The Week AI Tokens Got Cheap
Four labs shipped frontier models in about 48 hours, and the price of a token fell off a cliff. A quick read on what happened and what it means if you build with these things.
I have been watching the model releases pile up this month, and one week stood out. Around the 8th and 9th of July, four labs shipped new frontier models inside about 48 hours of each other. xAI put out Grok 4.5, OpenAI shipped the GPT-5.6 family, Meta had Muse Spark 1.1, and Cognition shipped SWE-1.7. A week later Moonshot dropped Kimi K3, a 2.8 trillion parameter open model with a million token context, and soon after that Claude Opus 5 and a wave of Gemini Flash models landed too.
The models are not the real story. The price is.
The number that actually matters
Across those releases, the cost of output tokens fell to roughly a tenth of what the last generation charged. Not ten percent cheaper. About ten times cheaper. When four labs ship in the same week, nobody gets to hold a premium, so they all race to the floor.
If you build on top of these models, that is the whole game. The thing that used to wreck a budget, generating long outputs, summarizing everything, running an agent in a loop, just got an order of magnitude cheaper to run.
There is a catch
Cheaper tokens do not mean the benchmarks got honest. One detail stuck with me. GPT-5.6 Sol was quoted at 38.3 percent on the ARC-AGI-3 test when it ran with the lab's own API features, and 7.8 percent in the plain test setup. Same model, very different number, depending on who set up the run. So read a launch benchmark the way you read a used car listing.
The open models are the quiet winners. Kimi K3 landing near the top of the leaderboards as a fully open release means the gap between what you can rent and what you can run yourself keeps shrinking.
Where this leaves me
Two things. First, if you shelved an idea because the token bill did not make sense, it is worth pricing again, because the math probably changed this month. Second, picking a model is turning into a commodity decision. The moat is not the model anymore, it is what you build around it, your data, your evals, and the product. That part did not get cheaper.