AI pricing split this week: frontier labs shipped models that are no longer always smarter-and-cheaper at once, and the market is now separating routine work into low-cost routing while reserving top-end spend for hard, long-horizon tasks. Grok 4.5, Meta’s Spark 1.1, and OpenAI’s 5.6/Luna are positioned around price, token efficiency, and agentic routing, not just raw intelligence. The bullish case is that cheaper models unlock far more usage and app creativity; the bearish counter is that enterprises are tightening budgets and the cheap-model race is getting subsidized by giant capex and proprietary silicon. The video closes on a market structure thesis: intelligence is no longer one commodity, and the winners will be the labs and routing layers that capture both price-sensitive mass use and premium frontier demand.
Within about 48 hours, Grok 4.5, OpenAI 5.6 Global, and Meta’s frontier model all shipped, marking a break from the last ~5 years when every model launch got smarter and cheaper at the same time.
That smarter-and-cheaper dynamic is said to have died this year with Anthropic Stable 5 and new frontier launches like GPT 5.6, forcing buyers to separate intelligence from price.
Fable 5 is described at about $20–$30 input and $80 output per million tokens, making companies spending tens to hundreds of millions question whether they need the smartest model for every task.
Grok 4.5 is pitched as a cheaper daily driver: $2 in and $6 out per million tokens, about a third of the price of the reference model while generating tokens faster per dollar.
Grok 4.5 is built on xAI’s version 9 foundation model at about 1.5 trillion parameters versus the all-year reference version 8 small at about 500 billion, roughly a 3x parameter jump.
Compared with Claude Opus 4.8, Grok 4.5 is said to use four times fewer tokens and up to 17x less price/cost for the same task, especially on agentic parallel work.
Cursor’s acquisition for $60 billion is used as the routing example: it supplies usage data, not code content, to help train Grok 4.5 to route requests more intelligently and pre-calculate model spend.
The intended Grok 4.5 use case is building and long tasks: a 90%–95% intelligence model at one-tenth the price is framed as enough for 80%–90% of work, with the frontier model reserved to orchestrate planning and design.
Elon says Grok 4.5 is not the smartest model but is a strong workhorse; another wave of models is already said to be coming, with Grok 5 or Grok 5.5 expected under a month away at 5–10 trillion parameters and a Mythos-class model possible in a couple months.
Meta’s first paid token model is presented as Muse Spark 1.1: $125 in and $425 out per million tokens, about 25% of the Frontier model’s price, and cheaper than GLM 5.2.
Muse Spark 1.1 is framed as an agentic, coding, omni model that takes video, images, and text, uses a master agent to plan and delegate, and can take over a computer/account for requested actions.
Meta’s model stack is tied to a much larger infrastructure bet: a reported $250 billion including chips over the next year, about 14 GB of data-center capability, MTIA 400 chips that are 400% faster and use 51% more HBM, and a stated plan to put 15 gigawatts on the ground.
The Silicon Token Expenditure Index, Bloomberg ticker SD LLMTK, is said to be down 20% over the last couple of weeks and months, showing enterprises spending less overall on AI tokens while companies like Uber, Meta, Microsoft, Amazon, Coinbase, and Walmart tighten usage or adopt cheaper open models.
A Goldman forecast is cited for token consumption rising 24x from 2026 to 2030 to 120 quadrillion tokens per month, while the counterweight is that cheap models often use more tokens per answer and enterprise budgets are already capping usage.
The conclusion is that intelligence has split into two markets: roughly 80% routine work will go to cost-per-intelligence routing, while the remaining hard 20% — frontier math, verification, and business-building — can still justify Fable-class, GPT 5.6, and GPT 6.0 spend.
No specific sponsor was detailed beyond a recruiting-style ask to refer people via X or email so the operation can become self-sufficient.