Meta has finally shipped something it is willing to call a frontier model. Muse Wonder, released Monday alongside a bigger Muse Wonder Full, was trained at the gigawatt-scale Prometheus data center in Ohio and is priced at $4 per million input tokens and $20 per million output tokens through the Meta Model API. Meta’s model card puts it in the neighborhood of the top closed models and claims a lead on its own memory benchmark. What it does not do is publish specific scores, parameter counts or a license, and the full safety report is promised “later.” Wonder Full is limited to Muse Code subscribers and isn’t served through the API at all. Read it as a serious entry rather than a settled one: the pitch is agents with memory and focus, and the proof will come from people using them.
Across the same news cycle, one of the earlier consumer-AI bets went the other way. Character.AI has filed for Chapter 11, listing $50–100 million in assets and liabilities, with Disney agreeing to buy some assets for $20 million plus a non-exclusive license to the technology. That is a long way down from the $2.7 billion Google paid in 2024 to license its models and bring back its founders. The filing blames lawsuits, including one over a teenager’s death, a Texas child-safety investigation, and the plain economics of a product where heavy engagement raises your inference bill. There is no buyer yet for the rest, and the CEO says that without a bid, “the platform will be wound down.” The lesson is uncomfortable for everyone selling companionship at scale: the risk was never only technical.
The agent needs a leash
The week’s other thread is who controls software that acts on your behalf. OpenAI’s Codex CLI, since version 0.157.0, starts a persistent background daemon and self-updater whenever you open an interactive session. A user issue asks for it to be opt-in; the daemon outlives your terminal, can restart mid-task, keeps the environment of whichever session started it, and some commenters compared it to malware. The thread shows no reply from OpenAI, and the opt-out reportedly isn’t in the config docs. It is probably an overreach in convenience rather than anything sinister, but it lands in a month when trust in coding agents is already thin.
Microsoft’s Brad Smith, meanwhile, argues the UK should lead AI rules, citing its AI Security Institute’s early access to frontier models. The more interesting part is the ask: binding requirements, including mandatory breach notification for rogue agents and outside review before release, because voluntary pledges mostly bind the firms that were already behaving. Whatever you think of a vendor writing its own regulatory wish list, the specifics are concrete enough to argue with.
The through-line is accountability. Meta wants you to trust an agent with memory, Character.AI shows what happens when a product outruns its safeguards, and Codex shows how easily a tool decides it knows better than its user. Everything else today, Amex’s proof-of-intent tokens included, is an attempt to give the agent a leash.
