous requirement, the plan that has to be right the first time. The cheap, fast model handles the mechanical majority, the boilerplate, the formatting, the rough first draft. Most teams are still running everything through a single model, usually the most expensive one, and paying for reasoning they never use.
Routing also means building for swappability, and in practice that is two habits. First, keep a thin layer between your product and any specific model, so switching is a configuration change rather than a rewrite. Second, and more important, keep an evaluation set built from your actual work, so that when a new model lands on a Tuesday you can re-qualify it against your real tasks by Wednesday afternoon instead of trusting someone else's benchmark. The teams that move fastest over the next year will be the ones that can try a new model in a day because they already built the harness to judge it.
Where the advantage actually lives now
Follow the logic to its end. If the model is a tiered, steadily cheapening, swappable utility, then the model cannot be your advantage. Your competitor can call the same one through the same API before the week is out.
What does not commoditize is everything around it. The proprietary data only you have. The workflow you have shaped to your customers and your domain. The judgment to know which task deserves the expensive model and which does not. And the verification that tells you the output is actually correct, which I keep returning to because it is the one part that does not get easier as the models get cheaper. That was the argument in Work Back from the Future State: the tools got cheap, which is exactly why the things that are not the tools got more valuable.
The wave pushes value up the stack. It moves the durable advantage off the model, which everyone rents, and onto the things you build, own, and understand.
The models will keep coming
There will be another wave, probably inside the month. Some model you have not heard of will top some benchmark you have not seen, the group chat will light up, and someone will ask whether the team should switch.
If your architecture is built around one anointed model, that news is a fire drill every single time. If it is built to route work to the right tier and to swap models on a day's notice, the same news is just another Tuesday, and a slightly cheaper Tuesday than the last one. Stop trying to pick the model that wins. Build the thing that outlasts all of them.