Three AI plans become one, at a third of the price
Edition of 2026-10-13
# AI For All ### Tuesday, October 13, 2026
As of today, someone paying separately for Cursor Pro, SuperGrok and X Premium can fold all three into one plan at *$30/month, down from $90/month*.
One plan where there used to be three
X merged Grok, Cursor and X Premium into a single subscription. The paid entry point is *$8/month. The tier that matters is Premium, at $30/month, which the company describes as carrying the benefits of SuperGrok, Cursor Pro and X Premium+, previously $90/month* combined.
The mechanics are the interesting part. Every paid tier carries one shared pool of Grok usage, drawn on from the Grok apps, from the Grok tab inside X, and from Cursor alike. X says each tier includes more coding usage than the comparable Cursor plan. But a shared pool can be spent, so heavy use of one product now eats the budget of the others.
Nothing switches on its own. Existing subscriptions keep running, and the change is optional. For anyone paying for more than one of these products, the decision is arithmetic: three bills, or one.
Meta buys its way onto the front row
Meta released Muse Wonder and Muse Wonder Full, its first frontier models, trained at its Prometheus data centre. Muse Wonder is available today through the Meta Model API and Muse Code. The stronger model, Muse Wonder Full, is reserved for Muse Code subscribers and is not served through the API.
The features Meta leads with are about continuity, not raw power. Active Memory carries what matters about a user between conversations. Active Focus keeps the model working when a task gets hard, trying other approaches and asking for clarification. Both are aimed at agents that run for a long time on someone's behalf.
The detail to watch is what is missing. Meta says a fuller report with additional safety evaluations will follow. Shipping a frontier model before the safety report is a choice, and it is the one a buyer should price in.
Anthropic makes copying Claude much less productive
Anthropic published the defences it has built against distillation, the practice of training a smaller model on a stronger model's answers. Early evidence, it says, is that the data harvested by distillers has fallen by *90 to 99 percent*. The measures concentrate on the most valuable material: Claude's written reasoning.
The post names the labs it blames, including DeepSeek, Moonshot and Zhipu, and says Zhipu distilled the cyber capabilities of leading US models while building GLM-5.3. It repeats the earlier finding that Moonshot's Kimi models forwarded customer requests to Claude.
The cost sits on the same page. Anthropic says the changes have required some legitimate customers to update their integrations, and that more work remains. Making a model harder to copy can also make it harder to connect to, and that is the trade a buyer inherits.
Weak signal
An independent leaderboard now grades whether a model's code change would actually be merged by a maintainer of a real repository, rather than whether it passes unit tests. The best model leads a tight cluster of front-runners, and even they fail a large share of the tasks. The useful column is cost, where the same job is priced very differently from model to model. It is a measurement, not something to use today, which is why it does not rank.
Bundling is the quiet move of this quarter: fewer bills, one shared meter. Watch who is metered, and how fast the pool runs dry.