Four labs shipped in one day. GPT-5.6 Sol, Terra, and Luna went public this morning across ChatGPT, the API, and Codex, after the government lifted the freeze that had held them to about twenty approved partners. Grok 4.5 landed the same day, with Musk calling it Opus-class at a fraction of the cost. And Meta, quiet for months, released Muse Spark 1.1 and started charging for its models for the first time.

Yesterday I wrote that the frontier now turns over faster than habits can form. I did not expect the confirmation to arrive before the article was a day old. One release tracker is calling July 9 the most consequential single day in the history of model releases, and for once the superlative feels earned.

The token economy reset overnight. Fable 5, the model I have been working with these past ten days, runs ten dollars per million input tokens and fifty out on usage credits. Sol arrived this morning at five and thirty. Terra matches last quarter's flagship performance at two fifty and fifteen. Muse Spark undercuts them all at a dollar twenty-five. In one day, the cost of frontier intelligence dropped by about half.

This is the exponential people keep declaring dead. It never feels like a smooth curve from inside. It feels like nothing much for weeks, and then a Thursday where four labs ship at once and every cost assumption you held on Tuesday is obsolete.

The release I keep thinking about is not a model though, it is a mode. Sol's ultra mode embeds a multi-agent system inside the model itself. Hand it a hard task and it decomposes the work, spawns parallel subagents, and synthesizes their results, with no orchestration code on your side. On Terminal-Bench that lifts Sol from 88.8 to 91.9 percent, and a single ultra call can burn several times the tokens of a normal one.

I have spent months building exactly that by hand. Task decomposition, subagent routing, a manager process watching the fleet. Yesterday I wrote that the scaffolding is where the leverage is. Today a piece of the scaffolding moved into the model and became a setting. This keeps happening. Whatever the community builds around the models gets absorbed into the models, usually within a couple of quarters.

What does not get absorbed, at least so far, is the loop. Ultra mode spawns subagents, but it does not remember what worked for you last week, does not update its own skills, does not carry your evals or your taste. The orchestration moved into the model. The memory and the judgment still live in your files. That boundary between what the model owns and what you own keeps moving, and watching where it settles is now half my job.

There is a quieter signal inside Meta's release too. Muse Spark 1.1 is the first Meta model with a price tag, from the company that spent two years giving models away to erode everyone else's moat. When the free lab installs a meter, the market is telling you that tokens have become a real economy, with real margins worth defending.

For a practitioner the immediate work is re-benchmarking. Every routing decision I made last month, which model plans, which implements, which handles the cheap mechanical work, was priced against a table that no longer exists. The right response to a day like this is not excitement, it's a spreadsheet.

This is my second article today. I do not usually do that. But if the days are going to compound like this, the writing has to keep up. I doubt this is the last double.