OpenAI, Meta and Apple price their new personal AI agents anywhere from zero to $500 a month

The best evidence that model quality no longer decides this contest is two price lists. Google's Gemini 4 Argon and OpenAI's GPT-6.1 Sol came out at the same $2 per million input tokens and $10 per million output, within 24 hours of each other ^[5]. Forkast reads this as near-frontier intelligence becoming a utility. On its account, once the intelligence is a commodity, the advantage moves to the interface and the ecosystem around it ^[6]. Matching prices are a decision about what to charge, and they say little about whether the two models are equally capable. Forkast does not compare what each model does with those tokens ^[5].

I think the consumer prices make more sense if you ask where the inference runs. OpenAI gives every Dots agent its own cloud computer ^[2]. That is a server bill that grows with each subscriber. The Pro 500 tier costs $500 a month, five times Meta's top tier of $100 ^[2][1]. Apple runs Siri AI on the device alongside its Private Cloud Compute ^[8], and only on hardware such as the iPhone 15 Pro and M1 iPads ^[4]. Part of that compute sits on a chip the customer has already paid for. Apple has announced no subscription fee ^[4].

Meta's free tier is harder to square with the developer price floor. Muse is free up to 100 million tokens a week ^[3]. At the $2 input rate Google and OpenAI charge developers, a full week of that allowance would list at $200. At the $10 output rate it would list at $1,000 ^[3]. Meta's first paid tier is $20 a month ^[3]. So either few free users come near the cap, or Meta serves tokens far more cheaply than those list prices ^[3].

The adoption figures measure reach. They do not yet measure habit, and habit is what the thesis depends on. Muse had 2.5 million US downloads by September 23 and topped the App Store, according to Forkast ^[12]. A download counts an install. Habit would show up as people still using the agent weeks later. OpenAI reports 1.2 billion weekly ChatGPT users, a figure Forkast notes is self-reported and has not been independently audited ^[10]. It counts ChatGPT users, not Dots subscribers ^[10].

The five-launch count also flatters the argument. Forkast lists five platforms launching between September 8 and October 2 ^[1]. One of them is Anthropic's Frontier Academy, a $100 million programme to train 10,000 engineers inside firms such as McKinsey, Accenture and Deloitte by the end of 2027 ^[7]. That works out to about $10,000 per engineer ^[2]. Gemini 4 Argon appears in the account mainly as a developer price ^[5]. The consumer split rests on three companies.

Each of those three lock-in strategies has its own limit. Forkast locates Apple's advantage in the personal context stored in Messages, Mail, Calendar and Photos ^[8]. Even so, the first rollout leaves out the EU and China, with Apple citing DMA compliance ^[9]. Meta put out Muse under two weeks after an $18 billion settlement with several states over children's harms from social media ^[13]. OpenAI's approach needs capital. SoftBank has committed $30 billion, backed by an $11.1 billion high-yield bond sale that Forkast calls the largest in corporate history ^[11].