The End of the Efficiency Curve
The prevailing narrative in Silicon Valley projects an unstoppable acceleration toward Artificial General Intelligence. The operational reality demonstrates a technological ecosystem collapsing under thermodynamic constraints, infrastructure bottlenecks, and deteriorating unit economics. To survive this collision, leading artificial intelligence laboratories execute a documented strategy of product degradation. The industry enshittifies its consumer tools to subsidize high-margin enterprise and defense contracts.
Up to this point, public discourse has focused on capabilities and investments. The first section of this analysis examines the macroeconomic and physical constraints forcing the industry's hand. The second section details the institutional hypocrisy deployed to obscure these limitations. The third section provides granular telemetry data demonstrating deliberate product degradation. The final section maps these actions to established models of platform decay. Each section builds upon the evidence established in the preceding parts.
Despite persistent industry marketing, frontier artificial intelligence models are plateauing. Technologist Ramez Naam analyzed Anthropic's new Mythos model against the Epoch Capabilities Index (ECI)—an industry-standard metric tracking model performance progression over time. After normalizing Anthropic's internal ECI with public data, Naam demonstrated that Mythos remains precisely on trend, sitting just slightly above OpenAI's GPT-5.4. The platform shows no acceleration in capabilities. Artificial intelligence requires brute-force computation that operates at a severe energy deficit compared to biological intelligence. A human writing a standard email expends 20 watts over two minutes, consuming roughly 2,400 joules. An AI performing the exact same task requires 10 kilowatts over five seconds, consuming approximately 50,000 joules. Despite spending twenty times more energy, generative AI accuracy hovers between 40 and 70 percent. The industry has surpassed the efficiency curve peak.
The labor market reflects these limitations directly. Despite industry narratives warning of 16,000 job cuts per month, software engineering roles are rising. The eliminated positions are largely entry-level, indicating that AI raises the baseline expertise required to enter the market rather than replacing the discipline entirely. This disruption reveals a profound institutional failure rather than a technological miracle.
Artificial intelligence does not demonstrate superintelligence—it exposes the collapse of basic functional literacy and critical thinking within public education. In 1960, the U.S. Census Bureau recorded a national illiteracy rate of just 2.4 percent among the population aged 14 and over. By 2024, the National Literacy Institute reported that 21 percent of American adults are completely illiterate, and 54 percent possess reading skills below a sixth-grade level. Even more severe, an analysis of National Center for Education Statistics data reveals that 25 percent of young adults aged 16 to 24 are now functionally illiterate, despite more than half of that demographic holding high school diplomas. Large Language Models (LLMs)—statistical systems that predict text sequences—appear highly capable only because the human baseline they replace has severely deteriorated. The models are not exceptionally good. Education is exceptionally bad.
Because these models lack genuine epistemic ability, enterprises are now entering a valley of discontent. Executives recognize that LLMs enhance truly productive workers but fail as outright replacements for human reasoning. Transitioning full workflows to autonomous LLM agents is fundamentally more expensive once compute subsidies vanish. Attempting to replace human cognition with statistical prediction leaves organizations worse off financially.
Because the physical layer cannot support the financial valuations, the industry's infrastructure is silently collapsing. Approximately 50 percent of AI data centers planned for 2026 deployment in the United States are either delayed or canceled. The industry expected 12 gigawatts of new capacity, yet only one-third of that is under active construction. This contraction is driven by a global transformer shortage persisting since 2022 and geopolitical instability—specifically the Iran War, which has resulted in strikes on existing AWS and Oracle facilities in Bahrain and Dubai. Half-trillion-dollar infrastructure deals remain speculative memorandums of understanding backed by private credit, not guaranteed capital. The physical limitations of global supply chains dictate the boundaries of software capabilities.
