In Today’s Issue:

🚢 The shipping week: Five releases, two structural breaks, and what was only announced

📈 The market squeeze: Why the middle of the market filled in, and why the top did not

💰 The price war: How competition is reaching the incumbents—in their own numbers

🖥️ The silicon race: Almost every frontier lab is now buying into silicon

🌍 Three theaters, three scarcities: American compute, Chinese chips, and European enforcement

Dear Readers,

On June 16, the Chinese lab Z.ai shipped GLM-5.2 under an open MIT license. On June 30, Anthropic released Claude Sonnet 5. On July 9, OpenAI moved its GPT-5.6 family into general availability, one day after xAI's Grok 4.5 went live, and on the very same day Meta shipped Muse Spark 1.1 and, for the first time in its history, asked businesses to pay for access to its models. Five providers put new models into general availability inside roughly three and a half weeks (Reuters; TechCrunch; Axios; Bloomberg, 07/09/2026). A market that read as a two-horse race between OpenAI and Anthropic just a quarter ago suddenly looks like a crowded field.

And yet the second fact of the week is just as stubborn as the first: none of the new arrivals took the capability crown. On the Artificial Analysis Intelligence Index, a neutral composite score that blends a model's performance across standard benchmarks, the newcomers sit at or below the incumbents' flagships, some of them well below (Artificial Analysis, 07/10/2026). The honest headline is cheaper, not better: capability near the frontier became available at a fraction of the price, from more suppliers, faster.

That distinction is where this DeepDive lives. If capable-enough intelligence is becoming cheap and interchangeable in the middle of the market, then pricing power on the model layer erodes even if nobody dethrones the best model. And money that cannot rest in the model layer has to go somewhere. The evidence points in two directions at once: downward, into compute and silicon, and outward, into the rules and infrastructure that govern deployment. So the question this essay tests: when the models themselves become interchangeable, does the contest simply move somewhere else, down to the chips that run them and out to the rules that govern them, and who is already winning on those layers?

All the best,

Kim Isenberg

Everything in this piece depends on one discipline: separating what shipped from what was merely announced. Inside a three-and-a-half-week window, June 16 to July 9, five providers put new models into general availability: GLM-5.2 from Z.ai on June 16, Claude Sonnet 5 from Anthropic on June 30, OpenAI's GPT-5.6 family (previewed June 26, generally available July 9), Grok 4.5 from xAI on July 8, and Muse Spark 1.1 from Meta on July 9 (Reuters; TechCrunch; Axios; Bloomberg, 07/09/2026). Two of those releases are structural events rather than routine version bumps, and they are the reason the duopoly framing no longer fits.

Meta's Muse Spark 1.1 is the first time Meta has charged businesses for access to its models (Bloomberg, 07/09/2026). Mark Zuckerberg described it as "the first time that we're doing a real serious API" and promised that "the pricing is going to be very aggressive and attractive" (Bloomberg, 07/09/2026). That is a clean break with the Llama era, when Meta gave its models away as open weights. Bloomberg frames it as a pivot toward closed models Meta can charge for, one that "required essentially rebuilding the technologies from scratch" after Alexandr Wang, formerly of Scale AI, came in to rebuild the lab now branded Meta Superintelligence Labs, backed by a record capital-spending plan of $125 billion to $145 billion for 2026 (Bloomberg, 07/09/2026; TechCrunch). The price is the real message: $1.25 per million input tokens and $4.25 per million output tokens, which Bloomberg reports is roughly a quarter of the cost advertised by other top models from OpenAI and Anthropic (TechCrunch, 07/09/2026; Axios, 07/09/2026). Tokens, for readers new to the jargon, are the small chunks of text a model reads and writes; per-token prices are the market's cleanest price signal. Zuckerberg's own gloss on the incumbents: "The pricing from some of the other labs is very extreme and has very high margins" (Bloomberg, 07/09/2026). Hold onto that quote, because the same man delivers the sharpest counterargument to this entire essay a few sections down.

Grok 4.5 is the second structural event: xAI's first genuine enterprise entry, wrapped in a corporate story that is itself evidence for where value is heading. xAI was absorbed by SpaceX, rebranded SpaceXAI, and went public on June 12, at a debut valuation reported anywhere between roughly $1.75 trillion and $2.2 trillion depending on the outlet and the trading day (CNBC, 06/23/2026; TechCrunch, 07/09/2026; Sherwood, 07/07/2026). The model lists at $2 per million input tokens and $6 per million output, with a 500,000-token context window (OpenRouter, 07/08/2026; Axios, 07/08/2026). Elon Musk positioned it as an "Opus-class" model that is faster, more token-efficient and cheaper. Axios, doing its own assessment, called it "a serious contender for business use that can be cheaper than rivals, even if it currently trails the performance of its competitors' best models" (Axios, 07/08/2026). It was not available in the EU at launch, and it was trained alongside Cursor, the coding tool xAI acquired (Axios, 07/08/2026).

Now the announced column, because that is where the loudest numbers live. Musk's next model is not shipped: a 2-trillion-parameter system he says will finish training in July and become available in August, a roughly 33 percent step up from the current 1.5-trillion-parameter V9 foundation model, with a further promise of entirely new foundation models each month through December 2026 (Crypto Briefing, 07/08/2026). That claim traces to a single Musk post relayed by one outlet, so treat it as a testable prediction, not a fact. Meta's next model, codenamed Watermelon, is in training with vastly more compute and no date or benchmark attached (Bloomberg, 07/09/2026; Axios, 07/09/2026). Huawei's forward chip roadmap is a roadmap, not delivered silicon. Everything structural in this essay rests on the shipped column alone.