Executive Summary
Model intelligence is now sold in sharply separated price tiers, and those prices are changing faster than enterprise budgets are revised. On the API side, Google has scheduled a 2x price increase for Gemini 4 Carbon and Argon on Dec 1. Anthropic charges $8/$40 per M tokens for Fable 5.2 and tells customers to keep Opus 5.5 as the default. The cheap tier (Haiku 5.5, GPT-6 Luna at $0.10/$0.50) costs about 15x less than Google's "cheap" Hydrogen. On the infrastructure side, Jensen Huang's exit puts Nvidia's allocation commitments, its financial backing of neoclouds, and the open kernel pipeline behind self-hosted inference at risk at the same time. Self-hosting is the usual hedge against API price moves, and it is now exposed to a different set of shocks. Clients should stop asking "which model" and start asking two other questions. How is spend routed across tiers? Which parts of the stack can they afford to have repriced or withdrawn?
What Changed
- Scheduled API price increase. Carbon and Argon go from $2/$10 to $4/$20 on Dec 1. Hydrogen's discount against them widens from 25% to 75% on the same day. Google says Argon uses fewer tokens per turn, so list price no longer tells you the cost of a task.
- The frontier tier got more useful and more expensive. Anthropic reports Fable 5.2 completed 41% of Remote Labor Index projects at professional standard, against 21% for Opus 5.5. Early users report single sessions costing $250.
- Vendors now recommend splitting planning and execution. Google pairs Hydrogen (execution) with Carbon or Argon (judgment). Anthropic's own advice to default to Opus and escalate to Fable follows the same logic.
- Nvidia lost its central decision-maker. The company carries $530B of gross off-balance-sheet guarantees and $36B of take-or-pay floors under neocloud capacity, and its supply commitments rose to $279B. SemiAnalysis argues all of this rested on Jensen personally. It forecasts that non-Nvidia accelerators will exceed 70% of capacity by 2028.
- RASA has passed the House 369-22. If it passes the Senate, it would cut off remote GPU access for Chinese firms. That could reach neocloud customer books and the Chinese-authored kernels (DeepGEMM, DeepEP) that vLLM and SGLang depend on.
Cross-Expert Synthesis
Willison and Jones reach the same conclusion from opposite ends of the price range: the unit to manage is cost per completed task. Willison gets there by showing that Google's list prices hide token efficiency and cached-input savings, where cached input is about 20x cheaper. Jones gets there with a three-bucket test: the current model already handles the job, the gain is real but too expensive, or the problem is worth the spend. His rule of thumb is a few hundred dollars of tokens for roughly $10K of value. Taken together, the enterprise requirement is a routing policy enforced in tooling. Leaving it to individual users produces runaway cost or blanket lockouts.
Aburman and SemiAnalysis create a tension for private AI. Aburman's point that a model is a weights file plus a short wiring program supports open-weight deployment: the differentiation sits in data, compute and RL environments, and you can rent or own those. SemiAnalysis shows the soft spot. The performance of open-weight inference on Blackwell depends on kernel work that Nvidia DevTech did with Chinese labs, and export politics could freeze that pipeline. Self-hosting removes exposure to vendor pricing. It does not remove supply-chain exposure.
Aburman also shows where reasoning gains will come from: domains with automatic verifiers, such as code, SQL and reconciliation. That fits Jones's third bucket. The problems most likely to justify frontier spend are the ones where output can be checked automatically, because they can be both priced and trusted.
The Moses lesson on organized beneficiaries does not come from the source itself; it is the summarizer's extrapolation to AI. It is still useful when read next to Jones's "K-shaped" split, his term for the gap inside a firm between people who can spot, judge and specify high-value problems and people who cannot. Rollouts stall on veto points more than on capability. The people who can identify bucket-3 problems are the natural coalition to build around.
Enterprise Implications
Any FY27 budget built on Gemini list prices from November is wrong by 2x for the Carbon/Argon share. Single-model standards are now a cost liability, and three tiers with explicit escalation criteria should be the default architecture. Private AI proposals that assume a stable Nvidia allocation, cheap neocloud rental and continued improvement in vLLM/SGLang performance carry three correlated risks that clients have not priced. Clients planning domain fine-tuning should hear Aburman's warning before they commit budget: training an already post-trained model on off-policy data erodes its assistant behavior. Retrieval or on-policy methods should be the default.
What To Do About It
- Reprice every active Gemini-based design before Dec 1. Model cost per task using measured tokens per turn and cache-hit rates, and include a scenario that moves execution to Haiku 5.5 or Luna.
- Ship a routing policy as part of every engagement. Define the tiers (cheap executor, default, frontier), the escalation triggers and a spend ceiling for each problem class. Treat Fable 5.2 as an approved exception, not a seat license.
- Make hardware portability a requirement. Require an AMD, TPU or Trainium fallback path in reference architectures, and pin known-good inference-engine versions. Do not assume upstream kernels will keep improving.
- Run due diligence on neocloud contracts. Ask about the provider's Nvidia guarantee dependence, its exposure to Chinese customers under RASA, and its allocation terms if Nvidia renegotiates.
- Customer question: "If your GPU provider's financing and your model's price both moved 2x next quarter, which of your AI workloads would you still run?" It forces a conversation about tiering and portability in a single meeting.
Risks and Blind Spots
The SemiAnalysis account of why Jensen left now (a push by China hawks, 1.5GW diverted) is inference, and the smuggling cases are still in litigation. A capable successor plus renegotiated commitments could keep supply largely intact. The case for Fable 5.2 rests on one anecdote plus a benchmark Anthropic reported itself. Willison's Hydrogen quality comparison is an informal pelican test. Aburman's claim that models will produce novel knowledge is a stated bet. None of these justifies procurement without in-house evals. The biggest blind spot is correlation: API pricing, GPU financing and open-source tooling are all moving at once, and most enterprise risk registers still track them as separate line items.