Even the people paid to follow this industry gave up trying to keep pace this week. Business Insider counted a "whirlwind 72 hours" of rival announcements. The investor Jason Calacanis complained on X that "it's like a better new operating system, laptop and CPU being launched every 14 days." An Indian business channel compared the experience of following AI news to being trapped in a family WhatsApp group. When the noise is this loud, the useful thing is not to catalogue it. It is to ask what all the announcements have in common, because underneath the product names there were only two moves, and every lab made both.
The first move: stop selling a model, start selling a finished job. OpenAI's headline release on 9 July was the GPT-5.6 family, but the piece that changes the pitch is ChatGPT Work, an agent designed to take on a whole task rather than answer a question. It pulls from your files, your apps and your connected business systems, and it is meant to come back with a deliverable. Codex now lives inside the ChatGPT desktop app. OpenAI notes that more than a million non-coding workers already use Codex for things that are not code, which is the tell: the coding agent was the beachhead, and the office is the actual market. Meta made the same bet from the other end, launching Muse Spark 1.1 for agentic coding alongside the Meta Model API, its first proprietary API in the United States. Mark Zuckerberg broke a three-year silence on X to announce it, which tells you how much the company wants developers to notice.
The second move: cut the price, and say so out loud. OpenAI renamed its tiers, replacing "mini" and "nano" with Sol, Terra and Luna, and the point of the renaming is the pricing ladder underneath. Sol sits at $5 per million input tokens and $30 output, Terra at $2.50 and $15, Luna at $1 and $6. The company's own framing of Terra is revealing: competitive with the older GPT-5.5 "while being 2x cheaper." That is not a capability claim. That is a discount. Zuckerberg described Muse Spark's pricing as "very aggressive and attractive", and the numbers back him, at $1.25 in and $4.25 out. Meanwhile MiniMax, the Shanghai open-weight developer, raised $2 billion, with plans to follow it with $6.5 billion in zero-coupon convertible bonds, on the strength of a model whose main selling point is that it runs its two inference phases nine and fifteen times faster than its predecessor.
Put the two moves together and the strategy is legible. If you are selling an agent that grinds through a long task on its own, you are selling tokens by the bucket, not by the sip. Cost per token stops being a line item and becomes the whole business model. The labs are not cutting prices out of generosity; they are cutting them because the product they now want you to buy consumes tokens at a rate that would be unaffordable at last year's rates. OpenAI's new "ultra" setting, which coordinates four agents in parallel by default, is fast precisely because it burns more.
There are two things worth watching that got lost in the announcement blur. The first is a safety detail hiding in OpenAI's own paperwork: the company rates all three tiers, including the cheap ones, at its "High" risk level for both cyber and biological or chemical misuse. Terra and Luna are cheap, not harmless, and a High-risk capability at $1 per million tokens is a different proposition from the same capability at $5. The second is that Meta's coding model, on the benchmarks available, does not clearly beat what Anthropic and OpenAI already ship. It is competing on price and distribution, not on quality, which is what companies do when they are behind and know it.
The uncomfortable implication for the labs is that this is what commoditisation looks like from the inside. When every announcement is an agent, and every agent is cheaper than the last one, the differentiator quietly stops being the model. It becomes who owns the workflow the agent plugs into, which is why OpenAI is turning ChatGPT into something that looks increasingly like an operating system, and why Meta is racing to put Muse Spark inside WhatsApp, Instagram and Facebook. The models are converging. The distribution is not.