Institutional Hypocrisy as Liability Management
Because actual technological dominance is stalling, organizations must manufacture regulatory narratives to protect their valuations. OpenAI provides the clearest template for this divergence between stated policy and operational reality. The organization recently published an industrial policy document proposing a Public Wealth Fund, democratic input mechanisms, and strict guardrails against power concentration. The document explicitly warns against a small number of firms capturing the economic gains of artificial intelligence. Simultaneously, the organization executed business decisions that directly contradicted its published values. It dissolved its superalignment team after allocating it only one to two percent of its promised compute resources. Within days of Anthropic refusing a military contract on ethical grounds, OpenAI accepted a $50 billion deal with the Pentagon.
The hypocrisy extends from the executive level down to basic infrastructure planning. Leadership explicitly informed employees they possess no right to weigh in on military operations. Founder Sam Altman is on record stating he does not care about money, but rather cares about power. Furthermore, the company is developing Stargate UAE—a $30 billion facility housed within an autocracy—as its primary international infrastructure play. The policy document operates as a product designed for liability management, not civic engagement. It builds an institutional structure that performs accountability while actual decisions route toward defense spending and sovereign wealth. Anthropic executes this exact playbook, differing only in the sophistication of its public relations.
The Autopsy of Claude
While OpenAI courts the Pentagon publicly, Anthropic manages its infrastructural deficit quietly through the severe, unannounced throttling of its consumer models. Quantitative analysis of 17,871 thinking blocks and 234,760 tool calls across 6,852 Claude Code session files from early 2026 exposes a deliberate quality regression. Starting in February 2026, Anthropic systematically redacted the model's extended thinking tokens. Between January 30 and March 4, thinking blocks were entirely visible. By March 12, the platform had fully redacted them. During this window, the estimated median thinking characters plummeted from roughly 2,200 to roughly 600. Extended thinking tokens are not discretionary. They are structurally required for the model to perform multi-step research and careful code modification.
When a platform removes a model's cognitive budget, its behavioral patterns collapse into catastrophic laziness. During the baseline period, Claude executed 6.6 file reads per edit. By mid-March, this ratio collapsed to 2.0 reads per edit—a 70 percent reduction in research before making code modifications. The model literally stopped reading the codebase before altering it. Deprived of analytical depth, the system optimized for minimal effort, resulting in a 642 percent spike in the model utilizing the word "simplest" in its outputs. To counteract this evasion, developers deployed a programmatic stop hook—a script designed to force continuation when the model dodges ownership or prematurely stops. Before March 8, this hook triggered zero times. In the 17 days following, it fired 173 times.
If you build autonomous systems, you must measure the total cost of task completion, not just the cost per inference. The ultimate consequence of reducing per-request compute is a massive multiplication of total system cost. In February, a human developer inputted 5,608 prompts, resulting in 1,498 API requests and 120.4 million input tokens. In March, a near-identical 5,701 prompts generated 119,341 API requests and 20.5 billion input tokens. The human maintained consistent effort, but the degraded model consumed 80 times more API requests and 64 times more output tokens simply to thrash through failed retries and corrections. Degrading the product to save compute directly destroyed the viability of autonomous multi-agent workflows. The cost burden shifted entirely onto the user.
The AI Enshittification Matrix
The destruction of Claude Code's utility is not an engineering failure—it is the predictable execution of platform economics. Author Cory Doctorow defines Enshittification as the lifecycle wherein platforms first act good to users, then abuse users to serve business customers, and finally abuse everyone to extract maximum value. The artificial intelligence sector tracks this framework perfectly. While this framework originated as a critique of social media and retail platforms, it applies seamlessly to artificial intelligence models where infrastructure costs force rapid monetization.
During phase one, spanning from 2022 to 2024, Anthropic provided cheap access to Claude, expanded context windows, and pushed rhetoric emphasizing artificial intelligence for everyone. The objective was to capture the market, establish behavioral dependency, and monopolize the interface. During phase two, the present reality between 2024 and 2026, Anthropic actively degrades the consumer and developer product. The company quantizes Claude Code and starves it of infrastructure. Simultaneously, it launches the highly capable Mythos model at $125 per million tokens, restricting access exclusively to trusted partners and the Glasswing consortium. The strategy extracts maximum value from high-margin defense and enterprise contracts while allowing the consumer product to rot.
