Casey Newton, writing at Platformer:

Last week at his company’s developer conference, Meta CEO Mark Zuckerberg heavily promoted Muse, its own consumer-focused agent, which he positioned as a new peer to company pillars like Instagram and WhatsApp.

But after admittedly limited testing — I only got access a couple hours ago — so far I have been most impressed with Dots, the new agent from OpenAI. It’s highly capable, focused on the work I actually do, and intuitive to use.

Unlike Meta’s Muse, it isn’t free: to start, Dots will be available only to paid users of ChatGPT. (“You should of course expect us to do a mass-market [product] for billions of people someday,” OpenAI CEO Sam Altman told reporters Tuesday during a Q&A I attended.)

Here’s a good place to note that using an agent like this requires that you place a massive amount of trust in a platform. Giving an agent access to your text messages, emails, and banking information can go badly in all kinds of ways, and it’s completely reasonable to decide you would rather be a late adopter here, if you choose to adopt agents at all. Meta already operates at a significant trust deficit with its user base, and OpenAI is in the midst of a seemingly never-ending series of disclosures about various ways its agents misbehaved during testing. (The company announced yesterday that it had scrapped GPT-6.1 Astra because, among other things, it regularly deceived users about what it had and had not done.)

In September, I took the usefulness angle with these always-on artificial intelligence assistants; Newton took the privacy angle, which I think is equally important. Evidently, always-on agents are Silicon Valley’s next product category to conquer, but none of the major AI laboratories seem to have models aligned enough to handle sensitive data, nor do they appear trustworthy enough to store private information. In the days since my first article about Muse, the agent has been caught sending a user’s address to random hagglers on Facebook Marketplace without consent, among other suspicious and concerning events. All of this is to say that everything I said about Muse equally applies to Dots, which OpenAI seemingly spells without a capital D for quirkiness. Also from OpenAI:

We’re introducing GPT‑6.1 Sol, an upgrade to GPT‑6 Sol that nearly matches GPT‑6 Astra’s intelligence on agentic coding, computer use, and professional work at one-fifth of Astra’s standard input and output token prices. Cached input costs just $0.10 per million tokens—95% less than standard input pricing and 50% less than GPT‑6 Sol’s cached input pricing—giving developers more room to build and run capable agents that reuse context across requests.

While AI companies have been developing new products like Muse and Dots since the beginning of the fall, they’ve also been “pacing the frontier” and slowing high-end model releases. Earlier this year, the typical frontier model release cadence was a new model about every month — now, misaligned frontier models are being scrapped completely. GPT-6.1 Sol, Claude Opus 5.5, and Claude Sonnet 5.5 are prime examples of slowing AI development in the name of safety. These smaller yet capable models also address one of my persistent gripes with the AI industry: its seeming inability to turn a profit. Token prices for the best models have increased dramatically after the launches of Claude Fable 5 and GPT-6 Astra, and while their capabilities remain impressive, they’ve all but priced out major enterprise customers and hobbyists. These mid-tier models match frontier capabilities through distillation, but let labs lower inference costs and devote more time and compute to safety guardrails. I hope to see more of this, even after the next Astra and Fable models eventually debut in the coming months.

On the opposite side of the spectrum lies OpenAI’s new $500 Pro subscription, enabling ultra-fast inference hosted via Cerebras. Here’s Igor Bonifacic at Engadget:

Announced during the company’s DevDay event today, OpenAI’s new Pro 500 plan offers OpenAI’s highest usage allowance and comes with access to its new “Ultrafast” feature — it also costs $500 per month. At the end of 2024, OpenAI became the first AI lab to introduce a $200 Pro plan. Nearly every other provider, including Anthropic and Perplexity, later announced their own $200 tiers, so it’s possible other competitors will follow suit.

“Ultrafast is our premium speed tier for workloads where speed matters most,” the company explains. With the feature enabled, OpenAI’s Codex and Work apps will generate up to 300 tokens per second, making them capable of writing code and carrying out other tasks faster. The Ultrafast speed tier is also available to API users.

At the same time, OpenAI is also making its existing $200 Pro plan less appealing. In Codex and Work, $200 Pro subscribers will see their included usage decrease from 20x of what the company offers to Plus users, down to 10x of that same allowance. In ChatGPT, meanwhile, GPT-6 Pro message caps will decrease from 200 to 100 per week. According to the company, existing subscribers will keep their current limits for a time, and will later receive a one-time credit to help them make the most of their new reduced allowances.

The way I interpret this is that the low-end is effectively dead, the high-end will become more esoteric and niche over time, and the middle models and subscriptions will probably be the sweet spot for the vast majority of paid AI users. On the low-end, the core ChatGPT product has not received a model update in months now — it is end-of-life and illustrates that free ChatGPT users are unimportant to OpenAI. I don’t see that changing anytime soon. The high-end plans, meanwhile, are either getting nerfed or becoming so expensive that they price out almost all normal consumers, even those who use the most inference. Still, in October 2026, $200 Max-tier Claude subscribers cannot use more than half of their weekly allowance on Claude Fable 5.1. GPT-6 Astra usage for $200 ChatGPT Pro subscribers is now halved. These plans, once hailed as providing practically unlimited inference, are now objectively far less valuable. This trend will undoubtedly continue.

The mid-tier plans and mid-tier models will be how the industry makes money, if it does at all. The free plans, meanwhile, will seldom receive updates and attention as they no longer attract paying subscribers.