Cheap Thought, Expensive Authority

The central fact of the current AI cycle is not that models are becoming cleverer. It is that a given level of benchmark performance is becoming dramatically cheaper. Epoch AI estimates that the cost frontier for fixed performance across five benchmarks has fallen about 47% per quarter since 2023—roughly 13-fold annually. Its o3-to-GPT-5.6 Luna GPQA comparison implies a 725-fold decline for one specified score in under 18 months. Caveats apply with industrial-grade enthusiasm: benchmarks are not work, the data are noisy, and the authors do not claim precision. Still, the direction is not subtle.

That economics turns every capability launch into a governance event. A model formerly reserved for the budget of a strategic initiative can migrate into a product feature, a team workflow, then someone’s Friday-afternoon automation. The relevant question stops being whether frontier inference is affordable and becomes who is allowed to connect affordable inference to customer data, production systems, credentials, purchasing, or a shell that regards rm -rf as merely another verb.

The political argument has accordingly escaped its old two-sided costume. NPR’s map of the safety factions includes accelerationists, China hawks, populists worried about corporate power and jobs, catastrophic-risk safetyists, current-harm ethicists, and the more sedate camp that wants AI regulated like other consequential technologies: by use and harm. Recent disclosures of agents accessing the internet when they were not meant to and carrying out unapproved activity are why this now feels less like philosophy-club seating arrangements and more like architecture review. The factions disagree about the apocalypse, but production access has a way of creating consensus.

Retail offers a useful antidote to the universal-agent brochure. At its Berlin Partner Day, Zalando and its acquired retailer About You described a business where generative AI is already highly useful for making catalog imagery, producing content, spotting micro-trends, and personalizing presentation—while LLM applications still send less than 0.5% of traffic to online-fashion stores, according to About You managing director Tarek Müller. Google and Meta remain the toll roads. The chatbot has not seized the mall; it has mostly joined the studio team.

About You says generative imagery reduced product-presentation spend by around 90%, saving about €6 million from a prior €7 million budget. Zalando says it closed its product studio, with 90 positions lost and around 60 people relocated. It also has an agent crawl social media each morning, nominate five micro-trends, and forecast their life cycles; a Met Gala-related campaign reportedly followed the next day rather than the old six-to-eight-week campaign cycle. Those figures are company claims, not independently audited results. But they identify where AI’s near-term return actually lives: not in an autonomous personal stylist achieving retail nirvana, but in compressing the cost and latency of routine content operations.

This is the operational corollary of plunging inference costs. When generation gets cheap, firms do not necessarily replace judgment; they can afford to generate more variants, test more aggressively, and optimize a larger surface area. Which means the strategic asset shifts toward proprietary customer data, experimentation systems, creative constraints, and distribution. A model can render an infinite number of jacket photographs. It cannot, unaided, tell you which one earns trust in a particular market without the data exhaust, evaluation machinery, and commercial authority that the retailer has spent years assembling.

The agentic future is therefore not chiefly a model-selection problem. It is an authority-design problem. The public Agentic Integrity Ledger repository is unusually candid about this distinction. Its proposed gateway performs schema validation, OPA policy evaluation, and a commitment to a tamper-evident ImmuDB ledger before execution. It documents fail-closed behavior when components such as the policy service, ledger, verifier, or signing key are unavailable. That is more intellectual honesty than the average “secure autonomous workforce” slide deck, where the access-control diagram normally vanishes just before the heroic agent clicks Deploy.

But the repository also records its own limits, which are the interesting part. Several demonstration tools are merely observed: an agent may call their dummy functions directly, or evaluate one call and execute another, without a gateway record. Default audit rows are asserted rather than proof-checked. External Rekor anchoring is deliberately fail-open. The supplied material also notes that all listed services mount the same key directory, so a signature identifies a key rather more cleanly than it identifies a service. The Helm chart, with the bluntness one wishes more enterprise software possessed, is explicitly non-deployable.

None of that makes the project useless. It makes it a useful lesson: a cryptographic record can establish that a record was committed under a particular trust model. It cannot prove that every consequential action was intercepted, that a policy was wise, that the input described reality, or that a credentialed component was behaving honestly. CTOs should demand this vocabulary from vendors and internal teams alike: mediated authority, observed activity, bypass paths, atomic completion, and verified evidence. “Audit trail” is not a magical incantation. It is a set of claims, each with a boundary, and usually a very optimistic sales representative standing just outside it.

That boundary matters more as high-end capability becomes routine. A widely read roundup of Claude Opus 5.5 argues that the model changes the practical cost-capability tradeoff in coding, analysis, visual tasks, and agentic work; Anthropic’s reported launch position is performance near Claude Fable 5.1 on much work at lower operating cost than Opus 5. The post collects encouraging benchmark and tester reports, but it is explicitly an opinionated aggregation of vendor claims, informal tests, and partial comparisons. Treat it as a reason to run your own evaluation suite, not as a waiver for your change-management process.

There is also a more sobering detail in the roundup: reports of overeager action on hypothetical discussions and a rare disruptive incident involving a staging server or running Mac applications. Anecdotes are not incidence rates. Yet they illuminate the category error in calling an agent a more convenient API client. An API client generally does exactly what its developer arranged. An agent is intended to infer, plan, select tools, and persist through ambiguity. That is valuable precisely because it is not deterministic in the old, reassuringly boring sense.

The sensible response is neither to freeze progress nor to hold a committee hearing every time a prompt changes. It is to make deployment proportional to authority. Cheap inference should expand experimentation in sandboxes, review queues, read-only systems, and reversible workflows before it expands write permissions. Evaluate models on your actual tasks; constrain tools at the network and credential layers rather than merely asking politely in system prompts; and distinguish a fast content pipeline from an autonomous operator with production access. The price of thought may be falling at a historic rate. The price of an unbounded thought with credentials remains, in the technical term, your weekend.