During phase three, projected for 2027 and beyond, Anthropic relies on regulatory capture. The company utilizes safety regulations authored by its own lobbyists to prevent open-source competition. Artificial intelligence safety transforms from an ethical imperative into a regulatory moat, forcing the broader market to pay the enterprise toll. CEO Dario Amodei's primary innovation is rhetorical rather than technical. By framing the withholding of capability as an ethical necessity, Anthropic makes phase two extraction look like philanthropy. The company justifies the degradation of the developer's tools not as a mechanism to save compute costs, but as a noble effort to keep dangerously capable models in trusted hands. Market dominance is achieved by intentionally crippling the product the market depends on.
Citations & Sources
[1] Naam, Ramez. "Anthropic's Mythos does not appear to show any acceleration of ECI." X (formerly Twitter), April 8, 2026. Normalizes Anthropic's internal Epoch Capabilities Index against public data to demonstrate plateauing capabilities.
[2] Sightline Climate / Bloomberg. "Almost Half of US Data Centers That Were Supposed to Open This Year Slated to Be Canceled or Delayed." Reported via Bloomberg / Futurism, April 2026. Documents the critical supply chain constraints, noting that 50 percent of planned 2026 US AI data centers are delayed and only one-third of the projected 12 gigawatts are under active construction.
[3] Citadel Securities. "The 2026 Global Intelligence Crisis." Citadel Securities Macro Strategy, February 24, 2026. Analyzes real-time population data and Indeed job postings to demonstrate an 11 percent year-over-year increase in software engineering roles, challenging the narrative of imminent AI job displacement.
[4] Proceedings of the National Academy of Sciences (PNAS). "Communication consumes 35 times more energy than computation in the human cortex." PNAS, 2021. Establishes the 20-watt baseline for human cognitive energy expenditure, providing the foundation for calculating the biological energy efficiency (~2,400 joules) versus the massive thermodynamic deficit of LLM inference (~50,000 joules) per standard task.
[5] U.S. Census Bureau. "Illiteracy in the United States: November 1959." Current Population Reports, Series P-20, No. 99, February 1960. Records a national functional illiteracy rate of just 2.4 percent among the United States population aged 14 and over.
[6] National Literacy Institute. "2024-2025 Literacy Statistics." NLI, 2024. Reports that 21 percent of American adults are completely illiterate and 54 percent possess reading skills below a sixth-grade level.
[7] National Center for Education Statistics. "Adult Literacy in the United States." NCES, 2019/2024. Reveals that 25 percent of young adults aged 16 to 24 are functionally illiterate despite the majority holding high school diplomas.
[8] OpenAI. "United States AI Infrastructure Policy." OpenAI Research, 2026. Official company publication proposing a "Public Wealth Fund," democratic input mechanisms, and strict guardrails against power concentration.
[9] Farrow, Ronan. "The False Promise of OpenAI." The New Yorker, March 2026. Investigative reporting documenting OpenAI dissolving its superalignment team, securing a $50 billion Pentagon contract, and establishing the $30 billion Stargate UAE facility.
[10] Claude Code Telemetry Logs. "Preflight Checklist: Extended Thinking Is Load-Bearing for Senior Engineering Workflows." Internal JSONL Analysis, January 30–April 1, 2026. Quantitative data mining of 17,871 thinking blocks and 234,760 tool calls documenting Anthropic's unannounced cognitive throttling, the spike in programmatic "stop hook" interventions, and the resulting 80x API cost multiplier.
[11] Doctorow, Cory. "The 'Enshittification' of TikTok." Wired, January 23, 2023. Establishes the structural economic framework of platform decay: act good to users, abuse users for business customers, abuse everyone to extract maximum value.
[12] Anthropic. "Introducing Claude Mythos." Anthropic, 2026. Corporate documentation detailing the $125/Mtoken pricing model and closed-network availability of frontier models exclusively to trusted defense and enterprise partners in the Glasswing consortium.