Executive Summary
Today's sources point the same way. The frontier model is turning into a perishable, repriced input, and the lasting value and risk are moving to two layers the enterprise controls: the harness (the routing, interfaces and controllers around the model) and the knowledge substrate (the code, chat and documents agents get access to). Anthropic's S1 shows buyers moving to new models within six months. Google's Gemini 4 tiers reverse their price order inside a month. A general coding agent drives a physical robot after nothing more than an interface change. Christiano argues that lab secrets leak in hours, which means no model lead lasts. For enterprise buyers, the advantage goes to whoever can switch models cheaply and govern agent permissions tightly. Picking the "right" model matters less.
What Changed
- Vendor filings now document model churn. About 80% of Anthropic usage runs on models less than six months old, and most revenue is swappable usage-based API spend. Customers are already switching models at the API layer.
- List prices stopped predicting cost. Gemini 4 has four tiers, four reasoning-effort levels, a launch discount ending December 1 that doubles Carbon to $4/$20, and Argon has dropped out of the lineup without a deprecation notice. Willison's own ratios don't add up, so even careful observers can't work out the cost structure.
- The coding agent has become a general-purpose actuator. It is Anthropic's only proven mega-vertical. With a controller underneath, it completed 24 of 25 physical robot trials. It is also the part of the stack that gets write access to repositories.
- Frontier-lab security is reportedly weak. One startup is reported to have gained pull-request access to OpenAI's monorepo, and another reached all of Anthropic's Slack in 12 hours. Both claims are unverified and the excerpt is truncated.
Cross-Expert Synthesis
Jaipuria, Willison and Jones all treat the model as interchangeable and put the durable work around it. Jaipuria expects open-weight token margins to stay under 40% and sees labs defending themselves through app-layer products. Willison's model-agnostic llm tool makes re-benchmarking routine. Jones cites a study where changing only the software around a fixed model raised scores substantially. That is three independent lines of evidence that integration quality now decides outcomes as much as model choice.
Christiano supplies the reason this keeps happening. If algorithmic insight leaks within hours, no lab holds a capability lead for long. Commoditization then becomes the stable state of the market. It is not a pricing phase that will pass.
That conclusion conflicts with the vendors' economics. Anthropic has about $900B in compute commitments, growth is slowing from more than 100% to 40-45% quarter over quarter, and the usage it relies on is easy to swap. Vendors therefore need stickiness, and they will pursue it through agents, workflow products (Claude Code, Cowork) and confusing pricing. Buyers should expect 12 months of heavy lock-in attempts dressed up as capability launches.
A second conflict runs between Jones and Christiano. Agents get more useful as they gain more authority: repository write access, camera control, actuators. Christiano's evidence says the systems those agents touch are where breaches happen.
Enterprise Implications
The main growth lever is diffusion, not intelligence. Even accounts spending over $100K put less than 0.1% of opex into Anthropic, and Jaipuria projects 10-30x growth in token spend from broader adoption alone. The architecture decisions enterprises make in the next 18 months will therefore lock in long-term cost structure. That matters more than which model they choose now.
For private AI, this changes the pitch. Open-weight inference is already priced at 10-20% of lab rates, and Anthropic reports about 85% gross margin on paid inference with no need to raise prices to recover losses. Self-hosting rarely wins on price, so it has to win on data control, latency, customization and post-training. The post-training point matches Jaipuria's view that inference providers will compete on customization.
The security finding also redraws the crown-jewel map. Slack and source control hold strategic knowledge, and they are exactly what enterprises are now connecting to RAG and coding agents. Write access to a monorepo is a supply-chain integrity risk, not just a confidentiality risk.
What To Do About It
- Make model routing a contract requirement. Every deployment should be able to swap models within days. Evaluate on cost per completed task across tier and effort level, not per-token list price.
- Freeze Gemini 4 volume commitments until after December 1. Get Argon's lifecycle status from Google in writing.
- Treat agent permissions as supply-chain controls. Coding agents get least-privilege repo scopes, mandatory human review of their pull requests, and separate credentials. Classify collaboration platforms as crown-jewel systems before connecting them to agents.
- Re-run evals on every frontier release. This includes hardware and robotics pilots, scored on end-to-end task time and intervention rate rather than success rate.
- Customer question: "Your vendor's own filing says 80% of usage moves to a newer model within six months. What in your architecture lets you move in a week, and which of your agents can open a pull request today without a human approving it?"
Risks and Blind Spots
The security evidence is the weakest of the four and has the largest consequences. The excerpt is truncated, describes no attack method, and doesn't say whether either incident was an authorized engagement. Don't put "12 hours" in a client deck. The robotics evidence is also thin (25 trials, mostly simulation, one lab), and latency remains the binding constraint: about 90% of time goes to model deliberation and successful runs take over an hour.
The commoditization thesis has a counter-case. If a frontier lead lasts longer than six months, buyers who built routing will have paid an abstraction cost for nothing. That cost is cheap insurance compared with lock-in.
Finally, Anthropic's efficiency numbers come from roadshow material prepared for investors. The S1 figures are the more reliable source and should be used where they differ.